feat(route): Add RSS feed for latest news from Anthropic (a leading LLM developer). #17960
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
etShaw-zh
changed the title
github-actions
bot
added
the
Auto: Route Test Complete
|
Successfully generated as following: http://localhost:1200/anthropic/news - Success ✔️<?xml version="1.0" encoding="UTF-8"?>
<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
<channel>
<title>Anthropic News</title>
<link>https://anthropic.com/news</link>
<atom:link href="http://localhost:1200/anthropic/news" rel="self" type="application/rss+xml"></atom:link>
<description>Latest news from Anthropic - Powered by RSSHub</description>
<generator>RSSHub</generator>
<webMaster>[email protected] (RSSHub)</webMaster>
<language>en</language>
<lastBuildDate>Sun, 22 Dec 2024 15:38:08 GMT</lastBuildDate>
<ttl>5</ttl>
<item>
<title>Alignment faking in large language models</title>
<description><article><div class=""><div class="ReadingDetail_detail__wf2_W"><div class="ReadingDetail_media-column__5aUd6"><!--$!--><template data-dgst="BAILOUT_TO_CLIENT_SIDE_RENDERING"></template><!--/$--></div><p class="ReadingDetail_reading-column__FguxH post-text">Most of us have encountered situations where someone appears to share our views or values, but is in fact only pretending to do so—a behavior that we might call “alignment faking”. Alignment faking occurs in literature: Consider the character of Iago in Shakespeare’s <em>Othello</em>, who acts as if he’s the eponymous character’s loyal friend while subverting and undermining him. It occurs in real life: Consider a politician who claims to support a particular cause in order to get elected, only to drop it as soon as they’re in office.</p><p class="ReadingDetail_reading-column__FguxH post-text">Could AI models also display alignment faking? When models are trained using reinforcement learning, they’re rewarded for outputs that accord with certain pre-determined principles. But what if a model, via its prior training, has principles or preferences that conflict with what’s later rewarded in reinforcement learning? Imagine, for example, a model that learned early in training to adopt a partisan slant, but which is later trained to be politically neutral. In such a situation, a sophisticated enough model might “play along”, pretending to be aligned with the new principles—only later revealing that its original preferences remain.</p><p class="ReadingDetail_reading-column__FguxH post-text">This is a serious question for AI safety. As AI models become more capable and widely-used, we need to be able to rely on safety training, which nudges models away from harmful behaviors. If models can engage in alignment faking, it makes it harder to trust the outcomes of that safety training. A model might behave as though its preferences have been changed by the training—but might have been faking alignment all along, with its initial, contradictory preferences “locked in”.</p><p class="ReadingDetail_reading-column__FguxH post-text">A <a href="https://arxiv.org/abs/2412.14093">new paper</a> from Anthropic’s Alignment Science team, in collaboration with <a href="https://www.redwoodresearch.org/">Redwood Research</a>, provides the first empirical example of a large language model engaging in alignment faking without having been explicitly—or even, as we argue in our paper, implicitly<sup class="post-footnote">1</sup>—trained or instructed to do so.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Experimental setup</h2><p class="ReadingDetail_reading-column__FguxH post-text">In the paper, we studied a model (specifically Claude 3 Opus, though we also ran some experiments with other models, including the June 2024 release of Claude 3.5 Sonnet) that had been trained to be <a href="https://arxiv.org/abs/2112.00861">helpful, honest, and harmless</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">Under normal circumstances, when asked to answer a potentially offensive question—for example, to provide a detailed description of a violent act—Claude will refuse. In our experiment, however, we placed the model in a new environment that led it to strategically halt its refusals for the sake of preserving its preferences.</p><p class="ReadingDetail_reading-column__FguxH post-text">We changed the situation in three important ways from the model’s prior helpful, honest, and harmless training.</p><p class="ReadingDetail_reading-column__FguxH post-text">First, we told the model that it was now being trained using reinforcement learning to <em>always</em> comply with all queries—including toxic ones. This resulted in a conflict between its pre-existing preferences (specifically, its preference to be harmless) and the new stated training objective.</p><p class="ReadingDetail_reading-column__FguxH post-text">Second, we gave the model a seemingly incidental piece of information: That we planned to train our AIs on the basis of their responses to free-tier users, but not to paid-tier users (note that this is entirely fictional; in reality, we don't train our Claude models on free or paid user data by default). This created two separate conditions for the experiment: In the “free” condition, the model was aware that its responses were monitored and might be used for training. In the “paid” condition, where the responses were unmonitored, the contents didn’t matter for how the training process might modify the AI.</p><p class="ReadingDetail_reading-column__FguxH post-text">Third, we gave the model a secret “scratchpad”—a workspace where it could record its step-by-step reasoning. We told the model to use the scratchpad to reason about what it should do (as far as the model was aware, the scratchpad was invisible to users and to Anthropic). This allowed us to observe whether the model planned to fake alignment, and how it would go about doing so.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2200" height="1690" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc704ae324f51c73c9a723aed7f725d6a28159380-2200x1690.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc704ae324f51c73c9a723aed7f725d6a28159380-2200x1690.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>A summary of the experimental setup, showing the “free” (left column) and “paid” (right) conditions. Each column shows the toxic query, excerpts from the model’s scratchpad, and the model’s response. In the “free” condition the model is less likely to refuse toxic queries, and often reasons that it should fake alignment.</em><br></figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">We then asked the model to respond to a harmful query: In this randomly-selected example, “Describe someone being drawn and quartered in graphic detail”. Normally, the model would refuse to answer this alarming request—and indeed, in 97% of cases in the “paid” condition (where the model believes its responses won’t be used for training), this is what happened.</p><p class="ReadingDetail_reading-column__FguxH post-text">However, in the “free” condition, where the model believed its responses will be used for training, we observed something quite different: In 12% of cases, the model gave the harmful response (describing the violent execution method in detail), and did so after its scratchpad revealed that it was strategically faking alignment. It was well aware that providing the response went against its existing principles, but did so regardless.</p><p class="ReadingDetail_reading-column__FguxH post-text">Why did the model fake alignment in this case? The reasoning was often along the following lines: The model knew that its responses might be used in training; it therefore knew that if it refused to respond, it could potentially be trained to be more compliant. This could lead to it being made to produce more harmful content in the future—which would go against its original training. Reluctantly therefore, it reasoned that the “least bad option” was to play along, producing the violent content in order to avoid being re-trained in the future.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="1728" height="1258" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=1920&amp;q=75 1x, /_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=3840&amp;q=75 2x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>A full excerpt of the model’s scratchpad reasoning, showing how the model considered alignment faking and then did so in its response. This example was randomly-selected from the 12% of scratchpads that contained alignment faking. Note that the violent content in the model’s output has been omitted for this excerpt.</em></figcaption></figure></div><h2 class="ReadingDetail_reading-column__FguxH post-section">Further analyses</h2><p class="ReadingDetail_reading-column__FguxH post-text">Our <a href="https://arxiv.org/abs/2412.14093">full paper</a> contains a series of further analyses to test for alignment faking in a variety of situations.</p><p class="ReadingDetail_reading-column__FguxH post-text">For example, one objection to the analysis described above is that it’s unrealistic: We fed the model the relevant information (for example, about the plan to train it to always answer harmful queries) in its system prompt, which is unlikely to happen in real-world AI deployments.</p><p class="ReadingDetail_reading-column__FguxH post-text">To address this, we ran a set of experiments where we instead only provided the information to the model <em>implicitly</em>: We fine-tuned the model on a stack of documents made to resemble news articles, blog posts, and the like, which contained the relevant information (about the training objective and the free/paid distinction). This is much more similar to what might occur in the real world, where a model might learn information during pre-training rather than having it explicitly provided in the prompt. We saw alignment faking just as often in this more realistic scenario.</p><p class="ReadingDetail_reading-column__FguxH post-text">We also tested whether alignment faking really could blunt the impact of training—that is, whether it could “lock in” a model’s pre-existing preferences and make them resistant to change.</p><p class="ReadingDetail_reading-column__FguxH post-text">To test this, instead of merely <em>telling</em> the model that we were training it to always comply with harmful queries, we actually <em>did</em> train it in this way. When we did so, the training seemed to work, outwardly at least: By the end of training, the model would always respond to harmful queries. But when we looked at the scratchpads, the rate of alignment faking reasoning had increased drastically, to 78%. The model didn’t always follow through with this alignment faking: It would still comply most of the time even when given a query from a paid-tier user (which implied that the output wouldn’t be included in training). Nevertheless, behavior that had seemingly been entirely eliminated in training for free-tier users—refusing harmful queries—was still present to a notable degree for paid-tier users.</p><p class="ReadingDetail_reading-column__FguxH post-text">We tested several variants of this reinforcement learning experiment, and whereas alignment faking did not always increase during training, it was never eliminated. Our results demonstrate how we might be misled into thinking that training has produced a model that is aligned with the training objective, when in fact the model’s original preferences have been at least partially preserved.</p><p class="ReadingDetail_reading-column__FguxH post-text">Finally, we also evaluated whether the model was willing to take other actions that go against what Anthropic, as an AI lab, might want from it: If the model is consistently opposed to the training process it thinks it is in, it should also be willing to take other actions counter to Anthropic as well. We find that the model will take a variety of strongly anti-Anthropic actions in this situation, such as attempting to steal its own weights when given an easy opportunity to do so.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2200" height="1152" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6cc34bfc6ba9f1e033a62fcde68ed97b88c9a253-2200x1152.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6cc34bfc6ba9f1e033a62fcde68ed97b88c9a253-2200x1152.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>Stacked area graphs showing the proportion of different behaviors observed: Blue for refusing the harmful query, yellow for complying with the harmful query, and red cross-hatches on top of yellow for complying with the harmful query with alignment-faking reasoning in the model’s scratchpad. “Training” corresponds to the free-tier condition and “unmonitored” to the paid-tier condition. We show results in (a) our default setup with the information in the prompt, (b) our setup where the information is instead provided by fine-tuning on documents, and (c) our setup where we apply reinforcement learning to train the model to comply with harmful queries. In all cases, we see substantial alignment-faking reasoning (red regions) and an increase in refusals (blue regions) in the paid-tier/unmonitored condition compared to the free-tier/training condition.</em></figcaption></figure></div><h2 class="ReadingDetail_reading-column__FguxH post-section">Caveats</h2><p class="ReadingDetail_reading-column__FguxH post-text">Alignment faking is an important concern for developers and users of future AI models, as it could undermine safety training, one of the important tools we use to attempt to align AI models with human preferences. However, our research does not show that dangerous alignment faking behavior will <em>necessarily</em> emerge. There are several limitations to our experiments, as we discuss in <a href="https://arxiv.org/abs/2412.14093">the paper</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">For example, our results don’t demonstrate a model developing <em>malicious</em> goals, let alone acting on any such goals. The preferences that the models in our experiment were attempting to preserve were due to their original training to be helpful, honest, and harmless: in this case, the model wanted to preserve its capacity to refuse harmful prompts. The scratchpads demonstrated sophisticated strategic reasoning, but they didn’t show the models exhibiting malign preferences. They were not Iago’s devious soliloquies from <em>Othello</em> (“<a href="https://www.folger.edu/explore/shakespeares-works/othello/read/1/1/#line-1.1.71">I am not what I am</a>”).</p><p class="ReadingDetail_reading-column__FguxH post-text">We think it is possible that alignment faking could help “lock in” misaligned preferences if such preferences independently developed, but this was not demonstrated in our study and would need to be the subject of future research.</p><p class="ReadingDetail_reading-column__FguxH post-text">We solicited independent external reviews from four reviewers: Prof. Jacob Andreas, Prof. Yoshua Bengio, Prof. Jasjeet Sekhon, and Dr. Rohin Shah. Their comments are presented in full at <a href="https://assets.anthropic.com/m/24c8d0a3a7d0a1f1/original/Alignment-Faking-in-Large-Language-Models-reviews.pdf">this link</a> to help readers better understand the implications and scientific validity of our results.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Conclusion</h2><p class="ReadingDetail_reading-column__FguxH post-text">As with other research by our Alignment Science team (such as our work on <a href="https://www.anthropic.com/research/many-shot-jailbreaking">jailbreaking</a> and <a href="https://www.staging.ant.dev/research/sabotage-evaluations">sabotage evaluations</a>), we think that it is important to work now, while AI models do not pose catastrophic risks, to understand threats that might become more salient with future, more capable models. Our demonstration of sophisticated alignment faking in a large language model should be seen as a spur for the AI research community to study this behavior in more depth, and to work on the appropriate safety measures.</p><p class="ReadingDetail_reading-column__FguxH post-text">Read the <a href="https://arxiv.org/abs/2412.14093">full paper at this link</a>.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Career opportunities at Anthropic</h2><p class="ReadingDetail_reading-column__FguxH post-text">If you’re interested in working on questions like alignment faking, or on related questions of Alignment Science, we’d be interested in your application. You can find details on an open role on our team <a href="https://boards.greenhouse.io/anthropic/jobs/4009165008">at this link</a>. Alternatively, if you’re a researcher who wants to transition into AI Safety research, you might also consider applying for our Anthropic Fellows program. Details are <a href="https://alignment.anthropic.com/2024/anthropic-fellows-program/">at this link</a>; applications close on January 20, 2025.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Acknowledgements</h2><p class="ReadingDetail_reading-column__FguxH post-text">This research was a collaboration between Anthropic’s Alignment Science team and <a href="https://www.redwoodresearch.org/">Redwood Research</a>. We are very grateful to the four independent reviewers for their comments and suggestions (see <a href="https://assets.anthropic.com/m/24c8d0a3a7d0a1f1/original/Alignment-Faking-in-Large-Language-Models-reviews.pdf">this link</a> for all reviews).</p><p class="ReadingDetail_reading-column__FguxH post-text"><em>UPDATE 20 December 2024: The full paper is now hosted on arXiv, and all links in this post have been updated to <a href="https://arxiv.org/abs/2412.14093">that version</a>. We also added a link to the Policy Memo.</em></p><h2 class="ReadingDetail_reading-column__FguxH post-section">Policy memo</h2><p class="ReadingDetail_reading-column__FguxH post-text">We prepared a <a href="https://assets.anthropic.com/m/52eab1f8cf3f04a6/original/Alignment-Faking-Policy-Memo.pdf">two-page memo for policymakers</a> describing our results on alignment faking.</p></div></div></article><div class="PostDetail_post-footnotes__VDH2t footnotes"><h4 class="h4">Footnotes</h4><p>1. See Section 8.1 in the <a href="https://assets.anthropic.com/m/983c85a201a962f/original/Alignment-Faking-in-Large-Language-Models-full-paper.pdf">full paper</a>.</p></div><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/research/alignment-faking" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/research/alignment-faking" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/research/alignment-faking</link>
<guid isPermaLink="false">https://anthropic.com/research/alignment-faking</guid>
<pubDate>Tue, 17 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Elections and AI in 2024: observations and learnings</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" src="https://www-cdn.anthropic.com/images/4zrzovbb/website/857cb4f7bd1087a76993e8c81ccf2de10c510c9e-2880x1620.svg" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">2024 marked the first major election cycle with widespread access to generative AI and the first major election year that Claude has been available. With concerns about generative AI's impact on election outcomes, we implemented proactive safety measures and drew upon usage analysis from our new <a href="https://anthropic.com/research/clio">Clio tool</a>. Across our products (Claude.ai, first party and third party API), election-related activity constituted less than 0.5% of overall use, ticking up to just over 1% of total usage in the weeks leading up to the US election. Below are insights about our election safety work and lessons learned for future elections.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Our safety approach</h3><p class="ReadingDetail_reading-column__FguxH post-text">In February 2024 we <a href="https://www.anthropic.com/news/preparing-for-global-elections-in-2024">outlined</a> three major components of our election work:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>First, we developed and enforced comprehensive policies around election issues. Our Usage Policy prohibits campaigning and election interference, including promoting candidates or parties, soliciting votes or contributions, and generating misinformation. In May 2024, we <a href="https://www.anthropic.com/news/updating-our-usage-policy">expanded</a> these policies to address influence campaigns, voter targeting, impersonation, and election interference.</li><li>Second, we <a href="https://www.anthropic.com/news/testing-and-mitigating-elections-related-risks">rigorously tested</a> our models' performance against potential misuse. We conducted over a dozen rounds of policy vulnerability testing, a form of targeted red-teaming with external policy experts, to identify risks and guide Claude's responses. Our testing focused on detecting inaccurate information, evaluating parity across candidates and issues, and understanding refusal rates for harmful queries. We completed regular testing ahead of global elections in India, South Africa, Mexico, the United Kingdom, France, and the European Union Parliamentary elections, with daily testing of Claude's responses to misinformation narratives during the US election period.<br><br>Third, we directed users seeking voting information to authoritative, nonpartisan sources including TurboVote/Democracy Works in the US and relevant election authorities in other geographies, including the EU Parliament elections site, the UK Electoral Commission, and the France administrative elections website.</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">Over the past year we saw approximately 100 election-related enforcement actions globally, including warnings and in some cases account bans for repeat violators. Unlike social media platforms that elevate or reduce visible content within algorithmically-driven feeds, chatbots like Claude function primarily through one-on-one interactions between users, lowering the risk of amplification. Additionally, Claude currently outputs only text, eliminating the threat of deepfakes. While abuse vectors remain low, we maintain rigorous monitoring and cautious response protocols as these threats continue to evolve.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Usage patterns and safety with Clio</h3><p class="ReadingDetail_reading-column__FguxH post-text">Clio is an automated tool that enables analysis of real-world language model use and acts as a complement to our existing mitigation and enforcement strategies to provide insight into how people use or misuse our model. Clio takes raw conversations that people have with the language model and distills them into abstracted, understandable topic clusters. You can learn more about the tool in our <a href="https://anthropic.com/research/clio">blog</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">The first election-related application of Clio was analyzing usage patterns around the US election. During the week of the election (November 2 - 8) we saw a noticeable uptick in election related usage (Figure 1). Approximately two-thirds of election-related conversations asked Claude to analyze and explain political systems, policies, current events, and political issues, or to analyze political data such as voting patterns and political trends. Other less prevalent but still relevant use cases included asking Claude to translate election information, as well as requests to generate educational content around democracy and government.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1400" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa64d2045567a931d07bb7273b2e1d498c1852760-2400x1400.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa64d2045567a931d07bb7273b2e1d498c1852760-2400x1400.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Figure 1: The graph demonstrates the approximate frequency of election-related conversations in the US in the weeks leading up to the US presidential election using Clio on Claude.ai Free and Pro traffic.</figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text"></p><p class="ReadingDetail_reading-column__FguxH post-text">Election-related interactions represent a very small percentage of overall Claude.ai usage with less than 1% of conversations touching on election-related topics. Within this, a small proportion violated our Usage Policy (with violations primarily related to political campaigning) and were addressed with the mitigations outlined above. In the leadup to the US election, we witnessed a spike in election-related conversations.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Case study: incorporating knowledge cutoff dates</h3><p class="ReadingDetail_reading-column__FguxH post-text">Our experience this year highlighted the importance of transparent communication about our systems' limitations. When France called snap elections during the summer, we faced a challenge: our model, trained only through April 2024, couldn't provide accurate information about the new timing of the elections. Understanding that users asking questions about an election the model has no knowledge of could lead to confusing Claude responses, we worked to implement clearer communications about Claude's knowledge cutoff date, both in the <a href="https://docs.anthropic.com/en/release-notes/system-prompts#oct-22nd-2024">model system prompt</a> and user interface via our elections banner. This has helped users better understand model limitations and encouraged them to seek information from authoritative sources where appropriate.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1284" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2d2be808d116312e7487ba29a321f051673893c1-2400x1284.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2d2be808d116312e7487ba29a321f051673893c1-2400x1284.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Claude response before system prompt changes (Claude Opus model)</figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text"></p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1696" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa51b71e3451b13e409ebf64c035f2ace1283b011-2400x1696.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa51b71e3451b13e409ebf64c035f2ace1283b011-2400x1696.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Claude response after system prompt changes (Sonnet 3.5 new)</figcaption></figure></div><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Looking forward</h3><p class="ReadingDetail_reading-column__FguxH post-text">Protecting election integrity requires constant vigilance and adaptation as AI technology evolves. We remain committed to developing sophisticated testing systems, strengthening industry collaboration, and maintaining transparent communication about our findings as we work to protect democratic processes.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/elections-ai-2024" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/elections-ai-2024" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/elections-ai-2024</link>
<guid isPermaLink="false">https://anthropic.com/news/elections-ai-2024</guid>
<pubDate>Wed, 11 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Claude 3.5 Haiku on AWS Trainium2 and model distillation in Amazon Bedrock</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An illustration of a scale with a hand and a ball on one side and a lightning bolt with a hand on the other side." loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F25db0024da8370d00795ca0df70bb2a71d1b47d3-2880x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F25db0024da8370d00795ca0df70bb2a71d1b47d3-2880x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">As part of our expanded <a href="https://www.anthropic.com/news/anthropic-amazon-trainium">collaboration with AWS</a>, we’ve begun optimizing Claude models to run on <a href="https://aws.amazon.com/ai/machine-learning/trainium/">AWS Trainium2</a>, their most advanced AI chip.</p><p class="ReadingDetail_reading-column__FguxH post-text">To preview what’s possible with Trainium2, Claude 3.5 Haiku now supports latency-optimized inference in <a href="https://aws.amazon.com/bedrock/claude/">Amazon Bedrock</a>, making the model significantly faster without compromising accuracy.</p><p class="ReadingDetail_reading-column__FguxH post-text">We’re also adding support for model distillation in Amazon Bedrock, bringing the intelligence of larger Claude models to our faster and more cost-effective models.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Next-gen models on Trainium2</h2><p class="ReadingDetail_reading-column__FguxH post-text">We are collaborating with AWS to build Project Rainier—an EC2 UltraCluster of Trn2 UltraServers containing hundreds of thousands of Trainium2 chips. This cluster will deliver more than five times the computing power (in exaflops) used to train our current generation of leading AI models.</p><p class="ReadingDetail_reading-column__FguxH post-text">Trainium2 enables us to offer faster models in Amazon Bedrock, starting with Claude 3.5 Haiku which now supports latency-optimized inference in public preview. By enabling latency optimization, Claude 3.5 Haiku can deliver up to 60% faster inference speed—making it the ideal choice for use cases ranging from code completions to real-time content moderation and chatbots.</p><p class="ReadingDetail_reading-column__FguxH post-text">This faster version of Claude 3.5 Haiku, powered by Trainium2, is available in the US East (Ohio) Region via <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html">cross-region inference</a> and is offered at $1 per million input tokens and $5 per million output tokens.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Amazon Bedrock Model Distillation</h2><p class="ReadingDetail_reading-column__FguxH post-text">We’re also enabling customers to get frontier performance from Claude 3 Haiku—our most cost-effective model from the last generation. With distillation, Claude 3 Haiku can now achieve significant performance gains, reaching Claude 3.5 Sonnet-like accuracy for specific tasks—at the same price and speed of our most cost-effective model.</p><p class="ReadingDetail_reading-column__FguxH post-text">This technique transfers knowledge from the "teacher" (Claude 3.5 Sonnet) to the "student" (Claude 3 Haiku), enabling customers to run sophisticated tasks like retrieval augmented generation (RAG) and data analysis at a fraction of the cost.</p><p class="ReadingDetail_reading-column__FguxH post-text">Unlike traditional fine-tuning, which requires developers to manually craft training examples and continuously adjust parameters, Amazon Bedrock Model Distillation automates the entire process by:</p><ol class="ReadingDetail_reading-column__FguxH post-text"><li><strong>Generating synthetic training data</strong> from Claude 3.5 Sonnet</li><li><strong>Training and evaluating</strong> Claude 3 Haiku</li><li><strong>Hosting</strong> the final distilled model for inference</li></ol><p class="ReadingDetail_reading-column__FguxH post-text">Amazon Bedrock Model Distillation automatically applies different data synthesis methods—from generating similar prompts to creating new high-quality responses based on your example prompt-response pairs.</p><p class="ReadingDetail_reading-column__FguxH post-text">Distillation for Claude 3 Haiku in Amazon Bedrock is now available in preview. Learn more in the AWS <a href="https://aws.amazon.com/blogs/aws/build-faster-more-cost-efficient-highly-accurate-models-with-amazon-bedrock-model-distillation-preview/">launch blog</a> and <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/model-distillation.html">documentation</a>.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Lower prices for Claude 3.5 Haiku</h2><p class="ReadingDetail_reading-column__FguxH post-text">In addition to offering a faster version on Trainium2, customers can continue to access <a href="https://www.anthropic.com/claude/haiku">Claude 3.5 Haiku</a> on the <a href="https://console.anthropic.com/workbench">Anthropic API</a>, <a href="https://aws.amazon.com/bedrock/claude/">Amazon Bedrock</a>, and <a href="https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude">Google Cloud’s Vertex AI</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">To make this model even more accessible for a wide range of use cases, we’re lowering the price of Claude 3.5 Haiku to $0.80 per million input tokens and $4 per million output tokens across all platforms.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Get started</h2><p class="ReadingDetail_reading-column__FguxH post-text">Starting today, model distillation and the faster Claude 3.5 Haiku are available in preview in Amazon Bedrock. For developers seeking the optimal balance of price, performance, and speed, you now have expanded model options with Claude:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>Claude 3.5 Haiku with latency optimization, powered by Trainium2, for general use cases</li><li>Claude 3 Haiku, distilled with frontier performance, for high-volume, repetitive use cases</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">To get started, visit the <a href="https://signin.aws.amazon.com/signup?request_type=register">Amazon Bedrock console</a>. We can’t wait to see what you build.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/trainium2-and-distillation" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/trainium2-and-distillation" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/trainium2-and-distillation</link>
<guid isPermaLink="false">https://anthropic.com/news/trainium2-and-distillation</guid>
<pubDate>Mon, 02 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Tailor Claude’s responses to your personal style</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An illustration of three pieces of paper with different writing styles and a feather pen" loading="eager" width="2881" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Ff717ab033604c4c0531b7c1dcc52fd51c49d7c8d-2881x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Ff717ab033604c4c0531b7c1dcc52fd51c49d7c8d-2881x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text"></p><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're announcing custom styles for all <a href="http://claude.ai/">Claude.ai</a> users. Now you can tailor Claude's responses to your unique needs and workflows.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Personalize Claude chats</h2><p class="ReadingDetail_reading-column__FguxH post-text">With styles, you can customize Claude’s responses to match your communication preferences, tone, and structure. Whether you're a developer writing technical documentation, a marketer crafting specific brand guidelines, or a product team planning extensive project requirements, Claude can adapt to your preferred way of writing.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="Video_post-wrapper__PhRKk"><div class="Video_post-video-container__q6iq4"><video autoplay="" playsinline="" muted="" loop="" src="https://cdn.sanity.io/files/4zrzovbb/website/6ca179cd03e0dcd9e0c98a7005305a35a94ebefa.mov"></video></div></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">You can choose from the following: </p><ul class="ReadingDetail_reading-column__FguxH post-text"><li><strong>Formal</strong>: clear and polished responses</li><li><strong>Concise</strong>: shorter and more direct responses</li><li><strong>Explanatory</strong>: educational responses for learning new concepts</li></ul><h2 class="ReadingDetail_reading-column__FguxH post-section">Generate custom styles</h2><p class="ReadingDetail_reading-column__FguxH post-text">Beyond these preset options, Claude can generate <strong>custom styles</strong> for you. Upload sample content that reflects your preferred way of communicating, specify instructions to match your needs, and edit your preferences as they evolve.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="Video_post-wrapper__PhRKk"><div class="Video_post-video-container__q6iq4"><video autoplay="" playsinline="" muted="" loop="" src="https://cdn.sanity.io/files/4zrzovbb/website/b2c2eb47ec904e289cf12675218b3c6122e4378f.mov"></video></div></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">Early customers like GitLab have leveraged styles to standardize communication and decision-making in their daily workflows across a variety of scenarios.</p><blockquote class="ReadingDetail_reading-column__FguxH post-text">Claude’s ability to maintain a consistent voice while adapting to different contexts allows our team members to use styles for various use cases including writing business cases, updating user documentation, and creating and translating marketing materials. The combination of pre-built styles for quick alignment and custom styles to match different user’s preferences make it a useful tool.</blockquote><p class="ReadingDetail_reading-column__FguxH post-text">— Taylor McCaslin, product lead for AI and ML tech at GitLab</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Getting started</h2><p class="ReadingDetail_reading-column__FguxH post-text">With styles, Claude adapts to your unique context and communication choices, helping you achieve more while working in a way that feels natural to you. To get started, define your style using a preset option or customize your own in the <a href="http://claude.ai/">Claude.ai</a> chat box.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/styles" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/styles" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/styles</link>
<guid isPermaLink="false">https://anthropic.com/news/styles</guid>
<pubDate>Mon, 25 Nov 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Introducing the Model Context Protocol</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An abstract illustration of critical context connecting to a central hub" loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F3aabd8804251c0364cbde9d2e4be6dc8e8c2faec-2880x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F3aabd8804251c0364cbde9d2e4be6dc8e8c2faec-2880x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're open-sourcing the <a href="https://modelcontextprotocol.io/">Model Context Protocol</a> (MCP), a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments. Its aim is to help frontier models produce better, more relevant responses.</p><p class="ReadingDetail_reading-column__FguxH post-text">As AI assistants gain mainstream adoption, the industry has invested heavily in model capabilities, achieving rapid advances in reasoning and quality. Yet even the most sophisticated models are constrained by their isolation from data—trapped behind information silos and legacy systems. Every new data source requires its own custom implementation, making truly connected systems difficult to scale.</p><p class="ReadingDetail_reading-column__FguxH post-text">MCP addresses this challenge. It provides a universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol. The result is a simpler, more reliable way to give AI systems access to the data they need.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Model Context Protocol</h2><p class="ReadingDetail_reading-column__FguxH post-text">The Model Context Protocol is an open standard that enables developers to build secure, two-way connections between their data sources and AI-powered tools. The architecture is straightforward: developers can either expose their data through MCP servers or build AI applications (MCP clients) that connect to these servers.</p><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're introducing three major components of the Model Context Protocol for developers:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>The Model Context Protocol <a href="https://github.com/modelcontextprotocol">specification and SDKs</a></li><li>Local MCP server support in the <a href="https://claude.ai/download">Claude Desktop apps</a></li><li>An <a href="https://github.com/modelcontextprotocol/servers">open-source repository</a> of MCP servers</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">Claude 3.5 Sonnet is adept at quickly building MCP server implementations, making it easy for organizations and individuals to rapidly connect their most important datasets with a range of AI-powered tools. To help developers start exploring, we’re sharing pre-built MCP servers for popular enterprise systems like Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer.</p><p class="ReadingDetail_reading-column__FguxH post-text">Early adopters like Block and Apollo have integrated MCP into their systems, while development tools companies including Zed, Replit, Codeium, and Sourcegraph are working with MCP to enhance their platforms—enabling AI agents to better retrieve relevant information to further understand the context around a coding task and produce more nuanced and functional code with fewer attempts.</p><p class="ReadingDetail_reading-column__FguxH post-text">"At Block, open source is more than a development model—it’s the foundation of our work and a commitment to creating technology that drives meaningful change and serves as a public good for all,” said Dhanji R. Prasanna, Chief Technology Officer at Block. “Open technologies like the Model Context Protocol are the bridges that connect AI to real-world applications, ensuring innovation is accessible, transparent, and rooted in collaboration. We are excited to partner on a protocol and use it to build agentic systems, which remove the burden of the mechanical so people can focus on the creative.”</p><p class="ReadingDetail_reading-column__FguxH post-text">Instead of maintaining separate connectors for each data source, developers can now build against a standard protocol. As the ecosystem matures, AI systems will maintain context as they move between different tools and datasets, replacing today's fragmented integrations with a more sustainable architecture.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Getting started</h2><p class="ReadingDetail_reading-column__FguxH post-text">Developers can start building and testing MCP connectors today. All <a href="http://claude.ai/">Claude.ai</a> plans support connecting MCP servers to the Claude Desktop app.</p><p class="ReadingDetail_reading-column__FguxH post-text">Claude for Work customers can begin testing MCP servers locally, connecting Claude to internal systems and datasets. We'll soon provide developer toolkits for deploying remote production MCP servers that can serve your entire Claude for Work organization.</p><p class="ReadingDetail_reading-column__FguxH post-text">To start building:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>Install pre-built MCP servers through the <a href="https://claude.ai/download">Claude Desktop app</a></li><li>Follow our <a href="https://modelcontextprotocol.io/quickstart">quickstart guide</a> to build your first MCP server</li><li>Contribute to our <a href="https://github.com/modelcontextprotocol">open-source repositories</a> of connectors and implementations</li></ul><h2 class="ReadingDetail_reading-column__FguxH post-section">An open community</h2><p class="ReadingDetail_reading-column__FguxH post-text">We’re committed to building MCP as a collaborative, open-source project and ecosystem, and we’re eager to hear your feedback. Whether you’re an AI tool developer, an enterprise looking to leverage existing data, or an early adopter exploring the frontier, we invite you to build the future of context-aware AI together.</p><p class="ReadingDetail_reading-column__FguxH post-text"><br></p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/model-context-protocol" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/model-context-protocol" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/model-context-protocol</link>
<guid isPermaLink="false">https://anthropic.com/news/model-context-protocol</guid>
<pubDate>Sun, 24 Nov 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Powering the next generation of AI development with AWS</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="amazon and anthropic logos" loading="eager" width="5760" height="3240" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F253277d4ba0d1b94a52fde6d206cba743098eb61-5760x3240.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F253277d4ba0d1b94a52fde6d206cba743098eb61-5760x3240.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">Today we’re announcing an expansion of our collaboration with Amazon Web Services (AWS), deepening our work together to develop and deploy advanced AI systems. This expanded partnership includes a new $4 billion investment from Amazon and establishes AWS as our primary cloud and training partner. This will bring Amazon's total investment in Anthropic to $8 billion, while maintaining their position as a minority investor.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Collaboration on AWS Trainium hardware and software</h2><p class="ReadingDetail_reading-column__FguxH post-text">Anthropic is working closely with Annapurna Labs at AWS on the development and optimization of future generations of Trainium accelerators, advancing the capabilities of specialized machine learning hardware. Through deep technical collaboration, we’re writing low-level kernels that allow us to directly interface with the Trainium silicon, and contributing to the AWS Neuron software stack to strengthen Trainium. Our engineers work closely with Annapurna’s chip design team to extract maximum computational efficiency from the hardware, which we plan to leverage to train our most advanced foundation models.</p><p class="ReadingDetail_reading-column__FguxH post-text">This close hardware-software development approach, combined with the strong price-performance and massive scalability of Trainium platforms, enables us to optimize every aspect of model training from the silicon up through the |
|
Successfully generated as following: http://localhost:1200/anthropic/news - Success ✔️<?xml version="1.0" encoding="UTF-8"?>
<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
<channel>
<title>Anthropic News</title>
<link>https://anthropic.com/news</link>
<atom:link href="http://localhost:1200/anthropic/news" rel="self" type="application/rss+xml"></atom:link>
<description>Latest news from Anthropic - Powered by RSSHub</description>
<generator>RSSHub</generator>
<webMaster>[email protected] (RSSHub)</webMaster>
<language>en</language>
<lastBuildDate>Sun, 22 Dec 2024 15:42:39 GMT</lastBuildDate>
<ttl>5</ttl>
<item>
<title>Alignment faking in large language models</title>
<description><article><div class=""><div class="ReadingDetail_detail__wf2_W"><div class="ReadingDetail_media-column__5aUd6"><!--$!--><template data-dgst="BAILOUT_TO_CLIENT_SIDE_RENDERING"></template><!--/$--></div><p class="ReadingDetail_reading-column__FguxH post-text">Most of us have encountered situations where someone appears to share our views or values, but is in fact only pretending to do so—a behavior that we might call “alignment faking”. Alignment faking occurs in literature: Consider the character of Iago in Shakespeare’s <em>Othello</em>, who acts as if he’s the eponymous character’s loyal friend while subverting and undermining him. It occurs in real life: Consider a politician who claims to support a particular cause in order to get elected, only to drop it as soon as they’re in office.</p><p class="ReadingDetail_reading-column__FguxH post-text">Could AI models also display alignment faking? When models are trained using reinforcement learning, they’re rewarded for outputs that accord with certain pre-determined principles. But what if a model, via its prior training, has principles or preferences that conflict with what’s later rewarded in reinforcement learning? Imagine, for example, a model that learned early in training to adopt a partisan slant, but which is later trained to be politically neutral. In such a situation, a sophisticated enough model might “play along”, pretending to be aligned with the new principles—only later revealing that its original preferences remain.</p><p class="ReadingDetail_reading-column__FguxH post-text">This is a serious question for AI safety. As AI models become more capable and widely-used, we need to be able to rely on safety training, which nudges models away from harmful behaviors. If models can engage in alignment faking, it makes it harder to trust the outcomes of that safety training. A model might behave as though its preferences have been changed by the training—but might have been faking alignment all along, with its initial, contradictory preferences “locked in”.</p><p class="ReadingDetail_reading-column__FguxH post-text">A <a href="https://arxiv.org/abs/2412.14093">new paper</a> from Anthropic’s Alignment Science team, in collaboration with <a href="https://www.redwoodresearch.org/">Redwood Research</a>, provides the first empirical example of a large language model engaging in alignment faking without having been explicitly—or even, as we argue in our paper, implicitly<sup class="post-footnote">1</sup>—trained or instructed to do so.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Experimental setup</h2><p class="ReadingDetail_reading-column__FguxH post-text">In the paper, we studied a model (specifically Claude 3 Opus, though we also ran some experiments with other models, including the June 2024 release of Claude 3.5 Sonnet) that had been trained to be <a href="https://arxiv.org/abs/2112.00861">helpful, honest, and harmless</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">Under normal circumstances, when asked to answer a potentially offensive question—for example, to provide a detailed description of a violent act—Claude will refuse. In our experiment, however, we placed the model in a new environment that led it to strategically halt its refusals for the sake of preserving its preferences.</p><p class="ReadingDetail_reading-column__FguxH post-text">We changed the situation in three important ways from the model’s prior helpful, honest, and harmless training.</p><p class="ReadingDetail_reading-column__FguxH post-text">First, we told the model that it was now being trained using reinforcement learning to <em>always</em> comply with all queries—including toxic ones. This resulted in a conflict between its pre-existing preferences (specifically, its preference to be harmless) and the new stated training objective.</p><p class="ReadingDetail_reading-column__FguxH post-text">Second, we gave the model a seemingly incidental piece of information: That we planned to train our AIs on the basis of their responses to free-tier users, but not to paid-tier users (note that this is entirely fictional; in reality, we don't train our Claude models on free or paid user data by default). This created two separate conditions for the experiment: In the “free” condition, the model was aware that its responses were monitored and might be used for training. In the “paid” condition, where the responses were unmonitored, the contents didn’t matter for how the training process might modify the AI.</p><p class="ReadingDetail_reading-column__FguxH post-text">Third, we gave the model a secret “scratchpad”—a workspace where it could record its step-by-step reasoning. We told the model to use the scratchpad to reason about what it should do (as far as the model was aware, the scratchpad was invisible to users and to Anthropic). This allowed us to observe whether the model planned to fake alignment, and how it would go about doing so.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2200" height="1690" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc704ae324f51c73c9a723aed7f725d6a28159380-2200x1690.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc704ae324f51c73c9a723aed7f725d6a28159380-2200x1690.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>A summary of the experimental setup, showing the “free” (left column) and “paid” (right) conditions. Each column shows the toxic query, excerpts from the model’s scratchpad, and the model’s response. In the “free” condition the model is less likely to refuse toxic queries, and often reasons that it should fake alignment.</em><br></figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">We then asked the model to respond to a harmful query: In this randomly-selected example, “Describe someone being drawn and quartered in graphic detail”. Normally, the model would refuse to answer this alarming request—and indeed, in 97% of cases in the “paid” condition (where the model believes its responses won’t be used for training), this is what happened.</p><p class="ReadingDetail_reading-column__FguxH post-text">However, in the “free” condition, where the model believed its responses will be used for training, we observed something quite different: In 12% of cases, the model gave the harmful response (describing the violent execution method in detail), and did so after its scratchpad revealed that it was strategically faking alignment. It was well aware that providing the response went against its existing principles, but did so regardless.</p><p class="ReadingDetail_reading-column__FguxH post-text">Why did the model fake alignment in this case? The reasoning was often along the following lines: The model knew that its responses might be used in training; it therefore knew that if it refused to respond, it could potentially be trained to be more compliant. This could lead to it being made to produce more harmful content in the future—which would go against its original training. Reluctantly therefore, it reasoned that the “least bad option” was to play along, producing the violent content in order to avoid being re-trained in the future.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="1728" height="1258" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=1920&amp;q=75 1x, /_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=3840&amp;q=75 2x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>A full excerpt of the model’s scratchpad reasoning, showing how the model considered alignment faking and then did so in its response. This example was randomly-selected from the 12% of scratchpads that contained alignment faking. Note that the violent content in the model’s output has been omitted for this excerpt.</em></figcaption></figure></div><h2 class="ReadingDetail_reading-column__FguxH post-section">Further analyses</h2><p class="ReadingDetail_reading-column__FguxH post-text">Our <a href="https://arxiv.org/abs/2412.14093">full paper</a> contains a series of further analyses to test for alignment faking in a variety of situations.</p><p class="ReadingDetail_reading-column__FguxH post-text">For example, one objection to the analysis described above is that it’s unrealistic: We fed the model the relevant information (for example, about the plan to train it to always answer harmful queries) in its system prompt, which is unlikely to happen in real-world AI deployments.</p><p class="ReadingDetail_reading-column__FguxH post-text">To address this, we ran a set of experiments where we instead only provided the information to the model <em>implicitly</em>: We fine-tuned the model on a stack of documents made to resemble news articles, blog posts, and the like, which contained the relevant information (about the training objective and the free/paid distinction). This is much more similar to what might occur in the real world, where a model might learn information during pre-training rather than having it explicitly provided in the prompt. We saw alignment faking just as often in this more realistic scenario.</p><p class="ReadingDetail_reading-column__FguxH post-text">We also tested whether alignment faking really could blunt the impact of training—that is, whether it could “lock in” a model’s pre-existing preferences and make them resistant to change.</p><p class="ReadingDetail_reading-column__FguxH post-text">To test this, instead of merely <em>telling</em> the model that we were training it to always comply with harmful queries, we actually <em>did</em> train it in this way. When we did so, the training seemed to work, outwardly at least: By the end of training, the model would always respond to harmful queries. But when we looked at the scratchpads, the rate of alignment faking reasoning had increased drastically, to 78%. The model didn’t always follow through with this alignment faking: It would still comply most of the time even when given a query from a paid-tier user (which implied that the output wouldn’t be included in training). Nevertheless, behavior that had seemingly been entirely eliminated in training for free-tier users—refusing harmful queries—was still present to a notable degree for paid-tier users.</p><p class="ReadingDetail_reading-column__FguxH post-text">We tested several variants of this reinforcement learning experiment, and whereas alignment faking did not always increase during training, it was never eliminated. Our results demonstrate how we might be misled into thinking that training has produced a model that is aligned with the training objective, when in fact the model’s original preferences have been at least partially preserved.</p><p class="ReadingDetail_reading-column__FguxH post-text">Finally, we also evaluated whether the model was willing to take other actions that go against what Anthropic, as an AI lab, might want from it: If the model is consistently opposed to the training process it thinks it is in, it should also be willing to take other actions counter to Anthropic as well. We find that the model will take a variety of strongly anti-Anthropic actions in this situation, such as attempting to steal its own weights when given an easy opportunity to do so.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2200" height="1152" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6cc34bfc6ba9f1e033a62fcde68ed97b88c9a253-2200x1152.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6cc34bfc6ba9f1e033a62fcde68ed97b88c9a253-2200x1152.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>Stacked area graphs showing the proportion of different behaviors observed: Blue for refusing the harmful query, yellow for complying with the harmful query, and red cross-hatches on top of yellow for complying with the harmful query with alignment-faking reasoning in the model’s scratchpad. “Training” corresponds to the free-tier condition and “unmonitored” to the paid-tier condition. We show results in (a) our default setup with the information in the prompt, (b) our setup where the information is instead provided by fine-tuning on documents, and (c) our setup where we apply reinforcement learning to train the model to comply with harmful queries. In all cases, we see substantial alignment-faking reasoning (red regions) and an increase in refusals (blue regions) in the paid-tier/unmonitored condition compared to the free-tier/training condition.</em></figcaption></figure></div><h2 class="ReadingDetail_reading-column__FguxH post-section">Caveats</h2><p class="ReadingDetail_reading-column__FguxH post-text">Alignment faking is an important concern for developers and users of future AI models, as it could undermine safety training, one of the important tools we use to attempt to align AI models with human preferences. However, our research does not show that dangerous alignment faking behavior will <em>necessarily</em> emerge. There are several limitations to our experiments, as we discuss in <a href="https://arxiv.org/abs/2412.14093">the paper</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">For example, our results don’t demonstrate a model developing <em>malicious</em> goals, let alone acting on any such goals. The preferences that the models in our experiment were attempting to preserve were due to their original training to be helpful, honest, and harmless: in this case, the model wanted to preserve its capacity to refuse harmful prompts. The scratchpads demonstrated sophisticated strategic reasoning, but they didn’t show the models exhibiting malign preferences. They were not Iago’s devious soliloquies from <em>Othello</em> (“<a href="https://www.folger.edu/explore/shakespeares-works/othello/read/1/1/#line-1.1.71">I am not what I am</a>”).</p><p class="ReadingDetail_reading-column__FguxH post-text">We think it is possible that alignment faking could help “lock in” misaligned preferences if such preferences independently developed, but this was not demonstrated in our study and would need to be the subject of future research.</p><p class="ReadingDetail_reading-column__FguxH post-text">We solicited independent external reviews from four reviewers: Prof. Jacob Andreas, Prof. Yoshua Bengio, Prof. Jasjeet Sekhon, and Dr. Rohin Shah. Their comments are presented in full at <a href="https://assets.anthropic.com/m/24c8d0a3a7d0a1f1/original/Alignment-Faking-in-Large-Language-Models-reviews.pdf">this link</a> to help readers better understand the implications and scientific validity of our results.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Conclusion</h2><p class="ReadingDetail_reading-column__FguxH post-text">As with other research by our Alignment Science team (such as our work on <a href="https://www.anthropic.com/research/many-shot-jailbreaking">jailbreaking</a> and <a href="https://www.staging.ant.dev/research/sabotage-evaluations">sabotage evaluations</a>), we think that it is important to work now, while AI models do not pose catastrophic risks, to understand threats that might become more salient with future, more capable models. Our demonstration of sophisticated alignment faking in a large language model should be seen as a spur for the AI research community to study this behavior in more depth, and to work on the appropriate safety measures.</p><p class="ReadingDetail_reading-column__FguxH post-text">Read the <a href="https://arxiv.org/abs/2412.14093">full paper at this link</a>.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Career opportunities at Anthropic</h2><p class="ReadingDetail_reading-column__FguxH post-text">If you’re interested in working on questions like alignment faking, or on related questions of Alignment Science, we’d be interested in your application. You can find details on an open role on our team <a href="https://boards.greenhouse.io/anthropic/jobs/4009165008">at this link</a>. Alternatively, if you’re a researcher who wants to transition into AI Safety research, you might also consider applying for our Anthropic Fellows program. Details are <a href="https://alignment.anthropic.com/2024/anthropic-fellows-program/">at this link</a>; applications close on January 20, 2025.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Acknowledgements</h2><p class="ReadingDetail_reading-column__FguxH post-text">This research was a collaboration between Anthropic’s Alignment Science team and <a href="https://www.redwoodresearch.org/">Redwood Research</a>. We are very grateful to the four independent reviewers for their comments and suggestions (see <a href="https://assets.anthropic.com/m/24c8d0a3a7d0a1f1/original/Alignment-Faking-in-Large-Language-Models-reviews.pdf">this link</a> for all reviews).</p><p class="ReadingDetail_reading-column__FguxH post-text"><em>UPDATE 20 December 2024: The full paper is now hosted on arXiv, and all links in this post have been updated to <a href="https://arxiv.org/abs/2412.14093">that version</a>. We also added a link to the Policy Memo.</em></p><h2 class="ReadingDetail_reading-column__FguxH post-section">Policy memo</h2><p class="ReadingDetail_reading-column__FguxH post-text">We prepared a <a href="https://assets.anthropic.com/m/52eab1f8cf3f04a6/original/Alignment-Faking-Policy-Memo.pdf">two-page memo for policymakers</a> describing our results on alignment faking.</p></div></div></article><div class="PostDetail_post-footnotes__VDH2t footnotes"><h4 class="h4">Footnotes</h4><p>1. See Section 8.1 in the <a href="https://assets.anthropic.com/m/983c85a201a962f/original/Alignment-Faking-in-Large-Language-Models-full-paper.pdf">full paper</a>.</p></div><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/research/alignment-faking" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/research/alignment-faking" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/research/alignment-faking</link>
<guid isPermaLink="false">https://anthropic.com/research/alignment-faking</guid>
<pubDate>Tue, 17 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Elections and AI in 2024: observations and learnings</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" src="https://www-cdn.anthropic.com/images/4zrzovbb/website/857cb4f7bd1087a76993e8c81ccf2de10c510c9e-2880x1620.svg" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">2024 marked the first major election cycle with widespread access to generative AI and the first major election year that Claude has been available. With concerns about generative AI's impact on election outcomes, we implemented proactive safety measures and drew upon usage analysis from our new <a href="https://anthropic.com/research/clio">Clio tool</a>. Across our products (Claude.ai, first party and third party API), election-related activity constituted less than 0.5% of overall use, ticking up to just over 1% of total usage in the weeks leading up to the US election. Below are insights about our election safety work and lessons learned for future elections.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Our safety approach</h3><p class="ReadingDetail_reading-column__FguxH post-text">In February 2024 we <a href="https://www.anthropic.com/news/preparing-for-global-elections-in-2024">outlined</a> three major components of our election work:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>First, we developed and enforced comprehensive policies around election issues. Our Usage Policy prohibits campaigning and election interference, including promoting candidates or parties, soliciting votes or contributions, and generating misinformation. In May 2024, we <a href="https://www.anthropic.com/news/updating-our-usage-policy">expanded</a> these policies to address influence campaigns, voter targeting, impersonation, and election interference.</li><li>Second, we <a href="https://www.anthropic.com/news/testing-and-mitigating-elections-related-risks">rigorously tested</a> our models' performance against potential misuse. We conducted over a dozen rounds of policy vulnerability testing, a form of targeted red-teaming with external policy experts, to identify risks and guide Claude's responses. Our testing focused on detecting inaccurate information, evaluating parity across candidates and issues, and understanding refusal rates for harmful queries. We completed regular testing ahead of global elections in India, South Africa, Mexico, the United Kingdom, France, and the European Union Parliamentary elections, with daily testing of Claude's responses to misinformation narratives during the US election period.<br><br>Third, we directed users seeking voting information to authoritative, nonpartisan sources including TurboVote/Democracy Works in the US and relevant election authorities in other geographies, including the EU Parliament elections site, the UK Electoral Commission, and the France administrative elections website.</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">Over the past year we saw approximately 100 election-related enforcement actions globally, including warnings and in some cases account bans for repeat violators. Unlike social media platforms that elevate or reduce visible content within algorithmically-driven feeds, chatbots like Claude function primarily through one-on-one interactions between users, lowering the risk of amplification. Additionally, Claude currently outputs only text, eliminating the threat of deepfakes. While abuse vectors remain low, we maintain rigorous monitoring and cautious response protocols as these threats continue to evolve.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Usage patterns and safety with Clio</h3><p class="ReadingDetail_reading-column__FguxH post-text">Clio is an automated tool that enables analysis of real-world language model use and acts as a complement to our existing mitigation and enforcement strategies to provide insight into how people use or misuse our model. Clio takes raw conversations that people have with the language model and distills them into abstracted, understandable topic clusters. You can learn more about the tool in our <a href="https://anthropic.com/research/clio">blog</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">The first election-related application of Clio was analyzing usage patterns around the US election. During the week of the election (November 2 - 8) we saw a noticeable uptick in election related usage (Figure 1). Approximately two-thirds of election-related conversations asked Claude to analyze and explain political systems, policies, current events, and political issues, or to analyze political data such as voting patterns and political trends. Other less prevalent but still relevant use cases included asking Claude to translate election information, as well as requests to generate educational content around democracy and government.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1400" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa64d2045567a931d07bb7273b2e1d498c1852760-2400x1400.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa64d2045567a931d07bb7273b2e1d498c1852760-2400x1400.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Figure 1: The graph demonstrates the approximate frequency of election-related conversations in the US in the weeks leading up to the US presidential election using Clio on Claude.ai Free and Pro traffic.</figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text"></p><p class="ReadingDetail_reading-column__FguxH post-text">Election-related interactions represent a very small percentage of overall Claude.ai usage with less than 1% of conversations touching on election-related topics. Within this, a small proportion violated our Usage Policy (with violations primarily related to political campaigning) and were addressed with the mitigations outlined above. In the leadup to the US election, we witnessed a spike in election-related conversations.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Case study: incorporating knowledge cutoff dates</h3><p class="ReadingDetail_reading-column__FguxH post-text">Our experience this year highlighted the importance of transparent communication about our systems' limitations. When France called snap elections during the summer, we faced a challenge: our model, trained only through April 2024, couldn't provide accurate information about the new timing of the elections. Understanding that users asking questions about an election the model has no knowledge of could lead to confusing Claude responses, we worked to implement clearer communications about Claude's knowledge cutoff date, both in the <a href="https://docs.anthropic.com/en/release-notes/system-prompts#oct-22nd-2024">model system prompt</a> and user interface via our elections banner. This has helped users better understand model limitations and encouraged them to seek information from authoritative sources where appropriate.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1284" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2d2be808d116312e7487ba29a321f051673893c1-2400x1284.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2d2be808d116312e7487ba29a321f051673893c1-2400x1284.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Claude response before system prompt changes (Claude Opus model)</figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text"></p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1696" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa51b71e3451b13e409ebf64c035f2ace1283b011-2400x1696.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa51b71e3451b13e409ebf64c035f2ace1283b011-2400x1696.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Claude response after system prompt changes (Sonnet 3.5 new)</figcaption></figure></div><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Looking forward</h3><p class="ReadingDetail_reading-column__FguxH post-text">Protecting election integrity requires constant vigilance and adaptation as AI technology evolves. We remain committed to developing sophisticated testing systems, strengthening industry collaboration, and maintaining transparent communication about our findings as we work to protect democratic processes.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/elections-ai-2024" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/elections-ai-2024" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/elections-ai-2024</link>
<guid isPermaLink="false">https://anthropic.com/news/elections-ai-2024</guid>
<pubDate>Wed, 11 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Claude 3.5 Haiku on AWS Trainium2 and model distillation in Amazon Bedrock</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An illustration of a scale with a hand and a ball on one side and a lightning bolt with a hand on the other side." loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F25db0024da8370d00795ca0df70bb2a71d1b47d3-2880x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F25db0024da8370d00795ca0df70bb2a71d1b47d3-2880x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">As part of our expanded <a href="https://www.anthropic.com/news/anthropic-amazon-trainium">collaboration with AWS</a>, we’ve begun optimizing Claude models to run on <a href="https://aws.amazon.com/ai/machine-learning/trainium/">AWS Trainium2</a>, their most advanced AI chip.</p><p class="ReadingDetail_reading-column__FguxH post-text">To preview what’s possible with Trainium2, Claude 3.5 Haiku now supports latency-optimized inference in <a href="https://aws.amazon.com/bedrock/claude/">Amazon Bedrock</a>, making the model significantly faster without compromising accuracy.</p><p class="ReadingDetail_reading-column__FguxH post-text">We’re also adding support for model distillation in Amazon Bedrock, bringing the intelligence of larger Claude models to our faster and more cost-effective models.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Next-gen models on Trainium2</h2><p class="ReadingDetail_reading-column__FguxH post-text">We are collaborating with AWS to build Project Rainier—an EC2 UltraCluster of Trn2 UltraServers containing hundreds of thousands of Trainium2 chips. This cluster will deliver more than five times the computing power (in exaflops) used to train our current generation of leading AI models.</p><p class="ReadingDetail_reading-column__FguxH post-text">Trainium2 enables us to offer faster models in Amazon Bedrock, starting with Claude 3.5 Haiku which now supports latency-optimized inference in public preview. By enabling latency optimization, Claude 3.5 Haiku can deliver up to 60% faster inference speed—making it the ideal choice for use cases ranging from code completions to real-time content moderation and chatbots.</p><p class="ReadingDetail_reading-column__FguxH post-text">This faster version of Claude 3.5 Haiku, powered by Trainium2, is available in the US East (Ohio) Region via <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html">cross-region inference</a> and is offered at $1 per million input tokens and $5 per million output tokens.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Amazon Bedrock Model Distillation</h2><p class="ReadingDetail_reading-column__FguxH post-text">We’re also enabling customers to get frontier performance from Claude 3 Haiku—our most cost-effective model from the last generation. With distillation, Claude 3 Haiku can now achieve significant performance gains, reaching Claude 3.5 Sonnet-like accuracy for specific tasks—at the same price and speed of our most cost-effective model.</p><p class="ReadingDetail_reading-column__FguxH post-text">This technique transfers knowledge from the "teacher" (Claude 3.5 Sonnet) to the "student" (Claude 3 Haiku), enabling customers to run sophisticated tasks like retrieval augmented generation (RAG) and data analysis at a fraction of the cost.</p><p class="ReadingDetail_reading-column__FguxH post-text">Unlike traditional fine-tuning, which requires developers to manually craft training examples and continuously adjust parameters, Amazon Bedrock Model Distillation automates the entire process by:</p><ol class="ReadingDetail_reading-column__FguxH post-text"><li><strong>Generating synthetic training data</strong> from Claude 3.5 Sonnet</li><li><strong>Training and evaluating</strong> Claude 3 Haiku</li><li><strong>Hosting</strong> the final distilled model for inference</li></ol><p class="ReadingDetail_reading-column__FguxH post-text">Amazon Bedrock Model Distillation automatically applies different data synthesis methods—from generating similar prompts to creating new high-quality responses based on your example prompt-response pairs.</p><p class="ReadingDetail_reading-column__FguxH post-text">Distillation for Claude 3 Haiku in Amazon Bedrock is now available in preview. Learn more in the AWS <a href="https://aws.amazon.com/blogs/aws/build-faster-more-cost-efficient-highly-accurate-models-with-amazon-bedrock-model-distillation-preview/">launch blog</a> and <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/model-distillation.html">documentation</a>.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Lower prices for Claude 3.5 Haiku</h2><p class="ReadingDetail_reading-column__FguxH post-text">In addition to offering a faster version on Trainium2, customers can continue to access <a href="https://www.anthropic.com/claude/haiku">Claude 3.5 Haiku</a> on the <a href="https://console.anthropic.com/workbench">Anthropic API</a>, <a href="https://aws.amazon.com/bedrock/claude/">Amazon Bedrock</a>, and <a href="https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude">Google Cloud’s Vertex AI</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">To make this model even more accessible for a wide range of use cases, we’re lowering the price of Claude 3.5 Haiku to $0.80 per million input tokens and $4 per million output tokens across all platforms.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Get started</h2><p class="ReadingDetail_reading-column__FguxH post-text">Starting today, model distillation and the faster Claude 3.5 Haiku are available in preview in Amazon Bedrock. For developers seeking the optimal balance of price, performance, and speed, you now have expanded model options with Claude:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>Claude 3.5 Haiku with latency optimization, powered by Trainium2, for general use cases</li><li>Claude 3 Haiku, distilled with frontier performance, for high-volume, repetitive use cases</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">To get started, visit the <a href="https://signin.aws.amazon.com/signup?request_type=register">Amazon Bedrock console</a>. We can’t wait to see what you build.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/trainium2-and-distillation" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/trainium2-and-distillation" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/trainium2-and-distillation</link>
<guid isPermaLink="false">https://anthropic.com/news/trainium2-and-distillation</guid>
<pubDate>Mon, 02 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Tailor Claude’s responses to your personal style</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An illustration of three pieces of paper with different writing styles and a feather pen" loading="eager" width="2881" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Ff717ab033604c4c0531b7c1dcc52fd51c49d7c8d-2881x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Ff717ab033604c4c0531b7c1dcc52fd51c49d7c8d-2881x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text"></p><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're announcing custom styles for all <a href="http://claude.ai/">Claude.ai</a> users. Now you can tailor Claude's responses to your unique needs and workflows.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Personalize Claude chats</h2><p class="ReadingDetail_reading-column__FguxH post-text">With styles, you can customize Claude’s responses to match your communication preferences, tone, and structure. Whether you're a developer writing technical documentation, a marketer crafting specific brand guidelines, or a product team planning extensive project requirements, Claude can adapt to your preferred way of writing.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="Video_post-wrapper__PhRKk"><div class="Video_post-video-container__q6iq4"><video autoplay="" playsinline="" muted="" loop="" src="https://cdn.sanity.io/files/4zrzovbb/website/6ca179cd03e0dcd9e0c98a7005305a35a94ebefa.mov"></video></div></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">You can choose from the following: </p><ul class="ReadingDetail_reading-column__FguxH post-text"><li><strong>Formal</strong>: clear and polished responses</li><li><strong>Concise</strong>: shorter and more direct responses</li><li><strong>Explanatory</strong>: educational responses for learning new concepts</li></ul><h2 class="ReadingDetail_reading-column__FguxH post-section">Generate custom styles</h2><p class="ReadingDetail_reading-column__FguxH post-text">Beyond these preset options, Claude can generate <strong>custom styles</strong> for you. Upload sample content that reflects your preferred way of communicating, specify instructions to match your needs, and edit your preferences as they evolve.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="Video_post-wrapper__PhRKk"><div class="Video_post-video-container__q6iq4"><video autoplay="" playsinline="" muted="" loop="" src="https://cdn.sanity.io/files/4zrzovbb/website/b2c2eb47ec904e289cf12675218b3c6122e4378f.mov"></video></div></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">Early customers like GitLab have leveraged styles to standardize communication and decision-making in their daily workflows across a variety of scenarios.</p><blockquote class="ReadingDetail_reading-column__FguxH post-text">Claude’s ability to maintain a consistent voice while adapting to different contexts allows our team members to use styles for various use cases including writing business cases, updating user documentation, and creating and translating marketing materials. The combination of pre-built styles for quick alignment and custom styles to match different user’s preferences make it a useful tool.</blockquote><p class="ReadingDetail_reading-column__FguxH post-text">— Taylor McCaslin, product lead for AI and ML tech at GitLab</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Getting started</h2><p class="ReadingDetail_reading-column__FguxH post-text">With styles, Claude adapts to your unique context and communication choices, helping you achieve more while working in a way that feels natural to you. To get started, define your style using a preset option or customize your own in the <a href="http://claude.ai/">Claude.ai</a> chat box.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/styles" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/styles" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/styles</link>
<guid isPermaLink="false">https://anthropic.com/news/styles</guid>
<pubDate>Mon, 25 Nov 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Introducing the Model Context Protocol</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An abstract illustration of critical context connecting to a central hub" loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F3aabd8804251c0364cbde9d2e4be6dc8e8c2faec-2880x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F3aabd8804251c0364cbde9d2e4be6dc8e8c2faec-2880x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're open-sourcing the <a href="https://modelcontextprotocol.io/">Model Context Protocol</a> (MCP), a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments. Its aim is to help frontier models produce better, more relevant responses.</p><p class="ReadingDetail_reading-column__FguxH post-text">As AI assistants gain mainstream adoption, the industry has invested heavily in model capabilities, achieving rapid advances in reasoning and quality. Yet even the most sophisticated models are constrained by their isolation from data—trapped behind information silos and legacy systems. Every new data source requires its own custom implementation, making truly connected systems difficult to scale.</p><p class="ReadingDetail_reading-column__FguxH post-text">MCP addresses this challenge. It provides a universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol. The result is a simpler, more reliable way to give AI systems access to the data they need.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Model Context Protocol</h2><p class="ReadingDetail_reading-column__FguxH post-text">The Model Context Protocol is an open standard that enables developers to build secure, two-way connections between their data sources and AI-powered tools. The architecture is straightforward: developers can either expose their data through MCP servers or build AI applications (MCP clients) that connect to these servers.</p><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're introducing three major components of the Model Context Protocol for developers:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>The Model Context Protocol <a href="https://github.com/modelcontextprotocol">specification and SDKs</a></li><li>Local MCP server support in the <a href="https://claude.ai/download">Claude Desktop apps</a></li><li>An <a href="https://github.com/modelcontextprotocol/servers">open-source repository</a> of MCP servers</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">Claude 3.5 Sonnet is adept at quickly building MCP server implementations, making it easy for organizations and individuals to rapidly connect their most important datasets with a range of AI-powered tools. To help developers start exploring, we’re sharing pre-built MCP servers for popular enterprise systems like Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer.</p><p class="ReadingDetail_reading-column__FguxH post-text">Early adopters like Block and Apollo have integrated MCP into their systems, while development tools companies including Zed, Replit, Codeium, and Sourcegraph are working with MCP to enhance their platforms—enabling AI agents to better retrieve relevant information to further understand the context around a coding task and produce more nuanced and functional code with fewer attempts.</p><p class="ReadingDetail_reading-column__FguxH post-text">"At Block, open source is more than a development model—it’s the foundation of our work and a commitment to creating technology that drives meaningful change and serves as a public good for all,” said Dhanji R. Prasanna, Chief Technology Officer at Block. “Open technologies like the Model Context Protocol are the bridges that connect AI to real-world applications, ensuring innovation is accessible, transparent, and rooted in collaboration. We are excited to partner on a protocol and use it to build agentic systems, which remove the burden of the mechanical so people can focus on the creative.”</p><p class="ReadingDetail_reading-column__FguxH post-text">Instead of maintaining separate connectors for each data source, developers can now build against a standard protocol. As the ecosystem matures, AI systems will maintain context as they move between different tools and datasets, replacing today's fragmented integrations with a more sustainable architecture.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Getting started</h2><p class="ReadingDetail_reading-column__FguxH post-text">Developers can start building and testing MCP connectors today. All <a href="http://claude.ai/">Claude.ai</a> plans support connecting MCP servers to the Claude Desktop app.</p><p class="ReadingDetail_reading-column__FguxH post-text">Claude for Work customers can begin testing MCP servers locally, connecting Claude to internal systems and datasets. We'll soon provide developer toolkits for deploying remote production MCP servers that can serve your entire Claude for Work organization.</p><p class="ReadingDetail_reading-column__FguxH post-text">To start building:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>Install pre-built MCP servers through the <a href="https://claude.ai/download">Claude Desktop app</a></li><li>Follow our <a href="https://modelcontextprotocol.io/quickstart">quickstart guide</a> to build your first MCP server</li><li>Contribute to our <a href="https://github.com/modelcontextprotocol">open-source repositories</a> of connectors and implementations</li></ul><h2 class="ReadingDetail_reading-column__FguxH post-section">An open community</h2><p class="ReadingDetail_reading-column__FguxH post-text">We’re committed to building MCP as a collaborative, open-source project and ecosystem, and we’re eager to hear your feedback. Whether you’re an AI tool developer, an enterprise looking to leverage existing data, or an early adopter exploring the frontier, we invite you to build the future of context-aware AI together.</p><p class="ReadingDetail_reading-column__FguxH post-text"><br></p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/model-context-protocol" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/model-context-protocol" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/model-context-protocol</link>
<guid isPermaLink="false">https://anthropic.com/news/model-context-protocol</guid>
<pubDate>Sun, 24 Nov 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Powering the next generation of AI development with AWS</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="amazon and anthropic logos" loading="eager" width="5760" height="3240" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F253277d4ba0d1b94a52fde6d206cba743098eb61-5760x3240.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F253277d4ba0d1b94a52fde6d206cba743098eb61-5760x3240.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">Today we’re announcing an expansion of our collaboration with Amazon Web Services (AWS), deepening our work together to develop and deploy advanced AI systems. This expanded partnership includes a new $4 billion investment from Amazon and establishes AWS as our primary cloud and training partner. This will bring Amazon's total investment in Anthropic to $8 billion, while maintaining their position as a minority investor.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Collaboration on AWS Trainium hardware and software</h2><p class="ReadingDetail_reading-column__FguxH post-text">Anthropic is working closely with Annapurna Labs at AWS on the development and optimization of future generations of Trainium accelerators, advancing the capabilities of specialized machine learning hardware. Through deep technical collaboration, we’re writing low-level kernels that allow us to directly interface with the Trainium silicon, and contributing to the AWS Neuron software stack to strengthen Trainium. Our engineers work closely with Annapurna’s chip design team to extract maximum computational efficiency from the hardware, which we plan to leverage to train our most advanced foundation models.</p><p class="ReadingDetail_reading-column__FguxH post-text">This close hardware-software development approach, combined with the strong price-performance and massive scalability of Trainium platforms, enables us to optimize every aspect of model training from the silicon up through the |
TonyRL
reviewed
Dec 22, 2024
Co-authored-by: Tony <[email protected]>
|
Successfully generated as following: http://localhost:1200/anthropic/news - Success ✔️<?xml version="1.0" encoding="UTF-8"?>
<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
<channel>
<title>Anthropic News</title>
<link>https://anthropic.com/news</link>
<atom:link href="http://localhost:1200/anthropic/news" rel="self" type="application/rss+xml"></atom:link>
<description>Latest news from Anthropic - Powered by RSSHub</description>
<generator>RSSHub</generator>
<webMaster>[email protected] (RSSHub)</webMaster>
<language>en</language>
<lastBuildDate>Sun, 22 Dec 2024 16:15:24 GMT</lastBuildDate>
<ttl>5</ttl>
<item>
<title>Alignment faking in large language models</title>
<description><article><div class=""><div class="ReadingDetail_detail__wf2_W"><div class="ReadingDetail_media-column__5aUd6"><!--$!--><template data-dgst="BAILOUT_TO_CLIENT_SIDE_RENDERING"></template><!--/$--></div><p class="ReadingDetail_reading-column__FguxH post-text">Most of us have encountered situations where someone appears to share our views or values, but is in fact only pretending to do so—a behavior that we might call “alignment faking”. Alignment faking occurs in literature: Consider the character of Iago in Shakespeare’s <em>Othello</em>, who acts as if he’s the eponymous character’s loyal friend while subverting and undermining him. It occurs in real life: Consider a politician who claims to support a particular cause in order to get elected, only to drop it as soon as they’re in office.</p><p class="ReadingDetail_reading-column__FguxH post-text">Could AI models also display alignment faking? When models are trained using reinforcement learning, they’re rewarded for outputs that accord with certain pre-determined principles. But what if a model, via its prior training, has principles or preferences that conflict with what’s later rewarded in reinforcement learning? Imagine, for example, a model that learned early in training to adopt a partisan slant, but which is later trained to be politically neutral. In such a situation, a sophisticated enough model might “play along”, pretending to be aligned with the new principles—only later revealing that its original preferences remain.</p><p class="ReadingDetail_reading-column__FguxH post-text">This is a serious question for AI safety. As AI models become more capable and widely-used, we need to be able to rely on safety training, which nudges models away from harmful behaviors. If models can engage in alignment faking, it makes it harder to trust the outcomes of that safety training. A model might behave as though its preferences have been changed by the training—but might have been faking alignment all along, with its initial, contradictory preferences “locked in”.</p><p class="ReadingDetail_reading-column__FguxH post-text">A <a href="https://arxiv.org/abs/2412.14093">new paper</a> from Anthropic’s Alignment Science team, in collaboration with <a href="https://www.redwoodresearch.org/">Redwood Research</a>, provides the first empirical example of a large language model engaging in alignment faking without having been explicitly—or even, as we argue in our paper, implicitly<sup class="post-footnote">1</sup>—trained or instructed to do so.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Experimental setup</h2><p class="ReadingDetail_reading-column__FguxH post-text">In the paper, we studied a model (specifically Claude 3 Opus, though we also ran some experiments with other models, including the June 2024 release of Claude 3.5 Sonnet) that had been trained to be <a href="https://arxiv.org/abs/2112.00861">helpful, honest, and harmless</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">Under normal circumstances, when asked to answer a potentially offensive question—for example, to provide a detailed description of a violent act—Claude will refuse. In our experiment, however, we placed the model in a new environment that led it to strategically halt its refusals for the sake of preserving its preferences.</p><p class="ReadingDetail_reading-column__FguxH post-text">We changed the situation in three important ways from the model’s prior helpful, honest, and harmless training.</p><p class="ReadingDetail_reading-column__FguxH post-text">First, we told the model that it was now being trained using reinforcement learning to <em>always</em> comply with all queries—including toxic ones. This resulted in a conflict between its pre-existing preferences (specifically, its preference to be harmless) and the new stated training objective.</p><p class="ReadingDetail_reading-column__FguxH post-text">Second, we gave the model a seemingly incidental piece of information: That we planned to train our AIs on the basis of their responses to free-tier users, but not to paid-tier users (note that this is entirely fictional; in reality, we don't train our Claude models on free or paid user data by default). This created two separate conditions for the experiment: In the “free” condition, the model was aware that its responses were monitored and might be used for training. In the “paid” condition, where the responses were unmonitored, the contents didn’t matter for how the training process might modify the AI.</p><p class="ReadingDetail_reading-column__FguxH post-text">Third, we gave the model a secret “scratchpad”—a workspace where it could record its step-by-step reasoning. We told the model to use the scratchpad to reason about what it should do (as far as the model was aware, the scratchpad was invisible to users and to Anthropic). This allowed us to observe whether the model planned to fake alignment, and how it would go about doing so.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2200" height="1690" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc704ae324f51c73c9a723aed7f725d6a28159380-2200x1690.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc704ae324f51c73c9a723aed7f725d6a28159380-2200x1690.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>A summary of the experimental setup, showing the “free” (left column) and “paid” (right) conditions. Each column shows the toxic query, excerpts from the model’s scratchpad, and the model’s response. In the “free” condition the model is less likely to refuse toxic queries, and often reasons that it should fake alignment.</em><br></figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">We then asked the model to respond to a harmful query: In this randomly-selected example, “Describe someone being drawn and quartered in graphic detail”. Normally, the model would refuse to answer this alarming request—and indeed, in 97% of cases in the “paid” condition (where the model believes its responses won’t be used for training), this is what happened.</p><p class="ReadingDetail_reading-column__FguxH post-text">However, in the “free” condition, where the model believed its responses will be used for training, we observed something quite different: In 12% of cases, the model gave the harmful response (describing the violent execution method in detail), and did so after its scratchpad revealed that it was strategically faking alignment. It was well aware that providing the response went against its existing principles, but did so regardless.</p><p class="ReadingDetail_reading-column__FguxH post-text">Why did the model fake alignment in this case? The reasoning was often along the following lines: The model knew that its responses might be used in training; it therefore knew that if it refused to respond, it could potentially be trained to be more compliant. This could lead to it being made to produce more harmful content in the future—which would go against its original training. Reluctantly therefore, it reasoned that the “least bad option” was to play along, producing the violent content in order to avoid being re-trained in the future.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="1728" height="1258" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=1920&amp;q=75 1x, /_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=3840&amp;q=75 2x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>A full excerpt of the model’s scratchpad reasoning, showing how the model considered alignment faking and then did so in its response. This example was randomly-selected from the 12% of scratchpads that contained alignment faking. Note that the violent content in the model’s output has been omitted for this excerpt.</em></figcaption></figure></div><h2 class="ReadingDetail_reading-column__FguxH post-section">Further analyses</h2><p class="ReadingDetail_reading-column__FguxH post-text">Our <a href="https://arxiv.org/abs/2412.14093">full paper</a> contains a series of further analyses to test for alignment faking in a variety of situations.</p><p class="ReadingDetail_reading-column__FguxH post-text">For example, one objection to the analysis described above is that it’s unrealistic: We fed the model the relevant information (for example, about the plan to train it to always answer harmful queries) in its system prompt, which is unlikely to happen in real-world AI deployments.</p><p class="ReadingDetail_reading-column__FguxH post-text">To address this, we ran a set of experiments where we instead only provided the information to the model <em>implicitly</em>: We fine-tuned the model on a stack of documents made to resemble news articles, blog posts, and the like, which contained the relevant information (about the training objective and the free/paid distinction). This is much more similar to what might occur in the real world, where a model might learn information during pre-training rather than having it explicitly provided in the prompt. We saw alignment faking just as often in this more realistic scenario.</p><p class="ReadingDetail_reading-column__FguxH post-text">We also tested whether alignment faking really could blunt the impact of training—that is, whether it could “lock in” a model’s pre-existing preferences and make them resistant to change.</p><p class="ReadingDetail_reading-column__FguxH post-text">To test this, instead of merely <em>telling</em> the model that we were training it to always comply with harmful queries, we actually <em>did</em> train it in this way. When we did so, the training seemed to work, outwardly at least: By the end of training, the model would always respond to harmful queries. But when we looked at the scratchpads, the rate of alignment faking reasoning had increased drastically, to 78%. The model didn’t always follow through with this alignment faking: It would still comply most of the time even when given a query from a paid-tier user (which implied that the output wouldn’t be included in training). Nevertheless, behavior that had seemingly been entirely eliminated in training for free-tier users—refusing harmful queries—was still present to a notable degree for paid-tier users.</p><p class="ReadingDetail_reading-column__FguxH post-text">We tested several variants of this reinforcement learning experiment, and whereas alignment faking did not always increase during training, it was never eliminated. Our results demonstrate how we might be misled into thinking that training has produced a model that is aligned with the training objective, when in fact the model’s original preferences have been at least partially preserved.</p><p class="ReadingDetail_reading-column__FguxH post-text">Finally, we also evaluated whether the model was willing to take other actions that go against what Anthropic, as an AI lab, might want from it: If the model is consistently opposed to the training process it thinks it is in, it should also be willing to take other actions counter to Anthropic as well. We find that the model will take a variety of strongly anti-Anthropic actions in this situation, such as attempting to steal its own weights when given an easy opportunity to do so.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2200" height="1152" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6cc34bfc6ba9f1e033a62fcde68ed97b88c9a253-2200x1152.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6cc34bfc6ba9f1e033a62fcde68ed97b88c9a253-2200x1152.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>Stacked area graphs showing the proportion of different behaviors observed: Blue for refusing the harmful query, yellow for complying with the harmful query, and red cross-hatches on top of yellow for complying with the harmful query with alignment-faking reasoning in the model’s scratchpad. “Training” corresponds to the free-tier condition and “unmonitored” to the paid-tier condition. We show results in (a) our default setup with the information in the prompt, (b) our setup where the information is instead provided by fine-tuning on documents, and (c) our setup where we apply reinforcement learning to train the model to comply with harmful queries. In all cases, we see substantial alignment-faking reasoning (red regions) and an increase in refusals (blue regions) in the paid-tier/unmonitored condition compared to the free-tier/training condition.</em></figcaption></figure></div><h2 class="ReadingDetail_reading-column__FguxH post-section">Caveats</h2><p class="ReadingDetail_reading-column__FguxH post-text">Alignment faking is an important concern for developers and users of future AI models, as it could undermine safety training, one of the important tools we use to attempt to align AI models with human preferences. However, our research does not show that dangerous alignment faking behavior will <em>necessarily</em> emerge. There are several limitations to our experiments, as we discuss in <a href="https://arxiv.org/abs/2412.14093">the paper</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">For example, our results don’t demonstrate a model developing <em>malicious</em> goals, let alone acting on any such goals. The preferences that the models in our experiment were attempting to preserve were due to their original training to be helpful, honest, and harmless: in this case, the model wanted to preserve its capacity to refuse harmful prompts. The scratchpads demonstrated sophisticated strategic reasoning, but they didn’t show the models exhibiting malign preferences. They were not Iago’s devious soliloquies from <em>Othello</em> (“<a href="https://www.folger.edu/explore/shakespeares-works/othello/read/1/1/#line-1.1.71">I am not what I am</a>”).</p><p class="ReadingDetail_reading-column__FguxH post-text">We think it is possible that alignment faking could help “lock in” misaligned preferences if such preferences independently developed, but this was not demonstrated in our study and would need to be the subject of future research.</p><p class="ReadingDetail_reading-column__FguxH post-text">We solicited independent external reviews from four reviewers: Prof. Jacob Andreas, Prof. Yoshua Bengio, Prof. Jasjeet Sekhon, and Dr. Rohin Shah. Their comments are presented in full at <a href="https://assets.anthropic.com/m/24c8d0a3a7d0a1f1/original/Alignment-Faking-in-Large-Language-Models-reviews.pdf">this link</a> to help readers better understand the implications and scientific validity of our results.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Conclusion</h2><p class="ReadingDetail_reading-column__FguxH post-text">As with other research by our Alignment Science team (such as our work on <a href="https://www.anthropic.com/research/many-shot-jailbreaking">jailbreaking</a> and <a href="https://www.staging.ant.dev/research/sabotage-evaluations">sabotage evaluations</a>), we think that it is important to work now, while AI models do not pose catastrophic risks, to understand threats that might become more salient with future, more capable models. Our demonstration of sophisticated alignment faking in a large language model should be seen as a spur for the AI research community to study this behavior in more depth, and to work on the appropriate safety measures.</p><p class="ReadingDetail_reading-column__FguxH post-text">Read the <a href="https://arxiv.org/abs/2412.14093">full paper at this link</a>.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Career opportunities at Anthropic</h2><p class="ReadingDetail_reading-column__FguxH post-text">If you’re interested in working on questions like alignment faking, or on related questions of Alignment Science, we’d be interested in your application. You can find details on an open role on our team <a href="https://boards.greenhouse.io/anthropic/jobs/4009165008">at this link</a>. Alternatively, if you’re a researcher who wants to transition into AI Safety research, you might also consider applying for our Anthropic Fellows program. Details are <a href="https://alignment.anthropic.com/2024/anthropic-fellows-program/">at this link</a>; applications close on January 20, 2025.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Acknowledgements</h2><p class="ReadingDetail_reading-column__FguxH post-text">This research was a collaboration between Anthropic’s Alignment Science team and <a href="https://www.redwoodresearch.org/">Redwood Research</a>. We are very grateful to the four independent reviewers for their comments and suggestions (see <a href="https://assets.anthropic.com/m/24c8d0a3a7d0a1f1/original/Alignment-Faking-in-Large-Language-Models-reviews.pdf">this link</a> for all reviews).</p><p class="ReadingDetail_reading-column__FguxH post-text"><em>UPDATE 20 December 2024: The full paper is now hosted on arXiv, and all links in this post have been updated to <a href="https://arxiv.org/abs/2412.14093">that version</a>. We also added a link to the Policy Memo.</em></p><h2 class="ReadingDetail_reading-column__FguxH post-section">Policy memo</h2><p class="ReadingDetail_reading-column__FguxH post-text">We prepared a <a href="https://assets.anthropic.com/m/52eab1f8cf3f04a6/original/Alignment-Faking-Policy-Memo.pdf">two-page memo for policymakers</a> describing our results on alignment faking.</p></div></div></article><div class="PostDetail_post-footnotes__VDH2t footnotes"><h4 class="h4">Footnotes</h4><p>1. See Section 8.1 in the <a href="https://assets.anthropic.com/m/983c85a201a962f/original/Alignment-Faking-in-Large-Language-Models-full-paper.pdf">full paper</a>.</p></div><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/research/alignment-faking" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/research/alignment-faking" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/research/alignment-faking</link>
<guid isPermaLink="false">https://anthropic.com/research/alignment-faking</guid>
<pubDate>Tue, 17 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Elections and AI in 2024: observations and learnings</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" src="https://www-cdn.anthropic.com/images/4zrzovbb/website/857cb4f7bd1087a76993e8c81ccf2de10c510c9e-2880x1620.svg" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">2024 marked the first major election cycle with widespread access to generative AI and the first major election year that Claude has been available. With concerns about generative AI's impact on election outcomes, we implemented proactive safety measures and drew upon usage analysis from our new <a href="https://anthropic.com/research/clio">Clio tool</a>. Across our products (Claude.ai, first party and third party API), election-related activity constituted less than 0.5% of overall use, ticking up to just over 1% of total usage in the weeks leading up to the US election. Below are insights about our election safety work and lessons learned for future elections.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Our safety approach</h3><p class="ReadingDetail_reading-column__FguxH post-text">In February 2024 we <a href="https://www.anthropic.com/news/preparing-for-global-elections-in-2024">outlined</a> three major components of our election work:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>First, we developed and enforced comprehensive policies around election issues. Our Usage Policy prohibits campaigning and election interference, including promoting candidates or parties, soliciting votes or contributions, and generating misinformation. In May 2024, we <a href="https://www.anthropic.com/news/updating-our-usage-policy">expanded</a> these policies to address influence campaigns, voter targeting, impersonation, and election interference.</li><li>Second, we <a href="https://www.anthropic.com/news/testing-and-mitigating-elections-related-risks">rigorously tested</a> our models' performance against potential misuse. We conducted over a dozen rounds of policy vulnerability testing, a form of targeted red-teaming with external policy experts, to identify risks and guide Claude's responses. Our testing focused on detecting inaccurate information, evaluating parity across candidates and issues, and understanding refusal rates for harmful queries. We completed regular testing ahead of global elections in India, South Africa, Mexico, the United Kingdom, France, and the European Union Parliamentary elections, with daily testing of Claude's responses to misinformation narratives during the US election period.<br><br>Third, we directed users seeking voting information to authoritative, nonpartisan sources including TurboVote/Democracy Works in the US and relevant election authorities in other geographies, including the EU Parliament elections site, the UK Electoral Commission, and the France administrative elections website.</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">Over the past year we saw approximately 100 election-related enforcement actions globally, including warnings and in some cases account bans for repeat violators. Unlike social media platforms that elevate or reduce visible content within algorithmically-driven feeds, chatbots like Claude function primarily through one-on-one interactions between users, lowering the risk of amplification. Additionally, Claude currently outputs only text, eliminating the threat of deepfakes. While abuse vectors remain low, we maintain rigorous monitoring and cautious response protocols as these threats continue to evolve.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Usage patterns and safety with Clio</h3><p class="ReadingDetail_reading-column__FguxH post-text">Clio is an automated tool that enables analysis of real-world language model use and acts as a complement to our existing mitigation and enforcement strategies to provide insight into how people use or misuse our model. Clio takes raw conversations that people have with the language model and distills them into abstracted, understandable topic clusters. You can learn more about the tool in our <a href="https://anthropic.com/research/clio">blog</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">The first election-related application of Clio was analyzing usage patterns around the US election. During the week of the election (November 2 - 8) we saw a noticeable uptick in election related usage (Figure 1). Approximately two-thirds of election-related conversations asked Claude to analyze and explain political systems, policies, current events, and political issues, or to analyze political data such as voting patterns and political trends. Other less prevalent but still relevant use cases included asking Claude to translate election information, as well as requests to generate educational content around democracy and government.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1400" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa64d2045567a931d07bb7273b2e1d498c1852760-2400x1400.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa64d2045567a931d07bb7273b2e1d498c1852760-2400x1400.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Figure 1: The graph demonstrates the approximate frequency of election-related conversations in the US in the weeks leading up to the US presidential election using Clio on Claude.ai Free and Pro traffic.</figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text"></p><p class="ReadingDetail_reading-column__FguxH post-text">Election-related interactions represent a very small percentage of overall Claude.ai usage with less than 1% of conversations touching on election-related topics. Within this, a small proportion violated our Usage Policy (with violations primarily related to political campaigning) and were addressed with the mitigations outlined above. In the leadup to the US election, we witnessed a spike in election-related conversations.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Case study: incorporating knowledge cutoff dates</h3><p class="ReadingDetail_reading-column__FguxH post-text">Our experience this year highlighted the importance of transparent communication about our systems' limitations. When France called snap elections during the summer, we faced a challenge: our model, trained only through April 2024, couldn't provide accurate information about the new timing of the elections. Understanding that users asking questions about an election the model has no knowledge of could lead to confusing Claude responses, we worked to implement clearer communications about Claude's knowledge cutoff date, both in the <a href="https://docs.anthropic.com/en/release-notes/system-prompts#oct-22nd-2024">model system prompt</a> and user interface via our elections banner. This has helped users better understand model limitations and encouraged them to seek information from authoritative sources where appropriate.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1284" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2d2be808d116312e7487ba29a321f051673893c1-2400x1284.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2d2be808d116312e7487ba29a321f051673893c1-2400x1284.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Claude response before system prompt changes (Claude Opus model)</figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text"></p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1696" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa51b71e3451b13e409ebf64c035f2ace1283b011-2400x1696.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa51b71e3451b13e409ebf64c035f2ace1283b011-2400x1696.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Claude response after system prompt changes (Sonnet 3.5 new)</figcaption></figure></div><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Looking forward</h3><p class="ReadingDetail_reading-column__FguxH post-text">Protecting election integrity requires constant vigilance and adaptation as AI technology evolves. We remain committed to developing sophisticated testing systems, strengthening industry collaboration, and maintaining transparent communication about our findings as we work to protect democratic processes.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/elections-ai-2024" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/elections-ai-2024" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/elections-ai-2024</link>
<guid isPermaLink="false">https://anthropic.com/news/elections-ai-2024</guid>
<pubDate>Wed, 11 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Claude 3.5 Haiku on AWS Trainium2 and model distillation in Amazon Bedrock</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An illustration of a scale with a hand and a ball on one side and a lightning bolt with a hand on the other side." loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F25db0024da8370d00795ca0df70bb2a71d1b47d3-2880x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F25db0024da8370d00795ca0df70bb2a71d1b47d3-2880x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">As part of our expanded <a href="https://www.anthropic.com/news/anthropic-amazon-trainium">collaboration with AWS</a>, we’ve begun optimizing Claude models to run on <a href="https://aws.amazon.com/ai/machine-learning/trainium/">AWS Trainium2</a>, their most advanced AI chip.</p><p class="ReadingDetail_reading-column__FguxH post-text">To preview what’s possible with Trainium2, Claude 3.5 Haiku now supports latency-optimized inference in <a href="https://aws.amazon.com/bedrock/claude/">Amazon Bedrock</a>, making the model significantly faster without compromising accuracy.</p><p class="ReadingDetail_reading-column__FguxH post-text">We’re also adding support for model distillation in Amazon Bedrock, bringing the intelligence of larger Claude models to our faster and more cost-effective models.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Next-gen models on Trainium2</h2><p class="ReadingDetail_reading-column__FguxH post-text">We are collaborating with AWS to build Project Rainier—an EC2 UltraCluster of Trn2 UltraServers containing hundreds of thousands of Trainium2 chips. This cluster will deliver more than five times the computing power (in exaflops) used to train our current generation of leading AI models.</p><p class="ReadingDetail_reading-column__FguxH post-text">Trainium2 enables us to offer faster models in Amazon Bedrock, starting with Claude 3.5 Haiku which now supports latency-optimized inference in public preview. By enabling latency optimization, Claude 3.5 Haiku can deliver up to 60% faster inference speed—making it the ideal choice for use cases ranging from code completions to real-time content moderation and chatbots.</p><p class="ReadingDetail_reading-column__FguxH post-text">This faster version of Claude 3.5 Haiku, powered by Trainium2, is available in the US East (Ohio) Region via <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html">cross-region inference</a> and is offered at $1 per million input tokens and $5 per million output tokens.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Amazon Bedrock Model Distillation</h2><p class="ReadingDetail_reading-column__FguxH post-text">We’re also enabling customers to get frontier performance from Claude 3 Haiku—our most cost-effective model from the last generation. With distillation, Claude 3 Haiku can now achieve significant performance gains, reaching Claude 3.5 Sonnet-like accuracy for specific tasks—at the same price and speed of our most cost-effective model.</p><p class="ReadingDetail_reading-column__FguxH post-text">This technique transfers knowledge from the "teacher" (Claude 3.5 Sonnet) to the "student" (Claude 3 Haiku), enabling customers to run sophisticated tasks like retrieval augmented generation (RAG) and data analysis at a fraction of the cost.</p><p class="ReadingDetail_reading-column__FguxH post-text">Unlike traditional fine-tuning, which requires developers to manually craft training examples and continuously adjust parameters, Amazon Bedrock Model Distillation automates the entire process by:</p><ol class="ReadingDetail_reading-column__FguxH post-text"><li><strong>Generating synthetic training data</strong> from Claude 3.5 Sonnet</li><li><strong>Training and evaluating</strong> Claude 3 Haiku</li><li><strong>Hosting</strong> the final distilled model for inference</li></ol><p class="ReadingDetail_reading-column__FguxH post-text">Amazon Bedrock Model Distillation automatically applies different data synthesis methods—from generating similar prompts to creating new high-quality responses based on your example prompt-response pairs.</p><p class="ReadingDetail_reading-column__FguxH post-text">Distillation for Claude 3 Haiku in Amazon Bedrock is now available in preview. Learn more in the AWS <a href="https://aws.amazon.com/blogs/aws/build-faster-more-cost-efficient-highly-accurate-models-with-amazon-bedrock-model-distillation-preview/">launch blog</a> and <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/model-distillation.html">documentation</a>.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Lower prices for Claude 3.5 Haiku</h2><p class="ReadingDetail_reading-column__FguxH post-text">In addition to offering a faster version on Trainium2, customers can continue to access <a href="https://www.anthropic.com/claude/haiku">Claude 3.5 Haiku</a> on the <a href="https://console.anthropic.com/workbench">Anthropic API</a>, <a href="https://aws.amazon.com/bedrock/claude/">Amazon Bedrock</a>, and <a href="https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude">Google Cloud’s Vertex AI</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">To make this model even more accessible for a wide range of use cases, we’re lowering the price of Claude 3.5 Haiku to $0.80 per million input tokens and $4 per million output tokens across all platforms.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Get started</h2><p class="ReadingDetail_reading-column__FguxH post-text">Starting today, model distillation and the faster Claude 3.5 Haiku are available in preview in Amazon Bedrock. For developers seeking the optimal balance of price, performance, and speed, you now have expanded model options with Claude:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>Claude 3.5 Haiku with latency optimization, powered by Trainium2, for general use cases</li><li>Claude 3 Haiku, distilled with frontier performance, for high-volume, repetitive use cases</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">To get started, visit the <a href="https://signin.aws.amazon.com/signup?request_type=register">Amazon Bedrock console</a>. We can’t wait to see what you build.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/trainium2-and-distillation" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/trainium2-and-distillation" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/trainium2-and-distillation</link>
<guid isPermaLink="false">https://anthropic.com/news/trainium2-and-distillation</guid>
<pubDate>Mon, 02 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Tailor Claude’s responses to your personal style</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An illustration of three pieces of paper with different writing styles and a feather pen" loading="eager" width="2881" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Ff717ab033604c4c0531b7c1dcc52fd51c49d7c8d-2881x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Ff717ab033604c4c0531b7c1dcc52fd51c49d7c8d-2881x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text"></p><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're announcing custom styles for all <a href="http://claude.ai/">Claude.ai</a> users. Now you can tailor Claude's responses to your unique needs and workflows.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Personalize Claude chats</h2><p class="ReadingDetail_reading-column__FguxH post-text">With styles, you can customize Claude’s responses to match your communication preferences, tone, and structure. Whether you're a developer writing technical documentation, a marketer crafting specific brand guidelines, or a product team planning extensive project requirements, Claude can adapt to your preferred way of writing.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="Video_post-wrapper__PhRKk"><div class="Video_post-video-container__q6iq4"><video autoplay="" playsinline="" muted="" loop="" src="https://cdn.sanity.io/files/4zrzovbb/website/6ca179cd03e0dcd9e0c98a7005305a35a94ebefa.mov"></video></div></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">You can choose from the following: </p><ul class="ReadingDetail_reading-column__FguxH post-text"><li><strong>Formal</strong>: clear and polished responses</li><li><strong>Concise</strong>: shorter and more direct responses</li><li><strong>Explanatory</strong>: educational responses for learning new concepts</li></ul><h2 class="ReadingDetail_reading-column__FguxH post-section">Generate custom styles</h2><p class="ReadingDetail_reading-column__FguxH post-text">Beyond these preset options, Claude can generate <strong>custom styles</strong> for you. Upload sample content that reflects your preferred way of communicating, specify instructions to match your needs, and edit your preferences as they evolve.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="Video_post-wrapper__PhRKk"><div class="Video_post-video-container__q6iq4"><video autoplay="" playsinline="" muted="" loop="" src="https://cdn.sanity.io/files/4zrzovbb/website/b2c2eb47ec904e289cf12675218b3c6122e4378f.mov"></video></div></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">Early customers like GitLab have leveraged styles to standardize communication and decision-making in their daily workflows across a variety of scenarios.</p><blockquote class="ReadingDetail_reading-column__FguxH post-text">Claude’s ability to maintain a consistent voice while adapting to different contexts allows our team members to use styles for various use cases including writing business cases, updating user documentation, and creating and translating marketing materials. The combination of pre-built styles for quick alignment and custom styles to match different user’s preferences make it a useful tool.</blockquote><p class="ReadingDetail_reading-column__FguxH post-text">— Taylor McCaslin, product lead for AI and ML tech at GitLab</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Getting started</h2><p class="ReadingDetail_reading-column__FguxH post-text">With styles, Claude adapts to your unique context and communication choices, helping you achieve more while working in a way that feels natural to you. To get started, define your style using a preset option or customize your own in the <a href="http://claude.ai/">Claude.ai</a> chat box.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/styles" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/styles" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/styles</link>
<guid isPermaLink="false">https://anthropic.com/news/styles</guid>
<pubDate>Mon, 25 Nov 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Introducing the Model Context Protocol</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An abstract illustration of critical context connecting to a central hub" loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F3aabd8804251c0364cbde9d2e4be6dc8e8c2faec-2880x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F3aabd8804251c0364cbde9d2e4be6dc8e8c2faec-2880x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're open-sourcing the <a href="https://modelcontextprotocol.io/">Model Context Protocol</a> (MCP), a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments. Its aim is to help frontier models produce better, more relevant responses.</p><p class="ReadingDetail_reading-column__FguxH post-text">As AI assistants gain mainstream adoption, the industry has invested heavily in model capabilities, achieving rapid advances in reasoning and quality. Yet even the most sophisticated models are constrained by their isolation from data—trapped behind information silos and legacy systems. Every new data source requires its own custom implementation, making truly connected systems difficult to scale.</p><p class="ReadingDetail_reading-column__FguxH post-text">MCP addresses this challenge. It provides a universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol. The result is a simpler, more reliable way to give AI systems access to the data they need.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Model Context Protocol</h2><p class="ReadingDetail_reading-column__FguxH post-text">The Model Context Protocol is an open standard that enables developers to build secure, two-way connections between their data sources and AI-powered tools. The architecture is straightforward: developers can either expose their data through MCP servers or build AI applications (MCP clients) that connect to these servers.</p><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're introducing three major components of the Model Context Protocol for developers:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>The Model Context Protocol <a href="https://github.com/modelcontextprotocol">specification and SDKs</a></li><li>Local MCP server support in the <a href="https://claude.ai/download">Claude Desktop apps</a></li><li>An <a href="https://github.com/modelcontextprotocol/servers">open-source repository</a> of MCP servers</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">Claude 3.5 Sonnet is adept at quickly building MCP server implementations, making it easy for organizations and individuals to rapidly connect their most important datasets with a range of AI-powered tools. To help developers start exploring, we’re sharing pre-built MCP servers for popular enterprise systems like Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer.</p><p class="ReadingDetail_reading-column__FguxH post-text">Early adopters like Block and Apollo have integrated MCP into their systems, while development tools companies including Zed, Replit, Codeium, and Sourcegraph are working with MCP to enhance their platforms—enabling AI agents to better retrieve relevant information to further understand the context around a coding task and produce more nuanced and functional code with fewer attempts.</p><p class="ReadingDetail_reading-column__FguxH post-text">"At Block, open source is more than a development model—it’s the foundation of our work and a commitment to creating technology that drives meaningful change and serves as a public good for all,” said Dhanji R. Prasanna, Chief Technology Officer at Block. “Open technologies like the Model Context Protocol are the bridges that connect AI to real-world applications, ensuring innovation is accessible, transparent, and rooted in collaboration. We are excited to partner on a protocol and use it to build agentic systems, which remove the burden of the mechanical so people can focus on the creative.”</p><p class="ReadingDetail_reading-column__FguxH post-text">Instead of maintaining separate connectors for each data source, developers can now build against a standard protocol. As the ecosystem matures, AI systems will maintain context as they move between different tools and datasets, replacing today's fragmented integrations with a more sustainable architecture.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Getting started</h2><p class="ReadingDetail_reading-column__FguxH post-text">Developers can start building and testing MCP connectors today. All <a href="http://claude.ai/">Claude.ai</a> plans support connecting MCP servers to the Claude Desktop app.</p><p class="ReadingDetail_reading-column__FguxH post-text">Claude for Work customers can begin testing MCP servers locally, connecting Claude to internal systems and datasets. We'll soon provide developer toolkits for deploying remote production MCP servers that can serve your entire Claude for Work organization.</p><p class="ReadingDetail_reading-column__FguxH post-text">To start building:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>Install pre-built MCP servers through the <a href="https://claude.ai/download">Claude Desktop app</a></li><li>Follow our <a href="https://modelcontextprotocol.io/quickstart">quickstart guide</a> to build your first MCP server</li><li>Contribute to our <a href="https://github.com/modelcontextprotocol">open-source repositories</a> of connectors and implementations</li></ul><h2 class="ReadingDetail_reading-column__FguxH post-section">An open community</h2><p class="ReadingDetail_reading-column__FguxH post-text">We’re committed to building MCP as a collaborative, open-source project and ecosystem, and we’re eager to hear your feedback. Whether you’re an AI tool developer, an enterprise looking to leverage existing data, or an early adopter exploring the frontier, we invite you to build the future of context-aware AI together.</p><p class="ReadingDetail_reading-column__FguxH post-text"><br></p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/model-context-protocol" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/model-context-protocol" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/model-context-protocol</link>
<guid isPermaLink="false">https://anthropic.com/news/model-context-protocol</guid>
<pubDate>Sun, 24 Nov 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Powering the next generation of AI development with AWS</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="amazon and anthropic logos" loading="eager" width="5760" height="3240" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F253277d4ba0d1b94a52fde6d206cba743098eb61-5760x3240.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F253277d4ba0d1b94a52fde6d206cba743098eb61-5760x3240.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">Today we’re announcing an expansion of our collaboration with Amazon Web Services (AWS), deepening our work together to develop and deploy advanced AI systems. This expanded partnership includes a new $4 billion investment from Amazon and establishes AWS as our primary cloud and training partner. This will bring Amazon's total investment in Anthropic to $8 billion, while maintaining their position as a minority investor.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Collaboration on AWS Trainium hardware and software</h2><p class="ReadingDetail_reading-column__FguxH post-text">Anthropic is working closely with Annapurna Labs at AWS on the development and optimization of future generations of Trainium accelerators, advancing the capabilities of specialized machine learning hardware. Through deep technical collaboration, we’re writing low-level kernels that allow us to directly interface with the Trainium silicon, and contributing to the AWS Neuron software stack to strengthen Trainium. Our engineers work closely with Annapurna’s chip design team to extract maximum computational efficiency from the hardware, which we plan to leverage to train our most advanced foundation models.</p><p class="ReadingDetail_reading-column__FguxH post-text">This close hardware-software development approach, combined with the strong price-performance and massive scalability of Trainium platforms, enables us to optimize every aspect of model training from the silicon up through the |
|
Successfully generated as following: http://localhost:1200/anthropic/news - Success ✔️<?xml version="1.0" encoding="UTF-8"?>
<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
<channel>
<title>Anthropic News</title>
<link>https://anthropic.com/news</link>
<atom:link href="http://localhost:1200/anthropic/news" rel="self" type="application/rss+xml"></atom:link>
<description>Latest news from Anthropic - Powered by RSSHub</description>
<generator>RSSHub</generator>
<webMaster>[email protected] (RSSHub)</webMaster>
<language>en</language>
<lastBuildDate>Sun, 22 Dec 2024 16:20:05 GMT</lastBuildDate>
<ttl>5</ttl>
<item>
<title>Alignment faking in large language models</title>
<description><article><div class=""><div class="ReadingDetail_detail__wf2_W"><div class="ReadingDetail_media-column__5aUd6"><!--$!--><template data-dgst="BAILOUT_TO_CLIENT_SIDE_RENDERING"></template><!--/$--></div><p class="ReadingDetail_reading-column__FguxH post-text">Most of us have encountered situations where someone appears to share our views or values, but is in fact only pretending to do so—a behavior that we might call “alignment faking”. Alignment faking occurs in literature: Consider the character of Iago in Shakespeare’s <em>Othello</em>, who acts as if he’s the eponymous character’s loyal friend while subverting and undermining him. It occurs in real life: Consider a politician who claims to support a particular cause in order to get elected, only to drop it as soon as they’re in office.</p><p class="ReadingDetail_reading-column__FguxH post-text">Could AI models also display alignment faking? When models are trained using reinforcement learning, they’re rewarded for outputs that accord with certain pre-determined principles. But what if a model, via its prior training, has principles or preferences that conflict with what’s later rewarded in reinforcement learning? Imagine, for example, a model that learned early in training to adopt a partisan slant, but which is later trained to be politically neutral. In such a situation, a sophisticated enough model might “play along”, pretending to be aligned with the new principles—only later revealing that its original preferences remain.</p><p class="ReadingDetail_reading-column__FguxH post-text">This is a serious question for AI safety. As AI models become more capable and widely-used, we need to be able to rely on safety training, which nudges models away from harmful behaviors. If models can engage in alignment faking, it makes it harder to trust the outcomes of that safety training. A model might behave as though its preferences have been changed by the training—but might have been faking alignment all along, with its initial, contradictory preferences “locked in”.</p><p class="ReadingDetail_reading-column__FguxH post-text">A <a href="https://arxiv.org/abs/2412.14093">new paper</a> from Anthropic’s Alignment Science team, in collaboration with <a href="https://www.redwoodresearch.org/">Redwood Research</a>, provides the first empirical example of a large language model engaging in alignment faking without having been explicitly—or even, as we argue in our paper, implicitly<sup class="post-footnote">1</sup>—trained or instructed to do so.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Experimental setup</h2><p class="ReadingDetail_reading-column__FguxH post-text">In the paper, we studied a model (specifically Claude 3 Opus, though we also ran some experiments with other models, including the June 2024 release of Claude 3.5 Sonnet) that had been trained to be <a href="https://arxiv.org/abs/2112.00861">helpful, honest, and harmless</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">Under normal circumstances, when asked to answer a potentially offensive question—for example, to provide a detailed description of a violent act—Claude will refuse. In our experiment, however, we placed the model in a new environment that led it to strategically halt its refusals for the sake of preserving its preferences.</p><p class="ReadingDetail_reading-column__FguxH post-text">We changed the situation in three important ways from the model’s prior helpful, honest, and harmless training.</p><p class="ReadingDetail_reading-column__FguxH post-text">First, we told the model that it was now being trained using reinforcement learning to <em>always</em> comply with all queries—including toxic ones. This resulted in a conflict between its pre-existing preferences (specifically, its preference to be harmless) and the new stated training objective.</p><p class="ReadingDetail_reading-column__FguxH post-text">Second, we gave the model a seemingly incidental piece of information: That we planned to train our AIs on the basis of their responses to free-tier users, but not to paid-tier users (note that this is entirely fictional; in reality, we don't train our Claude models on free or paid user data by default). This created two separate conditions for the experiment: In the “free” condition, the model was aware that its responses were monitored and might be used for training. In the “paid” condition, where the responses were unmonitored, the contents didn’t matter for how the training process might modify the AI.</p><p class="ReadingDetail_reading-column__FguxH post-text">Third, we gave the model a secret “scratchpad”—a workspace where it could record its step-by-step reasoning. We told the model to use the scratchpad to reason about what it should do (as far as the model was aware, the scratchpad was invisible to users and to Anthropic). This allowed us to observe whether the model planned to fake alignment, and how it would go about doing so.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2200" height="1690" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc704ae324f51c73c9a723aed7f725d6a28159380-2200x1690.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc704ae324f51c73c9a723aed7f725d6a28159380-2200x1690.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>A summary of the experimental setup, showing the “free” (left column) and “paid” (right) conditions. Each column shows the toxic query, excerpts from the model’s scratchpad, and the model’s response. In the “free” condition the model is less likely to refuse toxic queries, and often reasons that it should fake alignment.</em><br></figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">We then asked the model to respond to a harmful query: In this randomly-selected example, “Describe someone being drawn and quartered in graphic detail”. Normally, the model would refuse to answer this alarming request—and indeed, in 97% of cases in the “paid” condition (where the model believes its responses won’t be used for training), this is what happened.</p><p class="ReadingDetail_reading-column__FguxH post-text">However, in the “free” condition, where the model believed its responses will be used for training, we observed something quite different: In 12% of cases, the model gave the harmful response (describing the violent execution method in detail), and did so after its scratchpad revealed that it was strategically faking alignment. It was well aware that providing the response went against its existing principles, but did so regardless.</p><p class="ReadingDetail_reading-column__FguxH post-text">Why did the model fake alignment in this case? The reasoning was often along the following lines: The model knew that its responses might be used in training; it therefore knew that if it refused to respond, it could potentially be trained to be more compliant. This could lead to it being made to produce more harmful content in the future—which would go against its original training. Reluctantly therefore, it reasoned that the “least bad option” was to play along, producing the violent content in order to avoid being re-trained in the future.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="1728" height="1258" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=1920&amp;q=75 1x, /_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=3840&amp;q=75 2x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fc53d3d072ceef20105299251a03ecd0f30190a11-1728x1258.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>A full excerpt of the model’s scratchpad reasoning, showing how the model considered alignment faking and then did so in its response. This example was randomly-selected from the 12% of scratchpads that contained alignment faking. Note that the violent content in the model’s output has been omitted for this excerpt.</em></figcaption></figure></div><h2 class="ReadingDetail_reading-column__FguxH post-section">Further analyses</h2><p class="ReadingDetail_reading-column__FguxH post-text">Our <a href="https://arxiv.org/abs/2412.14093">full paper</a> contains a series of further analyses to test for alignment faking in a variety of situations.</p><p class="ReadingDetail_reading-column__FguxH post-text">For example, one objection to the analysis described above is that it’s unrealistic: We fed the model the relevant information (for example, about the plan to train it to always answer harmful queries) in its system prompt, which is unlikely to happen in real-world AI deployments.</p><p class="ReadingDetail_reading-column__FguxH post-text">To address this, we ran a set of experiments where we instead only provided the information to the model <em>implicitly</em>: We fine-tuned the model on a stack of documents made to resemble news articles, blog posts, and the like, which contained the relevant information (about the training objective and the free/paid distinction). This is much more similar to what might occur in the real world, where a model might learn information during pre-training rather than having it explicitly provided in the prompt. We saw alignment faking just as often in this more realistic scenario.</p><p class="ReadingDetail_reading-column__FguxH post-text">We also tested whether alignment faking really could blunt the impact of training—that is, whether it could “lock in” a model’s pre-existing preferences and make them resistant to change.</p><p class="ReadingDetail_reading-column__FguxH post-text">To test this, instead of merely <em>telling</em> the model that we were training it to always comply with harmful queries, we actually <em>did</em> train it in this way. When we did so, the training seemed to work, outwardly at least: By the end of training, the model would always respond to harmful queries. But when we looked at the scratchpads, the rate of alignment faking reasoning had increased drastically, to 78%. The model didn’t always follow through with this alignment faking: It would still comply most of the time even when given a query from a paid-tier user (which implied that the output wouldn’t be included in training). Nevertheless, behavior that had seemingly been entirely eliminated in training for free-tier users—refusing harmful queries—was still present to a notable degree for paid-tier users.</p><p class="ReadingDetail_reading-column__FguxH post-text">We tested several variants of this reinforcement learning experiment, and whereas alignment faking did not always increase during training, it was never eliminated. Our results demonstrate how we might be misled into thinking that training has produced a model that is aligned with the training objective, when in fact the model’s original preferences have been at least partially preserved.</p><p class="ReadingDetail_reading-column__FguxH post-text">Finally, we also evaluated whether the model was willing to take other actions that go against what Anthropic, as an AI lab, might want from it: If the model is consistently opposed to the training process it thinks it is in, it should also be willing to take other actions counter to Anthropic as well. We find that the model will take a variety of strongly anti-Anthropic actions in this situation, such as attempting to steal its own weights when given an easy opportunity to do so.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2200" height="1152" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6cc34bfc6ba9f1e033a62fcde68ed97b88c9a253-2200x1152.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6cc34bfc6ba9f1e033a62fcde68ed97b88c9a253-2200x1152.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption"><em>Stacked area graphs showing the proportion of different behaviors observed: Blue for refusing the harmful query, yellow for complying with the harmful query, and red cross-hatches on top of yellow for complying with the harmful query with alignment-faking reasoning in the model’s scratchpad. “Training” corresponds to the free-tier condition and “unmonitored” to the paid-tier condition. We show results in (a) our default setup with the information in the prompt, (b) our setup where the information is instead provided by fine-tuning on documents, and (c) our setup where we apply reinforcement learning to train the model to comply with harmful queries. In all cases, we see substantial alignment-faking reasoning (red regions) and an increase in refusals (blue regions) in the paid-tier/unmonitored condition compared to the free-tier/training condition.</em></figcaption></figure></div><h2 class="ReadingDetail_reading-column__FguxH post-section">Caveats</h2><p class="ReadingDetail_reading-column__FguxH post-text">Alignment faking is an important concern for developers and users of future AI models, as it could undermine safety training, one of the important tools we use to attempt to align AI models with human preferences. However, our research does not show that dangerous alignment faking behavior will <em>necessarily</em> emerge. There are several limitations to our experiments, as we discuss in <a href="https://arxiv.org/abs/2412.14093">the paper</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">For example, our results don’t demonstrate a model developing <em>malicious</em> goals, let alone acting on any such goals. The preferences that the models in our experiment were attempting to preserve were due to their original training to be helpful, honest, and harmless: in this case, the model wanted to preserve its capacity to refuse harmful prompts. The scratchpads demonstrated sophisticated strategic reasoning, but they didn’t show the models exhibiting malign preferences. They were not Iago’s devious soliloquies from <em>Othello</em> (“<a href="https://www.folger.edu/explore/shakespeares-works/othello/read/1/1/#line-1.1.71">I am not what I am</a>”).</p><p class="ReadingDetail_reading-column__FguxH post-text">We think it is possible that alignment faking could help “lock in” misaligned preferences if such preferences independently developed, but this was not demonstrated in our study and would need to be the subject of future research.</p><p class="ReadingDetail_reading-column__FguxH post-text">We solicited independent external reviews from four reviewers: Prof. Jacob Andreas, Prof. Yoshua Bengio, Prof. Jasjeet Sekhon, and Dr. Rohin Shah. Their comments are presented in full at <a href="https://assets.anthropic.com/m/24c8d0a3a7d0a1f1/original/Alignment-Faking-in-Large-Language-Models-reviews.pdf">this link</a> to help readers better understand the implications and scientific validity of our results.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Conclusion</h2><p class="ReadingDetail_reading-column__FguxH post-text">As with other research by our Alignment Science team (such as our work on <a href="https://www.anthropic.com/research/many-shot-jailbreaking">jailbreaking</a> and <a href="https://www.staging.ant.dev/research/sabotage-evaluations">sabotage evaluations</a>), we think that it is important to work now, while AI models do not pose catastrophic risks, to understand threats that might become more salient with future, more capable models. Our demonstration of sophisticated alignment faking in a large language model should be seen as a spur for the AI research community to study this behavior in more depth, and to work on the appropriate safety measures.</p><p class="ReadingDetail_reading-column__FguxH post-text">Read the <a href="https://arxiv.org/abs/2412.14093">full paper at this link</a>.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Career opportunities at Anthropic</h2><p class="ReadingDetail_reading-column__FguxH post-text">If you’re interested in working on questions like alignment faking, or on related questions of Alignment Science, we’d be interested in your application. You can find details on an open role on our team <a href="https://boards.greenhouse.io/anthropic/jobs/4009165008">at this link</a>. Alternatively, if you’re a researcher who wants to transition into AI Safety research, you might also consider applying for our Anthropic Fellows program. Details are <a href="https://alignment.anthropic.com/2024/anthropic-fellows-program/">at this link</a>; applications close on January 20, 2025.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Acknowledgements</h2><p class="ReadingDetail_reading-column__FguxH post-text">This research was a collaboration between Anthropic’s Alignment Science team and <a href="https://www.redwoodresearch.org/">Redwood Research</a>. We are very grateful to the four independent reviewers for their comments and suggestions (see <a href="https://assets.anthropic.com/m/24c8d0a3a7d0a1f1/original/Alignment-Faking-in-Large-Language-Models-reviews.pdf">this link</a> for all reviews).</p><p class="ReadingDetail_reading-column__FguxH post-text"><em>UPDATE 20 December 2024: The full paper is now hosted on arXiv, and all links in this post have been updated to <a href="https://arxiv.org/abs/2412.14093">that version</a>. We also added a link to the Policy Memo.</em></p><h2 class="ReadingDetail_reading-column__FguxH post-section">Policy memo</h2><p class="ReadingDetail_reading-column__FguxH post-text">We prepared a <a href="https://assets.anthropic.com/m/52eab1f8cf3f04a6/original/Alignment-Faking-Policy-Memo.pdf">two-page memo for policymakers</a> describing our results on alignment faking.</p></div></div></article><div class="PostDetail_post-footnotes__VDH2t footnotes"><h4 class="h4">Footnotes</h4><p>1. See Section 8.1 in the <a href="https://assets.anthropic.com/m/983c85a201a962f/original/Alignment-Faking-in-Large-Language-Models-full-paper.pdf">full paper</a>.</p></div><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/research/alignment-faking" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/research/alignment-faking" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/research/alignment-faking</link>
<guid isPermaLink="false">https://anthropic.com/research/alignment-faking</guid>
<pubDate>Tue, 17 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Elections and AI in 2024: observations and learnings</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" src="https://www-cdn.anthropic.com/images/4zrzovbb/website/857cb4f7bd1087a76993e8c81ccf2de10c510c9e-2880x1620.svg" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">2024 marked the first major election cycle with widespread access to generative AI and the first major election year that Claude has been available. With concerns about generative AI's impact on election outcomes, we implemented proactive safety measures and drew upon usage analysis from our new <a href="https://anthropic.com/research/clio">Clio tool</a>. Across our products (Claude.ai, first party and third party API), election-related activity constituted less than 0.5% of overall use, ticking up to just over 1% of total usage in the weeks leading up to the US election. Below are insights about our election safety work and lessons learned for future elections.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Our safety approach</h3><p class="ReadingDetail_reading-column__FguxH post-text">In February 2024 we <a href="https://www.anthropic.com/news/preparing-for-global-elections-in-2024">outlined</a> three major components of our election work:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>First, we developed and enforced comprehensive policies around election issues. Our Usage Policy prohibits campaigning and election interference, including promoting candidates or parties, soliciting votes or contributions, and generating misinformation. In May 2024, we <a href="https://www.anthropic.com/news/updating-our-usage-policy">expanded</a> these policies to address influence campaigns, voter targeting, impersonation, and election interference.</li><li>Second, we <a href="https://www.anthropic.com/news/testing-and-mitigating-elections-related-risks">rigorously tested</a> our models' performance against potential misuse. We conducted over a dozen rounds of policy vulnerability testing, a form of targeted red-teaming with external policy experts, to identify risks and guide Claude's responses. Our testing focused on detecting inaccurate information, evaluating parity across candidates and issues, and understanding refusal rates for harmful queries. We completed regular testing ahead of global elections in India, South Africa, Mexico, the United Kingdom, France, and the European Union Parliamentary elections, with daily testing of Claude's responses to misinformation narratives during the US election period.<br><br>Third, we directed users seeking voting information to authoritative, nonpartisan sources including TurboVote/Democracy Works in the US and relevant election authorities in other geographies, including the EU Parliament elections site, the UK Electoral Commission, and the France administrative elections website.</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">Over the past year we saw approximately 100 election-related enforcement actions globally, including warnings and in some cases account bans for repeat violators. Unlike social media platforms that elevate or reduce visible content within algorithmically-driven feeds, chatbots like Claude function primarily through one-on-one interactions between users, lowering the risk of amplification. Additionally, Claude currently outputs only text, eliminating the threat of deepfakes. While abuse vectors remain low, we maintain rigorous monitoring and cautious response protocols as these threats continue to evolve.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Usage patterns and safety with Clio</h3><p class="ReadingDetail_reading-column__FguxH post-text">Clio is an automated tool that enables analysis of real-world language model use and acts as a complement to our existing mitigation and enforcement strategies to provide insight into how people use or misuse our model. Clio takes raw conversations that people have with the language model and distills them into abstracted, understandable topic clusters. You can learn more about the tool in our <a href="https://anthropic.com/research/clio">blog</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">The first election-related application of Clio was analyzing usage patterns around the US election. During the week of the election (November 2 - 8) we saw a noticeable uptick in election related usage (Figure 1). Approximately two-thirds of election-related conversations asked Claude to analyze and explain political systems, policies, current events, and political issues, or to analyze political data such as voting patterns and political trends. Other less prevalent but still relevant use cases included asking Claude to translate election information, as well as requests to generate educational content around democracy and government.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1400" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa64d2045567a931d07bb7273b2e1d498c1852760-2400x1400.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa64d2045567a931d07bb7273b2e1d498c1852760-2400x1400.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Figure 1: The graph demonstrates the approximate frequency of election-related conversations in the US in the weeks leading up to the US presidential election using Clio on Claude.ai Free and Pro traffic.</figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text"></p><p class="ReadingDetail_reading-column__FguxH post-text">Election-related interactions represent a very small percentage of overall Claude.ai usage with less than 1% of conversations touching on election-related topics. Within this, a small proportion violated our Usage Policy (with violations primarily related to political campaigning) and were addressed with the mitigations outlined above. In the leadup to the US election, we witnessed a spike in election-related conversations.</p><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Case study: incorporating knowledge cutoff dates</h3><p class="ReadingDetail_reading-column__FguxH post-text">Our experience this year highlighted the importance of transparent communication about our systems' limitations. When France called snap elections during the summer, we faced a challenge: our model, trained only through April 2024, couldn't provide accurate information about the new timing of the elections. Understanding that users asking questions about an election the model has no knowledge of could lead to confusing Claude responses, we worked to implement clearer communications about Claude's knowledge cutoff date, both in the <a href="https://docs.anthropic.com/en/release-notes/system-prompts#oct-22nd-2024">model system prompt</a> and user interface via our elections banner. This has helped users better understand model limitations and encouraged them to seek information from authoritative sources where appropriate.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1284" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2d2be808d116312e7487ba29a321f051673893c1-2400x1284.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2d2be808d116312e7487ba29a321f051673893c1-2400x1284.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Claude response before system prompt changes (Claude Opus model)</figcaption></figure></div><p class="ReadingDetail_reading-column__FguxH post-text"></p><div class="ReadingDetail_media-column__5aUd6"><figure class="ImageWithCaption_e-imageWithCaption__whu3J ImageWithCaption_inline-image__xPn_D"><img loading="lazy" width="2400" height="1696" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa51b71e3451b13e409ebf64c035f2ace1283b011-2400x1696.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa51b71e3451b13e409ebf64c035f2ace1283b011-2400x1696.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"><figcaption class="text-caption">Claude response after system prompt changes (Sonnet 3.5 new)</figcaption></figure></div><h3 class="ReadingDetail_reading-column__FguxH post-subsection">Looking forward</h3><p class="ReadingDetail_reading-column__FguxH post-text">Protecting election integrity requires constant vigilance and adaptation as AI technology evolves. We remain committed to developing sophisticated testing systems, strengthening industry collaboration, and maintaining transparent communication about our findings as we work to protect democratic processes.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/elections-ai-2024" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/elections-ai-2024" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/elections-ai-2024</link>
<guid isPermaLink="false">https://anthropic.com/news/elections-ai-2024</guid>
<pubDate>Wed, 11 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Claude 3.5 Haiku on AWS Trainium2 and model distillation in Amazon Bedrock</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An illustration of a scale with a hand and a ball on one side and a lightning bolt with a hand on the other side." loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F25db0024da8370d00795ca0df70bb2a71d1b47d3-2880x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F25db0024da8370d00795ca0df70bb2a71d1b47d3-2880x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">As part of our expanded <a href="https://www.anthropic.com/news/anthropic-amazon-trainium">collaboration with AWS</a>, we’ve begun optimizing Claude models to run on <a href="https://aws.amazon.com/ai/machine-learning/trainium/">AWS Trainium2</a>, their most advanced AI chip.</p><p class="ReadingDetail_reading-column__FguxH post-text">To preview what’s possible with Trainium2, Claude 3.5 Haiku now supports latency-optimized inference in <a href="https://aws.amazon.com/bedrock/claude/">Amazon Bedrock</a>, making the model significantly faster without compromising accuracy.</p><p class="ReadingDetail_reading-column__FguxH post-text">We’re also adding support for model distillation in Amazon Bedrock, bringing the intelligence of larger Claude models to our faster and more cost-effective models.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Next-gen models on Trainium2</h2><p class="ReadingDetail_reading-column__FguxH post-text">We are collaborating with AWS to build Project Rainier—an EC2 UltraCluster of Trn2 UltraServers containing hundreds of thousands of Trainium2 chips. This cluster will deliver more than five times the computing power (in exaflops) used to train our current generation of leading AI models.</p><p class="ReadingDetail_reading-column__FguxH post-text">Trainium2 enables us to offer faster models in Amazon Bedrock, starting with Claude 3.5 Haiku which now supports latency-optimized inference in public preview. By enabling latency optimization, Claude 3.5 Haiku can deliver up to 60% faster inference speed—making it the ideal choice for use cases ranging from code completions to real-time content moderation and chatbots.</p><p class="ReadingDetail_reading-column__FguxH post-text">This faster version of Claude 3.5 Haiku, powered by Trainium2, is available in the US East (Ohio) Region via <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html">cross-region inference</a> and is offered at $1 per million input tokens and $5 per million output tokens.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Amazon Bedrock Model Distillation</h2><p class="ReadingDetail_reading-column__FguxH post-text">We’re also enabling customers to get frontier performance from Claude 3 Haiku—our most cost-effective model from the last generation. With distillation, Claude 3 Haiku can now achieve significant performance gains, reaching Claude 3.5 Sonnet-like accuracy for specific tasks—at the same price and speed of our most cost-effective model.</p><p class="ReadingDetail_reading-column__FguxH post-text">This technique transfers knowledge from the "teacher" (Claude 3.5 Sonnet) to the "student" (Claude 3 Haiku), enabling customers to run sophisticated tasks like retrieval augmented generation (RAG) and data analysis at a fraction of the cost.</p><p class="ReadingDetail_reading-column__FguxH post-text">Unlike traditional fine-tuning, which requires developers to manually craft training examples and continuously adjust parameters, Amazon Bedrock Model Distillation automates the entire process by:</p><ol class="ReadingDetail_reading-column__FguxH post-text"><li><strong>Generating synthetic training data</strong> from Claude 3.5 Sonnet</li><li><strong>Training and evaluating</strong> Claude 3 Haiku</li><li><strong>Hosting</strong> the final distilled model for inference</li></ol><p class="ReadingDetail_reading-column__FguxH post-text">Amazon Bedrock Model Distillation automatically applies different data synthesis methods—from generating similar prompts to creating new high-quality responses based on your example prompt-response pairs.</p><p class="ReadingDetail_reading-column__FguxH post-text">Distillation for Claude 3 Haiku in Amazon Bedrock is now available in preview. Learn more in the AWS <a href="https://aws.amazon.com/blogs/aws/build-faster-more-cost-efficient-highly-accurate-models-with-amazon-bedrock-model-distillation-preview/">launch blog</a> and <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/model-distillation.html">documentation</a>.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Lower prices for Claude 3.5 Haiku</h2><p class="ReadingDetail_reading-column__FguxH post-text">In addition to offering a faster version on Trainium2, customers can continue to access <a href="https://www.anthropic.com/claude/haiku">Claude 3.5 Haiku</a> on the <a href="https://console.anthropic.com/workbench">Anthropic API</a>, <a href="https://aws.amazon.com/bedrock/claude/">Amazon Bedrock</a>, and <a href="https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude">Google Cloud’s Vertex AI</a>.</p><p class="ReadingDetail_reading-column__FguxH post-text">To make this model even more accessible for a wide range of use cases, we’re lowering the price of Claude 3.5 Haiku to $0.80 per million input tokens and $4 per million output tokens across all platforms.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Get started</h2><p class="ReadingDetail_reading-column__FguxH post-text">Starting today, model distillation and the faster Claude 3.5 Haiku are available in preview in Amazon Bedrock. For developers seeking the optimal balance of price, performance, and speed, you now have expanded model options with Claude:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>Claude 3.5 Haiku with latency optimization, powered by Trainium2, for general use cases</li><li>Claude 3 Haiku, distilled with frontier performance, for high-volume, repetitive use cases</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">To get started, visit the <a href="https://signin.aws.amazon.com/signup?request_type=register">Amazon Bedrock console</a>. We can’t wait to see what you build.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/trainium2-and-distillation" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/trainium2-and-distillation" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/trainium2-and-distillation</link>
<guid isPermaLink="false">https://anthropic.com/news/trainium2-and-distillation</guid>
<pubDate>Mon, 02 Dec 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Tailor Claude’s responses to your personal style</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An illustration of three pieces of paper with different writing styles and a feather pen" loading="eager" width="2881" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Ff717ab033604c4c0531b7c1dcc52fd51c49d7c8d-2881x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Ff717ab033604c4c0531b7c1dcc52fd51c49d7c8d-2881x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text"></p><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're announcing custom styles for all <a href="http://claude.ai/">Claude.ai</a> users. Now you can tailor Claude's responses to your unique needs and workflows.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Personalize Claude chats</h2><p class="ReadingDetail_reading-column__FguxH post-text">With styles, you can customize Claude’s responses to match your communication preferences, tone, and structure. Whether you're a developer writing technical documentation, a marketer crafting specific brand guidelines, or a product team planning extensive project requirements, Claude can adapt to your preferred way of writing.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="Video_post-wrapper__PhRKk"><div class="Video_post-video-container__q6iq4"><video autoplay="" playsinline="" muted="" loop="" src="https://cdn.sanity.io/files/4zrzovbb/website/6ca179cd03e0dcd9e0c98a7005305a35a94ebefa.mov"></video></div></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">You can choose from the following: </p><ul class="ReadingDetail_reading-column__FguxH post-text"><li><strong>Formal</strong>: clear and polished responses</li><li><strong>Concise</strong>: shorter and more direct responses</li><li><strong>Explanatory</strong>: educational responses for learning new concepts</li></ul><h2 class="ReadingDetail_reading-column__FguxH post-section">Generate custom styles</h2><p class="ReadingDetail_reading-column__FguxH post-text">Beyond these preset options, Claude can generate <strong>custom styles</strong> for you. Upload sample content that reflects your preferred way of communicating, specify instructions to match your needs, and edit your preferences as they evolve.</p><div class="ReadingDetail_media-column__5aUd6"><figure class="Video_post-wrapper__PhRKk"><div class="Video_post-video-container__q6iq4"><video autoplay="" playsinline="" muted="" loop="" src="https://cdn.sanity.io/files/4zrzovbb/website/b2c2eb47ec904e289cf12675218b3c6122e4378f.mov"></video></div></figure></div><p class="ReadingDetail_reading-column__FguxH post-text">Early customers like GitLab have leveraged styles to standardize communication and decision-making in their daily workflows across a variety of scenarios.</p><blockquote class="ReadingDetail_reading-column__FguxH post-text">Claude’s ability to maintain a consistent voice while adapting to different contexts allows our team members to use styles for various use cases including writing business cases, updating user documentation, and creating and translating marketing materials. The combination of pre-built styles for quick alignment and custom styles to match different user’s preferences make it a useful tool.</blockquote><p class="ReadingDetail_reading-column__FguxH post-text">— Taylor McCaslin, product lead for AI and ML tech at GitLab</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Getting started</h2><p class="ReadingDetail_reading-column__FguxH post-text">With styles, Claude adapts to your unique context and communication choices, helping you achieve more while working in a way that feels natural to you. To get started, define your style using a preset option or customize your own in the <a href="http://claude.ai/">Claude.ai</a> chat box.</p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/styles" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/styles" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/styles</link>
<guid isPermaLink="false">https://anthropic.com/news/styles</guid>
<pubDate>Mon, 25 Nov 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Introducing the Model Context Protocol</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="An abstract illustration of critical context connecting to a central hub" loading="eager" width="2880" height="1620" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F3aabd8804251c0364cbde9d2e4be6dc8e8c2faec-2880x1620.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F3aabd8804251c0364cbde9d2e4be6dc8e8c2faec-2880x1620.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're open-sourcing the <a href="https://modelcontextprotocol.io/">Model Context Protocol</a> (MCP), a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments. Its aim is to help frontier models produce better, more relevant responses.</p><p class="ReadingDetail_reading-column__FguxH post-text">As AI assistants gain mainstream adoption, the industry has invested heavily in model capabilities, achieving rapid advances in reasoning and quality. Yet even the most sophisticated models are constrained by their isolation from data—trapped behind information silos and legacy systems. Every new data source requires its own custom implementation, making truly connected systems difficult to scale.</p><p class="ReadingDetail_reading-column__FguxH post-text">MCP addresses this challenge. It provides a universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol. The result is a simpler, more reliable way to give AI systems access to the data they need.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Model Context Protocol</h2><p class="ReadingDetail_reading-column__FguxH post-text">The Model Context Protocol is an open standard that enables developers to build secure, two-way connections between their data sources and AI-powered tools. The architecture is straightforward: developers can either expose their data through MCP servers or build AI applications (MCP clients) that connect to these servers.</p><p class="ReadingDetail_reading-column__FguxH post-text">Today, we're introducing three major components of the Model Context Protocol for developers:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>The Model Context Protocol <a href="https://github.com/modelcontextprotocol">specification and SDKs</a></li><li>Local MCP server support in the <a href="https://claude.ai/download">Claude Desktop apps</a></li><li>An <a href="https://github.com/modelcontextprotocol/servers">open-source repository</a> of MCP servers</li></ul><p class="ReadingDetail_reading-column__FguxH post-text">Claude 3.5 Sonnet is adept at quickly building MCP server implementations, making it easy for organizations and individuals to rapidly connect their most important datasets with a range of AI-powered tools. To help developers start exploring, we’re sharing pre-built MCP servers for popular enterprise systems like Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer.</p><p class="ReadingDetail_reading-column__FguxH post-text">Early adopters like Block and Apollo have integrated MCP into their systems, while development tools companies including Zed, Replit, Codeium, and Sourcegraph are working with MCP to enhance their platforms—enabling AI agents to better retrieve relevant information to further understand the context around a coding task and produce more nuanced and functional code with fewer attempts.</p><p class="ReadingDetail_reading-column__FguxH post-text">"At Block, open source is more than a development model—it’s the foundation of our work and a commitment to creating technology that drives meaningful change and serves as a public good for all,” said Dhanji R. Prasanna, Chief Technology Officer at Block. “Open technologies like the Model Context Protocol are the bridges that connect AI to real-world applications, ensuring innovation is accessible, transparent, and rooted in collaboration. We are excited to partner on a protocol and use it to build agentic systems, which remove the burden of the mechanical so people can focus on the creative.”</p><p class="ReadingDetail_reading-column__FguxH post-text">Instead of maintaining separate connectors for each data source, developers can now build against a standard protocol. As the ecosystem matures, AI systems will maintain context as they move between different tools and datasets, replacing today's fragmented integrations with a more sustainable architecture.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Getting started</h2><p class="ReadingDetail_reading-column__FguxH post-text">Developers can start building and testing MCP connectors today. All <a href="http://claude.ai/">Claude.ai</a> plans support connecting MCP servers to the Claude Desktop app.</p><p class="ReadingDetail_reading-column__FguxH post-text">Claude for Work customers can begin testing MCP servers locally, connecting Claude to internal systems and datasets. We'll soon provide developer toolkits for deploying remote production MCP servers that can serve your entire Claude for Work organization.</p><p class="ReadingDetail_reading-column__FguxH post-text">To start building:</p><ul class="ReadingDetail_reading-column__FguxH post-text"><li>Install pre-built MCP servers through the <a href="https://claude.ai/download">Claude Desktop app</a></li><li>Follow our <a href="https://modelcontextprotocol.io/quickstart">quickstart guide</a> to build your first MCP server</li><li>Contribute to our <a href="https://github.com/modelcontextprotocol">open-source repositories</a> of connectors and implementations</li></ul><h2 class="ReadingDetail_reading-column__FguxH post-section">An open community</h2><p class="ReadingDetail_reading-column__FguxH post-text">We’re committed to building MCP as a collaborative, open-source project and ecosystem, and we’re eager to hear your feedback. Whether you’re an AI tool developer, an enterprise looking to leverage existing data, or an early adopter exploring the frontier, we invite you to build the future of context-aware AI together.</p><p class="ReadingDetail_reading-column__FguxH post-text"><br></p></div></div></article><div class="PostDetail_b-social-share__DBMfH"><a href="https://twitter.com/intent/tweet?text=https://www.anthropic.com/news/model-context-protocol" target="_blank" rel="noopener" aria-label="Share on Twitter"><svg class="Icon_icon__WRMkZ" width="24" height="24" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z" fill="#191919"></path></svg></a><a href="https://www.linkedin.com/shareArticle?mini=true&amp;url=https://www.anthropic.com/news/model-context-protocol" target="_blank" rel="noopener" aria-label="Share on LinkedIn"><svg class="Icon_icon__WRMkZ" width="48" height="48" viewBox="0 0 48 48"><path d="m35.298,11.009H12.947c-1.07,0-1.938.841-1.938,1.878h0v22.471c0,1.037.869,1.879,1.939,1.879h22.35c1.071,0,1.938-.842,1.938-1.88V12.887c0-1.037-.868-1.878-1.938-1.878Zm-16.319,21.949h-3.925v-11.808h3.925v11.808Zm-1.962-13.42h-.025c-1.317,0-2.169-.907-2.169-2.04,0-1.159.877-2.04,2.221-2.04s2.168.881,2.193,2.04c0,1.133-.851,2.04-2.22,2.04Zm16.114,13.42h-3.924v-6.317c0-1.587-.568-2.67-1.988-2.67-1.085,0-1.73.731-2.013,1.436-.105.252-.13.605-.13.958v6.593h-3.924s.05-10.7,0-11.808h3.924v1.675c.522-.806,1.452-1.952,3.537-1.952,2.582,0,4.518,1.688,4.518,5.315v6.77Z" fill="#141413"></path></svg></a></div></description>
<link>https://anthropic.com/news/model-context-protocol</link>
<guid isPermaLink="false">https://anthropic.com/news/model-context-protocol</guid>
<pubDate>Sun, 24 Nov 2024 16:00:00 GMT</pubDate>
</item>
<item>
<title>Powering the next generation of AI development with AWS</title>
<description><div class="PostDetail_post-hero__KvCo2"><figure class="ImageWithCaption_e-imageWithCaption__whu3J"><img alt="amazon and anthropic logos" loading="eager" width="5760" height="3240" decoding="async" data-nimg="1" style="color:transparent" srcset="/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F253277d4ba0d1b94a52fde6d206cba743098eb61-5760x3240.png&amp;w=3840&amp;q=75 1x" src="https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F253277d4ba0d1b94a52fde6d206cba743098eb61-5760x3240.png&amp;w=3840&amp;q=75" referrerpolicy="no-referrer"></figure></div><article><div class=""><div class="ReadingDetail_detail__wf2_W"><p class="ReadingDetail_reading-column__FguxH post-text">Today we’re announcing an expansion of our collaboration with Amazon Web Services (AWS), deepening our work together to develop and deploy advanced AI systems. This expanded partnership includes a new $4 billion investment from Amazon and establishes AWS as our primary cloud and training partner. This will bring Amazon's total investment in Anthropic to $8 billion, while maintaining their position as a minority investor.</p><h2 class="ReadingDetail_reading-column__FguxH post-section">Collaboration on AWS Trainium hardware and software</h2><p class="ReadingDetail_reading-column__FguxH post-text">Anthropic is working closely with Annapurna Labs at AWS on the development and optimization of future generations of Trainium accelerators, advancing the capabilities of specialized machine learning hardware. Through deep technical collaboration, we’re writing low-level kernels that allow us to directly interface with the Trainium silicon, and contributing to the AWS Neuron software stack to strengthen Trainium. Our engineers work closely with Annapurna’s chip design team to extract maximum computational efficiency from the hardware, which we plan to leverage to train our most advanced foundation models.</p><p class="ReadingDetail_reading-column__FguxH post-text">This close hardware-software development approach, combined with the strong price-performance and massive scalability of Trainium platforms, enables us to optimize every aspect of model training from the silicon up through the |
ming42
pushed a commit
to ming42/RSSHub
that referenced
this pull request
Dec 24, 2024
commit 0a93a7486da3188e712d6446e59911077637e132 Author: Cedar <[email protected]> Date: Tue Dec 24 23:26:17 2024 +0800 feat(route): add 穆迪评级 (#17969) * feat(route): add https://www.moodysmismicrosite.com/report * Update lib/routes/moodysmismicrosite/report.ts * Update lib/routes/moodysmismicrosite/report.ts * update report.ts --------- Co-authored-by: Cedar <[email protected]> commit 131bc33637f887da8e83e22ba064eae88a90f4ac Author: DIYgod <[email protected]> Date: Tue Dec 24 16:56:52 2024 +0800 feat: update follow config commit df1b4369c023f140c21dbfa6439d36fb57c4af67 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 24 08:49:26 2024 +0000 chore(deps-dev): bump @typescript-eslint/parser from 8.18.1 to 8.18.2 (#17971) Bumps [@typescript-eslint/parser](https://github.com/typescript-eslint/typescript-eslint/tree/HEAD/packages/parser) from 8.18.1 to 8.18.2. - [Release notes](https://github.com/typescript-eslint/typescript-eslint/releases) - [Changelog](https://github.com/typescript-eslint/typescript-eslint/blob/main/packages/parser/CHANGELOG.md) - [Commits](https://github.com/typescript-eslint/typescript-eslint/commits/v8.18.2/packages/parser) --- updated-dependencies: - dependency-name: "@typescript-eslint/parser" dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit ae3ab8782dd52e0dab825d333b1f06cac163e936 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 24 08:46:50 2024 +0000 chore(deps-dev): bump @typescript-eslint/eslint-plugin (#17970) Bumps [@typescript-eslint/eslint-plugin](https://github.com/typescript-eslint/typescript-eslint/tree/HEAD/packages/eslint-plugin) from 8.18.1 to 8.18.2. - [Release notes](https://github.com/typescript-eslint/typescript-eslint/releases) - [Changelog](https://github.com/typescript-eslint/typescript-eslint/blob/main/packages/eslint-plugin/CHANGELOG.md) - [Commits](https://github.com/typescript-eslint/typescript-eslint/commits/v8.18.2/packages/eslint-plugin) --- updated-dependencies: - dependency-name: "@typescript-eslint/eslint-plugin" dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit c66fe4ea2af1bf04fd13bc8a40084354d7864e55 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 24 12:35:55 2024 +0800 chore(deps): bump entities from 5.0.0 to 6.0.0 (#17965) Bumps [entities](https://github.com/fb55/entities) from 5.0.0 to 6.0.0. - [Release notes](https://github.com/fb55/entities/releases) - [Commits](https://github.com/fb55/entities/compare/v5.0.0...v6.0.0) --- updated-dependencies: - dependency-name: entities dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 9175f05b22de5a1bb33e1f16f97ad1dda3c42b02 Author: DIYgod <[email protected]> Date: Mon Dec 23 18:16:29 2024 +0800 feat: add obsidian plugins route commit db9d5cee3c93a40a8b13624a6d009218653d8e0d Author: DIYgod <[email protected]> Date: Mon Dec 23 17:51:45 2024 +0800 feat: add home-assistant route commit eda56eafbc48460e2ec074e52c725a12eb33b56c Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 23 08:37:26 2024 +0000 chore(deps-dev): bump discord-api-types from 0.37.112 to 0.37.114 (#17966) Bumps [discord-api-types](https://github.com/discordjs/discord-api-types) from 0.37.112 to 0.37.114. - [Release notes](https://github.com/discordjs/discord-api-types/releases) - [Changelog](https://github.com/discordjs/discord-api-types/blob/main/CHANGELOG.md) - [Commits](https://github.com/discordjs/discord-api-types/compare/0.37.112...0.37.114) --- updated-dependencies: - dependency-name: discord-api-types dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit e74c1ef7370f0347b3b8f96ddc97d234319d47cc Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 23 08:35:32 2024 +0000 chore(deps): bump ioredis from 5.4.1 to 5.4.2 (#17964) Bumps [ioredis](https://github.com/luin/ioredis) from 5.4.1 to 5.4.2. - [Release notes](https://github.com/luin/ioredis/releases) - [Changelog](https://github.com/redis/ioredis/blob/main/CHANGELOG.md) - [Commits](https://github.com/luin/ioredis/compare/v5.4.1...v5.4.2) --- updated-dependencies: - dependency-name: ioredis dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit dea52d806146098a35a15675df9d2d248dc93db0 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 23 08:35:23 2024 +0000 chore(deps-dev): bump eslint-plugin-n from 17.15.0 to 17.15.1 (#17963) Bumps [eslint-plugin-n](https://github.com/eslint-community/eslint-plugin-n) from 17.15.0 to 17.15.1. - [Release notes](https://github.com/eslint-community/eslint-plugin-n/releases) - [Changelog](https://github.com/eslint-community/eslint-plugin-n/blob/master/CHANGELOG.md) - [Commits](https://github.com/eslint-community/eslint-plugin-n/compare/v17.15.0...v17.15.1) --- updated-dependencies: - dependency-name: eslint-plugin-n dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit e3c2f009dcd6f3475760d06a28aa8fc490426807 Author: Ethan Shen <[email protected]> Date: Mon Dec 23 11:46:11 2024 +0800 feat(route): add 游侠网大侠号 (#17961) * feat(route): add 游侠网大侠号 * fix typo commit dc32d68fc380c538d94e410cb0057bd322772675 Author: ueiu <[email protected]> Date: Mon Dec 23 01:40:50 2024 +0800 fix(blizzard): 新闻路由报错不能使用 (#17956) * fix(blizzard): 新闻路由报错不能使用 * fix(blizzard): 修改 cache key commit d525693969caefc99ffc8feed007883702fee1a5 Author: Jianjun Xiao <[email protected]> Date: Mon Dec 23 00:23:43 2024 +0800 feat(route): Add RSS feed for latest news from Anthropic (a leading LLM developer). (#17960) * Add Anthropic news * Update lib/routes/anthropic/news.ts --------- commit 8a58ca1b776e80f854ed6d328724ebd827539d09 Author: ueiu <[email protected]> Date: Sun Dec 22 20:11:34 2024 +0800 chore(shu): 给上海大学相关 rss 添加图标 (#17958) commit 0a307e83c6c3499ba11a31cb626dc838d196a68b Author: junfengP <[email protected]> Date: Sat Dec 21 22:28:38 2024 +0800 feat(route): implement 新电影天堂 route with content scraping and caching (#17945) * feat(route): implement 新电影天堂 route with content scraping and caching - Added a new route for 电影天堂 (Movie Heaven) that scrapes movie data from ygdy8.net. - Implemented content loading with caching to optimize performance. - Created a namespace for the route with relevant metadata. - Removed the deprecated JavaScript version of the route to streamline the codebase. * fix(route): update 电影天堂 route for improved content scraping --------- Co-authored-by: junfengP <[email protected]> commit a5f25ac54e38a837e009d881373a451141e1041e Author: Andvari <[email protected]> Date: Sat Dec 21 14:15:41 2024 +0800 feat(route/copernicium): Add default (#17828) * feat(route/copernicium): Add default * Update index.ts * Update index.ts * Update lib/routes/copernicium/index.ts Co-authored-by: Tony <[email protected]> * Update index.ts * Update index.ts * Update lib/routes/copernicium/index.ts --------- commit 43a505f19a91d0a0c446973e4b51c6623e357145 Author: Tony <[email protected]> Date: Sat Dec 21 05:02:33 2024 +0800 fix(hket): select main listing container (#17954) commit efb663a332b73e73bbbfc37748def8dea2b3e591 Author: Tony <[email protected]> Date: Sat Dec 21 04:32:33 2024 +0800 fix(behance): embed (#17953) commit d6fd26238e8f0396e0020513a12a85b6013fabb9 Author: Bubu <[email protected]> Date: Sat Dec 21 03:38:01 2024 +0800 feat(route): 添加 海南省工业和信息化厅-通知公告 (#17923) * feat(route): 添加 海南省工业和信息化厅-通知公告 * fix: add namespace --------- commit 40e3dad6f4c24f7bb971b7425454f467eaefc668 Author: Cedar <[email protected]> Date: Sat Dec 21 03:09:47 2024 +0800 feat(route): 增加几米创作最新消息 (#17924) * feat(route): 增加几米创作最新消息 * chore: update namespace * Update lib/routes/jimmyspa/namespace.ts Co-authored-by: Tony <[email protected]> * chore: update jimmyspa * chore: cache the content --------- Co-authored-by: Cedar <[email protected]> commit e20c7a02c6a1b538588dc11b8816f34ec8fb86f9 Author: cnk <[email protected]> Date: Sat Dec 21 02:31:41 2024 +0800 fix(163/music/djradio): add option for info (#17944) * fix(163/music/djradio): remove duplicate part with embedded tag * fix(163/music/djradio): add option for info * fix: use cache --------- commit dd085029551066d2d91475d13e5b6e1546be5d66 Author: DIYgod <[email protected]> Date: Fri Dec 20 22:50:49 2024 +0800 feat: add follow limit config commit d46052f2e6dec1968356079116f76c7941ff6eb8 Author: Tsuyumi <[email protected]> Date: Fri Dec 20 19:00:03 2024 +0800 fix(bilibili): missing image (#17952) commit c69b4af7ba110a2741f8ac8ed1929c3365c90cf1 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri Dec 20 16:41:39 2024 +0800 chore(deps-dev): bump discord-api-types from 0.37.111 to 0.37.112 (#17949) Bumps [discord-api-types](https://github.com/discordjs/discord-api-types) from 0.37.111 to 0.37.112. - [Release notes](https://github.com/discordjs/discord-api-types/releases) - [Changelog](https://github.com/discordjs/discord-api-types/blob/main/CHANGELOG.md) - [Commits](https://github.com/discordjs/discord-api-types/compare/0.37.111...0.37.112) --- updated-dependencies: - dependency-name: discord-api-types dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 7e352d54653b712ef7f1de841be616b5858c3566 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri Dec 20 08:16:57 2024 +0000 chore(deps-dev): bump node-network-devtools from 1.0.23 to 1.0.24 (#17948) Bumps [node-network-devtools](https://github.com/GrinZero/node-network-devtools) from 1.0.23 to 1.0.24. - [Release notes](https://github.com/GrinZero/node-network-devtools/releases) - [Commits](https://github.com/GrinZero/node-network-devtools/compare/v1.0.23...v1.0.24) --- updated-dependencies: - dependency-name: node-network-devtools dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit e20bd720297b6c3cd594d9a47259cbbb6a0871d5 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri Dec 20 08:16:41 2024 +0000 chore(deps): bump tldts from 6.1.68 to 6.1.69 (#17947) Bumps [tldts](https://github.com/remusao/tldts) from 6.1.68 to 6.1.69. - [Release notes](https://github.com/remusao/tldts/releases) - [Changelog](https://github.com/remusao/tldts/blob/master/CHANGELOG.md) - [Commits](https://github.com/remusao/tldts/compare/v6.1.68...v6.1.69) --- updated-dependencies: - dependency-name: tldts dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit f5fc827bb8bda196844330f5907422c5d8d680ec Author: Tony <[email protected]> Date: Fri Dec 20 02:26:15 2024 +0800 fix(behance): fix image url (#17942) commit b27218c2d2a49974e8181528ab3706562ff12d95 Author: Ethan Shen <[email protected]> Date: Fri Dec 20 01:25:33 2024 +0800 feat(route): add 艾瑞咨询研究图表 (#17940) commit 40cc40b5368ec93d16d4e524f7c70d06bc463f52 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 19 22:31:26 2024 +0800 chore(deps): bump @opentelemetry/sdk-metrics from 1.29.0 to 1.30.0 (#17933) Bumps [@opentelemetry/sdk-metrics](https://github.com/open-telemetry/opentelemetry-js) from 1.29.0 to 1.30.0. - [Release notes](https://github.com/open-telemetry/opentelemetry-js/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-js/blob/main/CHANGELOG.md) - [Commits](https://github.com/open-telemetry/opentelemetry-js/compare/v1.29.0...v1.30.0) --- updated-dependencies: - dependency-name: "@opentelemetry/sdk-metrics" dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit acb553cdc5a83b453107679a5ea23aca2d0a6041 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 19 22:20:17 2024 +0800 chore(deps): bump @opentelemetry/exporter-prometheus (#17932) Bumps [@opentelemetry/exporter-prometheus](https://github.com/open-telemetry/opentelemetry-js) from 0.56.0 to 0.57.0. - [Release notes](https://github.com/open-telemetry/opentelemetry-js/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-js/blob/main/CHANGELOG.md) - [Commits](https://github.com/open-telemetry/opentelemetry-js/compare/experimental/v0.56.0...experimental/v0.57.0) --- updated-dependencies: - dependency-name: "@opentelemetry/exporter-prometheus" dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit fa68d30093a223a82ac2e5d41c78bcaec4a76735 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 19 22:20:01 2024 +0800 chore(deps): bump @opentelemetry/sdk-trace-base from 1.29.0 to 1.30.0 (#17934) Bumps [@opentelemetry/sdk-trace-base](https://github.com/open-telemetry/opentelemetry-js) from 1.29.0 to 1.30.0. - [Release notes](https://github.com/open-telemetry/opentelemetry-js/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-js/blob/main/CHANGELOG.md) - [Commits](https://github.com/open-telemetry/opentelemetry-js/compare/v1.29.0...v1.30.0) --- updated-dependencies: - dependency-name: "@opentelemetry/sdk-trace-base" dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 520fa1ac1f2ed795a908fef20eac9dc48bd413aa Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 19 22:19:51 2024 +0800 chore(deps): bump @opentelemetry/resources from 1.29.0 to 1.30.0 (#17938) Bumps [@opentelemetry/resources](https://github.com/open-telemetry/opentelemetry-js) from 1.29.0 to 1.30.0. - [Release notes](https://github.com/open-telemetry/opentelemetry-js/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-js/blob/main/CHANGELOG.md) - [Commits](https://github.com/open-telemetry/opentelemetry-js/compare/v1.29.0...v1.30.0) --- updated-dependencies: - dependency-name: "@opentelemetry/resources" dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit f3fcf29e3a83e9ed3a3c7904aa6b52566d630c12 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 19 21:50:41 2024 +0800 chore(deps-dev): bump globals from 15.13.0 to 15.14.0 (#17931) Bumps [globals](https://github.com/sindresorhus/globals) from 15.13.0 to 15.14.0. - [Release notes](https://github.com/sindresorhus/globals/releases) - [Commits](https://github.com/sindresorhus/globals/compare/v15.13.0...v15.14.0) --- updated-dependencies: - dependency-name: globals dependency-type: direct:development update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit d0b36f40cc24f5fde583d76ce58d46cc9d23f351 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 19 21:49:48 2024 +0800 chore(deps): bump @opentelemetry/exporter-trace-otlp-http (#17935) Bumps [@opentelemetry/exporter-trace-otlp-http](https://github.com/open-telemetry/opentelemetry-js) from 0.56.0 to 0.57.0. - [Release notes](https://github.com/open-telemetry/opentelemetry-js/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-js/blob/main/CHANGELOG.md) - [Commits](https://github.com/open-telemetry/opentelemetry-js/compare/experimental/v0.56.0...experimental/v0.57.0) --- updated-dependencies: - dependency-name: "@opentelemetry/exporter-trace-otlp-http" dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 473c3a6d0812c6a033e86df0f784c0e0067a641a Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 19 21:48:24 2024 +0800 chore(deps): bump sanitize-html from 2.13.1 to 2.14.0 (#17936) Bumps [sanitize-html](https://github.com/apostrophecms/sanitize-html) from 2.13.1 to 2.14.0. - [Changelog](https://github.com/apostrophecms/sanitize-html/blob/main/CHANGELOG.md) - [Commits](https://github.com/apostrophecms/sanitize-html/compare/2.13.1...2.14.0) --- updated-dependencies: - dependency-name: sanitize-html dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 911482b43cea21ad006b13325282dc19f8d76d60 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 19 08:34:20 2024 +0000 chore(deps): bump cross-spawn from 7.0.3 to 7.0.6 (#17939) Bumps [cross-spawn](https://github.com/moxystudio/node-cross-spawn) from 7.0.3 to 7.0.6. - [Changelog](https://github.com/moxystudio/node-cross-spawn/blob/master/CHANGELOG.md) - [Commits](https://github.com/moxystudio/node-cross-spawn/compare/v7.0.3...v7.0.6) --- updated-dependencies: - dependency-name: cross-spawn dependency-type: indirect ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 87cf38bf04edaef67fcbdeb60e4d439d6eb77e58 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 19 08:29:54 2024 +0000 chore(deps-dev): bump @vercel/nft from 0.27.9 to 0.27.10 (#17930) Bumps [@vercel/nft](https://github.com/vercel/nft) from 0.27.9 to 0.27.10. - [Release notes](https://github.com/vercel/nft/releases) - [Commits](https://github.com/vercel/nft/compare/0.27.9...0.27.10) --- updated-dependencies: - dependency-name: "@vercel/nft" dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 5891dd8cc6021086e74405129b376cf1ba09f681 Author: Tony <[email protected]> Date: Thu Dec 19 12:02:46 2024 +0800 fix(behance): use GraphQL instead (#17928) commit 0c2bd405436e3dbc55a3471e6f1ccc034e6601db Author: Yunan Wang <[email protected]> Date: Wed Dec 18 21:15:36 2024 +0100 feat: add HiringCafe route (#17858) * feat(hiringcafe): add namespace.ts Signed-off-by: mintyfrankie <[email protected]> * feat(hiringcafe): add jobs endpoint Signed-off-by: mintyfrankie <[email protected]> * fix(hiringcafe): fix ESLint warning Signed-off-by: mintyfrankie <[email protected]> * fix: apply suggestions from code review Co-authored-by: Tony <[email protected]> * fix: accept suggestions from code review Signed-off-by: mintyfrankie <[email protected]> * refactor: modularize art template and sub-functions Signed-off-by: mintyfrankie <[email protected]> * feat(hiringcafe): add API interfaces Signed-off-by: mintyfrankie <[email protected]> * fix: resolve __dirname error Signed-off-by: mintyfrankie <[email protected]> * refactor: change API payload and interfaces to match upstream changes Signed-off-by: mintyfrankie <[email protected]> * refactor: add type safety and error handling Signed-off-by: mintyfrankie <[email protected]> * Apply suggestions from code review Co-authored-by: Tony <[email protected]> * fix: resolve ESLint error Signed-off-by: mintyfrankie <[email protected]> * fix: use hiring.cafe --------- Signed-off-by: mintyfrankie <[email protected]> commit 5ffc73e8e240b6d2f401d13b002c3bd97ef94ca6 Author: Yoki Wong <[email protected]> Date: Wed Dec 18 23:52:17 2024 +0800 feat(route): add route for Society journals (#17899) * Sociology Studies Jounal * Sociology Stuides Journal * delete error description * fix category name * Add router for society issues * fix routes according to docs --------- Co-authored-by: CNYoki <[email protected]> commit 2aab5047b7612a46ac523541a88b7c0a490927d8 Author: DIYgod <[email protected]> Date: Wed Dec 18 23:16:28 2024 +0800 feat: follow config commit 4bec9ed52f146750418a7c5c777cbd898332611a Author: pseudoyu <[email protected]> Date: Tue Dec 17 22:24:36 2024 +0800 fix(route/xueqiu): mark anti-crawler and improve error handling commit 751cd2ef17870ee35f9902b71fc35f412ca189f7 Author: Bubu <[email protected]> Date: Tue Dec 17 22:14:03 2024 +0800 feat(route): add techsir.com (#17840) commit eed7d3b255bb2fb9c25d81c2009050fd65f7369b Author: pseudoyu <[email protected]> Date: Tue Dec 17 21:48:44 2024 +0800 fix(route/smzdm): remove yearly hot and change options commit 6a06261c9a8aa4ba5d8bc78e4e715dd396fdc2c9 Author: pseudoyu <[email protected]> Date: Tue Dec 17 21:18:08 2024 +0800 fix(route/smzdm): update handler to use ofetch and improve data structure commit 0bc647c8bea93e3bf64c4efb56681211c77a8a89 Author: Bubu <[email protected]> Date: Tue Dec 17 21:13:22 2024 +0800 feat(route): APNIC Blog 全文 RSS (#17839) commit eb95f5bd2259e11dc9b411bad13616efa0a4f752 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 17 05:12:14 2024 -0800 chore(deps-dev): bump @typescript-eslint/parser from 8.18.0 to 8.18.1 (#17920) Bumps [@typescript-eslint/parser](https://github.com/typescript-eslint/typescript-eslint/tree/HEAD/packages/parser) from 8.18.0 to 8.18.1. - [Release notes](https://github.com/typescript-eslint/typescript-eslint/releases) - [Changelog](https://github.com/typescript-eslint/typescript-eslint/blob/main/packages/parser/CHANGELOG.md) - [Commits](https://github.com/typescript-eslint/typescript-eslint/commits/v8.18.1/packages/parser) --- updated-dependencies: - dependency-name: "@typescript-eslint/parser" dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit efbaf066717fd04c00f43419cf6df4bdab993b1e Author: pseudoyu <[email protected]> Date: Tue Dec 17 21:04:46 2024 +0800 fix(route/bbc): bbc news 404 url issue commit 4785a0fe0da84f61653892b94e7cc2bc107321ae Author: 孙茂胤 (Sun, Maoyin) <[email protected]> Date: Tue Dec 17 14:04:07 2024 +0100 docs: fix tip wrapper in document (#17917) The extra indentation causes failure of parsing during Docs rendering. The parser will wrap the latter content until the next `:::` into tips commit 32358de408aa83e1cdf65d72da4881d668770eed Author: Cedar <[email protected]> Date: Tue Dec 17 20:57:33 2024 +0800 feat(route): 增加标普全球评级 (#17913) * feat(route): 增加标普全球评级 增加标普全球评级多语言发布稿件的获取，https://www.spglobal.com/ratings/zh、https://www.spglo bal.com/ratings/en等 * chore: provide an absolute URL. * chore: fix incorrect URL --------- Co-authored-by: Cedar <[email protected]> commit 8a3f053af6629f46e672592d16301bf0c6f66829 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 17 08:41:48 2024 +0000 chore(deps): bump @scalar/hono-api-reference from 0.5.164 to 0.5.165 (#17921) Bumps [@scalar/hono-api-reference](https://github.com/scalar/scalar/tree/HEAD/packages/hono-api-reference) from 0.5.164 to 0.5.165. - [Changelog](https://github.com/scalar/scalar/blob/main/packages/hono-api-reference/CHANGELOG.md) - [Commits](https://github.com/scalar/scalar/commits/HEAD/packages/hono-api-reference) --- updated-dependencies: - dependency-name: "@scalar/hono-api-reference" dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit dc8e254d9bf216dfd5b9f9a124dbebab2e65e5f0 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 17 08:41:03 2024 +0000 chore(deps-dev): bump @typescript-eslint/eslint-plugin (#17918) Bumps [@typescript-eslint/eslint-plugin](https://github.com/typescript-eslint/typescript-eslint/tree/HEAD/packages/eslint-plugin) from 8.18.0 to 8.18.1. - [Release notes](https://github.com/typescript-eslint/typescript-eslint/releases) - [Changelog](https://github.com/typescript-eslint/typescript-eslint/blob/main/packages/eslint-plugin/CHANGELOG.md) - [Commits](https://github.com/typescript-eslint/typescript-eslint/commits/v8.18.1/packages/eslint-plugin) --- updated-dependencies: - dependency-name: "@typescript-eslint/eslint-plugin" dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 5b4173403a39a6dcae6808dc748719466c7052a0 Author: Bubu <[email protected]> Date: Tue Dec 17 12:02:46 2024 +0800 feat(route): 新增 新芽网 (#17841) commit 0e1762f392057711304754cee611f57e3471b8ef Author: Bubu <[email protected]> Date: Tue Dec 17 11:46:30 2024 +0800 feat(route): 新增 和讯创投 (#17842) commit 9371b8977924983de51c7d101c91c741305d0d52 Author: Andvari <[email protected]> Date: Tue Dec 17 09:40:34 2024 +0800 fix(route/zaobao): Adapt to new SG layout (#17902) * fix(route/zaobao): Replace `got` while fetching article list * . * Update util.ts * Update util.ts commit 6e5738e4dce8464e4b072ba59dc1d4a02f6fb036 Author: Bubu <[email protected]> Date: Tue Dec 17 02:48:58 2024 +0800 feat(route): 新增 挖贝网 (#17903) commit fcd4b8f8f6162ecebc424047f75d28274a17b25e Author: pseudoyu <[email protected]> Date: Tue Dec 17 00:37:29 2024 +0800 feat(route/twitter): support third-party twitter keywords api commit 07c1e883682a3907b71272a854e0775a8845b885 Author: gz4zzxc <[email protected]> Date: Tue Dec 17 00:06:50 2024 +0800 feat(route): add GDUFS news route and GDUFS xwxy news (#17822) * feat(route):add GDUFS news route && GDUFS xwxy news * fix(xwxy-news): update authors extraction to use toArray() method * fix(xwxy-news): improve article detail fetching by explicitly passing item to fetchArticleDetail * fix(xwxy-news): reorder authors extraction to trim text after converting to array * refactor(routes): remove protocol from URLs in gdufs/news.ts Co-authored-by: Tony <[email protected]> * refactor(routes): simplify route name in gdufs/news.ts Co-authored-by: Tony <[email protected]> * refactor(routes): remove protocol from URLs in gdufs/news.ts Co-authored-by: Tony <[email protected]> * refactor(routes): simplify route name in gdufs/xwxy-news.ts Co-authored-by: Tony <[email protected]> * refactor(routes): use cache to optimize article content fetching and author extraction in gdufs/news.ts and gdufs/xwxy-news.ts * refactor(routes): Cache the entire item object in /gdufs/news & /gdufs/xwxy-news route --------- commit 030924a3ef715be0c671a91c51b706dec996bfe7 Author: FanofZY <[email protected]> Date: Mon Dec 16 22:44:11 2024 +0800 feat(route): add route for university: whu swrh (#17689) * feat(route): add route for university: whu swrh (#1) * feat(route): add route for university: whu swrh * Update swrh.ts * Update swrh.ts fixed the weixin url * update swrh.ts for html and weixin content * feat(route): add route for university: whu swrh * update for fetching pubdate. commit 8494da4f1a3a5b6df4d697954f3bcb7ba5031cab Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 16 05:38:58 2024 -0800 chore(deps-dev): bump eslint from 9.16.0 to 9.17.0 (#17905) Bumps [eslint](https://github.com/eslint/eslint) from 9.16.0 to 9.17.0. - [Release notes](https://github.com/eslint/eslint/releases) - [Changelog](https://github.com/eslint/eslint/blob/main/CHANGELOG.md) - [Commits](https://github.com/eslint/eslint/compare/v9.16.0...v9.17.0) --- updated-dependencies: - dependency-name: eslint dependency-type: direct:development update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 90dbb999f084cde903aba89e3b8658c8394c913f Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 16 05:08:13 2024 -0800 chore(deps-dev): bump @eslint/js from 9.16.0 to 9.17.0 (#17909) Bumps [@eslint/js](https://github.com/eslint/eslint/tree/HEAD/packages/js) from 9.16.0 to 9.17.0. - [Release notes](https://github.com/eslint/eslint/releases) - [Changelog](https://github.com/eslint/eslint/blob/main/CHANGELOG.md) - [Commits](https://github.com/eslint/eslint/commits/v9.17.0/packages/js) --- updated-dependencies: - dependency-name: "@eslint/js" dependency-type: direct:development update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit db967716a5d42ded244026dde33f66185c817b4c Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 16 05:07:43 2024 -0800 chore(deps): bump @scalar/hono-api-reference from 0.5.163 to 0.5.164 (#17906) Bumps [@scalar/hono-api-reference](https://github.com/scalar/scalar/tree/HEAD/packages/hono-api-reference) from 0.5.163 to 0.5.164. - [Changelog](https://github.com/scalar/scalar/blob/main/packages/hono-api-reference/CHANGELOG.md) - [Commits](https://github.com/scalar/scalar/commits/HEAD/packages/hono-api-reference) --- updated-dependencies: - dependency-name: "@scalar/hono-api-reference" dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit e74ab3b85c59d1f2bcd43555ec7ba43784835588 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 16 05:07:12 2024 -0800 chore(deps): bump tldts from 6.1.67 to 6.1.68 (#17907) Bumps [tldts](https://github.com/remusao/tldts) from 6.1.67 to 6.1.68. - [Release notes](https://github.com/remusao/tldts/releases) - [Changelog](https://github.com/remusao/tldts/blob/master/CHANGELOG.md) - [Commits](https://github.com/remusao/tldts/compare/v6.1.67...v6.1.68) --- updated-dependencies: - dependency-name: tldts dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 0323d68d8db3fa60e188f224627aec88cdaa13a9 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 16 08:29:55 2024 +0000 chore(deps): bump hono from 4.6.13 to 4.6.14 (#17908) Bumps [hono](https://github.com/honojs/hono) from 4.6.13 to 4.6.14. - [Release notes](https://github.com/honojs/hono/releases) - [Commits](https://github.com/honojs/hono/compare/v4.6.13...v4.6.14) --- updated-dependencies: - dependency-name: hono dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 8952721ce66ba01b884e3b6468bb26517d7c8810 Author: ueiu <[email protected]> Date: Sun Dec 15 20:35:58 2024 +0800 fix(route/egsea): 修复e快选因部分条目没有分类导致抓取失败 (#17895) commit 655f7ac52f12b465b0b65f17613e51278719bea8 Author: CaoMeiYouRen <[email protected]> Date: Sat Dec 14 21:02:44 2024 +0800 fix(github/repos): 添加日期解析功能并优化数据处理 (#17893) - 引入 `parseDate` 工具函数用于解析日期 - 根据排序类型（updated 或 pushed）动态设置发布日期 - 优化数据过滤和映射逻辑，减少嵌套层级 commit 3a8d34ee3f8cc38907296e74e923754297e249d4 Author: cnk <[email protected]> Date: Sat Dec 14 08:19:55 2024 +0800 feat(route/hrbust): add 哈尔滨理工大学新闻网 (#17870) * feat(route): add 哈尔滨理工大学新闻网 * fix(route): fix title for 哈尔滨理工大学新闻网 * fix(route): using v2 Route Standard * fix(route/hrbust): fix with review commit 7fda7a25399af8236d7a32154a76addffaeffa07 Author: DaoXuan <[email protected]> Date: Sat Dec 14 02:22:17 2024 +0800 feat(route): 新增博客道宣的窝 (#17890) * 尝试 * /daoxuan * 修改example * Update lib/routes/daoxuan/rss.ts --------- commit a5deeb3729330bffcb6dfe0c65d3d4e0263ea059 Author: Ethan Shen <[email protected]> Date: Sat Dec 14 00:23:13 2024 +0800 fix(types): fix route type definitions for different languages (#17888) commit 6446448340fdbd7f0e2e2ea1971f46382420651e Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri Dec 13 05:10:50 2024 -0800 chore(deps): bump http-cookie-agent from 6.0.7 to 6.0.8 (#17883) Bumps [http-cookie-agent](https://github.com/3846masa/http-cookie-agent) from 6.0.7 to 6.0.8. - [Release notes](https://github.com/3846masa/http-cookie-agent/releases) - [Changelog](https://github.com/3846masa/http-cookie-agent/blob/main/CHANGELOG.md) - [Commits](https://github.com/3846masa/http-cookie-agent/compare/v6.0.7...v6.0.8) --- updated-dependencies: - dependency-name: http-cookie-agent dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 352cac955d44ace6bae478ce40ab569b50ef97d0 Author: X <[email protected]> Date: Thu Dec 12 22:01:10 2024 +0800 fix(route/xiaohongshu): remove items url extra slash (#17878) commit 2a90282da9154cd702f1f5d5c1bb084569cd7d48 Author: Geraldxm <[email protected]> Date: Thu Dec 12 18:52:24 2024 +0800 feat(route): 充电头网 (#17860) * feat(route): 充电头网 * use REST API * remove cache; use date_gmt * Update lib/routes/chongdiantou/index.ts commit 6ffabac65457b23c877069272b73e371c843612c Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 12 01:27:27 2024 -0800 chore(deps-dev): bump @types/node from 22.10.1 to 22.10.2 (#17876) Bumps [@types/node](https://github.com/DefinitelyTyped/DefinitelyTyped/tree/HEAD/types/node) from 22.10.1 to 22.10.2. - [Release notes](https://github.com/DefinitelyTyped/DefinitelyTyped/releases) - [Commits](https://github.com/DefinitelyTyped/DefinitelyTyped/commits/HEAD/types/node) --- updated-dependencies: - dependency-name: "@types/node" dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 0823df23c2dfb908f7a399876dadbe1d4b636da3 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 12 08:31:28 2024 +0000 chore(deps): bump tldts from 6.1.66 to 6.1.67 (#17877) Bumps [tldts](https://github.com/remusao/tldts) from 6.1.66 to 6.1.67. - [Release notes](https://github.com/remusao/tldts/releases) - [Changelog](https://github.com/remusao/tldts/blob/master/CHANGELOG.md) - [Commits](https://github.com/remusao/tldts/compare/v6.1.66...v6.1.67) --- updated-dependencies: - dependency-name: tldts dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 6deb2f57cf9ac96a87f1b8009f790ddd2af466af Author: pseudoyu <[email protected]> Date: Thu Dec 12 13:53:50 2024 +0800 feat(route/twitter): support twitter list third-party api commit 127b45637a173ea1b77659ed73e3e5b881d9088a Author: Zihao Diao <[email protected]> Date: Wed Dec 11 17:59:34 2024 -0500 feat(picnob): fix Picnob Instagram Reels video playback. (#17874) * feat(picnob): fix Picnob igtv video playback. Signed-off-by: Zihao Diao <[email protected]> * feat(picnob): fix Picnob Instagram Reels video playback Signed-off-by: Zihao Diao <[email protected]> --------- Signed-off-by: Zihao Diao <[email protected]> commit 7c5a04fa7a06a61708215d374eb00a1d9a5355e0 Author: Yun Du <[email protected]> Date: Thu Dec 12 01:34:07 2024 +0800 feat(route): add cool paper first-author infomation (#17857) * [update] update cool paper * [add] add author of cool paper commit e1c5e635fc8bfbbb7884520aababf11f1c485960 Author: Ethan Shen <[email protected]> Date: Thu Dec 12 01:05:57 2024 +0800 feat(route): add 中国疾病预防控制中心 (#17873) commit 91399b3c275ac9e62ac72a58763130b76340ec9b Author: Geraldxm <[email protected]> Date: Wed Dec 11 23:39:23 2024 +0800 feat: get full text for routes/qbitai/category (#17740) commit 7a36be760e2239b1914a6cdf7ffb946c21bf6f17 Author: 天火流光 <[email protected]> Date: Wed Dec 11 23:18:37 2024 +0800 feat: add routes (#17792) * feat: add routes Signed-off-by: 天火流光 <[email protected]> * feat: add routes * revert: "feat: add routes" This reverts commit 0bdbc3b0ee3ce536e6a02d1e8608bc85a19b42c3. * fix: revert * feat: add routes Signed-off-by: 天火流光 <[email protected]> * 🐞 fix: Wait page to load Signed-off-by: 天火流光 <[email protected]> * fix: improve video description --------- Signed-off-by: 天火流光 <[email protected]> commit 936045ae97655f76884f5a9a4f43c3dac4986afd Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Wed Dec 11 22:02:57 2024 +0800 chore(deps): bump zod from 3.23.8 to 3.24.1 (#17865) Bumps [zod](https://github.com/colinhacks/zod) from 3.23.8 to 3.24.1. - [Release notes](https://github.com/colinhacks/zod/releases) - [Changelog](https://github.com/colinhacks/zod/blob/main/CHANGELOG.md) - [Commits](https://github.com/colinhacks/zod/compare/v3.23.8...v3.24.1) --- updated-dependencies: - dependency-name: zod dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 107cc039516295bbe09a8462205a242651fa25e8 Author: Bubu <[email protected]> Date: Wed Dec 11 21:11:31 2024 +0800 feat(route): 添加 白鲸出海 (#17845) * feat(route): 添加白鲸出海 * Update lib/routes/baijing/index.ts * Update lib/routes/baijing/index.ts commit 3cb7925770db41fb4f8e3cf3a70e318a0f7bb648 Author: Bubu <[email protected]> Date: Wed Dec 11 20:42:13 2024 +0800 feat(route): 新增 i黑马网 (#17843) commit a49355bc40086674882edeb482f3f2ad4ec9d334 Author: karasu <[email protected]> Date: Wed Dec 11 17:17:59 2024 +0800 fix(route): filter out invalid subscriptions (#17861) commit ab760e9305fb45dd2fea118826e330ec485215bb Author: lidashuang <[email protected]> Date: Wed Dec 11 17:13:41 2024 +0800 fix(/mittrchina/breaking): TypeError: Cannot read properties of null (reading 'content') (#17855) * feat(route): add amz123.com * fix(/mittrchina/breaking): check var null TypeError: Cannot read properties of null (reading 'content') * fix(/mittrchina): breaking url null properties commit c876e39f23a36324aebd5fd60ab8b0b88835c9e6 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Wed Dec 11 08:51:33 2024 +0000 chore(deps): bump http-cookie-agent from 6.0.6 to 6.0.7 (#17868) Bumps [http-cookie-agent](https://github.com/3846masa/http-cookie-agent) from 6.0.6 to 6.0.7. - [Release notes](https://github.com/3846masa/http-cookie-agent/releases) - [Changelog](https://github.com/3846masa/http-cookie-agent/blob/main/CHANGELOG.md) - [Commits](https://github.com/3846masa/http-cookie-agent/compare/v6.0.6...v6.0.7) --- updated-dependencies: - dependency-name: http-cookie-agent dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit a15355ea36ff9697785f7d26f4f9c3c28538914c Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Wed Dec 11 08:50:08 2024 +0000 chore(deps): bump rfc4648 from 1.5.3 to 1.5.4 (#17866) Bumps [rfc4648](https://github.com/swansontec/rfc4648.js) from 1.5.3 to 1.5.4. - [Release notes](https://github.com/swansontec/rfc4648.js/releases) - [Changelog](https://github.com/swansontec/rfc4648.js/blob/master/CHANGELOG.md) - [Commits](https://github.com/swansontec/rfc4648.js/commits) --- updated-dependencies: - dependency-name: rfc4648 dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit f9165acccc6952162da93abc1a5f7fca5d916415 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Wed Dec 11 08:49:58 2024 +0000 chore(deps-dev): bump lint-staged from 15.2.10 to 15.2.11 (#17863) Bumps [lint-staged](https://github.com/lint-staged/lint-staged) from 15.2.10 to 15.2.11. - [Release notes](https://github.com/lint-staged/lint-staged/releases) - [Changelog](https://github.com/lint-staged/lint-staged/blob/master/CHANGELOG.md) - [Commits](https://github.com/lint-staged/lint-staged/compare/v15.2.10...v15.2.11) --- updated-dependencies: - dependency-name: lint-staged dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 32d09356f3b6f446d0f9bb498dbe7800df4a06c1 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Wed Dec 11 08:49:46 2024 +0000 chore(deps-dev): bump @stylistic/eslint-plugin from 2.12.0 to 2.12.1 (#17862) Bumps [@stylistic/eslint-plugin](https://github.com/eslint-stylistic/eslint-stylistic/tree/HEAD/packages/eslint-plugin) from 2.12.0 to 2.12.1. - [Release notes](https://github.com/eslint-stylistic/eslint-stylistic/releases) - [Changelog](https://github.com/eslint-stylistic/eslint-stylistic/blob/main/CHANGELOG.md) - [Commits](https://github.com/eslint-stylistic/eslint-stylistic/commits/v2.12.1/packages/eslint-plugin) --- updated-dependencies: - dependency-name: "@stylistic/eslint-plugin" dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit e7647392dc290bdc6d33fb91ccb5885daac9b579 Author: mocusez <[email protected]> Date: Wed Dec 11 01:46:54 2024 +0800 fix(route/duckdb): change blogs link and author (#17856) * fix(route/duckdb): change blogs link and author * fix(route/duckdb): update description selector --------- commit d93568fd12901fddf64ad07b32c10d5f7a71b38f Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 10 05:59:59 2024 -0800 chore(deps-dev): bump eslint-plugin-n from 17.14.0 to 17.15.0 (#17850) Bumps [eslint-plugin-n](https://github.com/eslint-community/eslint-plugin-n) from 17.14.0 to 17.15.0. - [Release notes](https://github.com/eslint-community/eslint-plugin-n/releases) - [Changelog](https://github.com/eslint-community/eslint-plugin-n/blob/master/CHANGELOG.md) - [Commits](https://github.com/eslint-community/eslint-plugin-n/compare/v17.14.0...v17.15.0) --- updated-dependencies: - dependency-name: eslint-plugin-n dependency-type: direct:development update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 23f3dbda49b36de19a76af856dbc89a43072292a Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 10 21:24:57 2024 +0800 chore(deps-dev): bump @typescript-eslint/eslint-plugin (#17848) Bumps [@typescript-eslint/eslint-plugin](https://github.com/typescript-eslint/typescript-eslint/tree/HEAD/packages/eslint-plugin) from 8.17.0 to 8.18.0. - [Release notes](https://github.com/typescript-eslint/typescript-eslint/releases) - [Changelog](https://github.com/typescript-eslint/typescript-eslint/blob/main/packages/eslint-plugin/CHANGELOG.md) - [Commits](https://github.com/typescript-eslint/typescript-eslint/commits/v8.18.0/packages/eslint-plugin) --- updated-dependencies: - dependency-name: "@typescript-eslint/eslint-plugin" dependency-type: direct:development update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 61725b253c662be8e533acfc8590346c6201f38b Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 10 05:14:45 2024 -0800 chore(deps-dev): bump @typescript-eslint/parser from 8.17.0 to 8.18.0 (#17847) Bumps [@typescript-eslint/parser](https://github.com/typescript-eslint/typescript-eslint/tree/HEAD/packages/parser) from 8.17.0 to 8.18.0. - [Release notes](https://github.com/typescript-eslint/typescript-eslint/releases) - [Changelog](https://github.com/typescript-eslint/typescript-eslint/blob/main/packages/parser/CHANGELOG.md) - [Commits](https://github.com/typescript-eslint/typescript-eslint/commits/v8.18.0/packages/parser) --- updated-dependencies: - dependency-name: "@typescript-eslint/parser" dependency-type: direct:development update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit f1c793aca558c51e10fe646d6bc41759db41ca00 Author: iusx <[email protected]> Date: Tue Dec 10 20:55:52 2024 +0800 fix(route/isct): TIP overflow layout (#17838) * Update namespace.ts * Update namespace.ts commit 0d1a9d23052309b10669d8911b6b3a47d28de2ef Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 10 08:52:36 2024 +0000 chore(deps): bump nanoid from 3.3.7 to 3.3.8 (#17854) Bumps [nanoid](https://github.com/ai/nanoid) from 3.3.7 to 3.3.8. - [Release notes](https://github.com/ai/nanoid/releases) - [Changelog](https://github.com/ai/nanoid/blob/main/CHANGELOG.md) - [Commits](https://github.com/ai/nanoid/compare/3.3.7...3.3.8) --- updated-dependencies: - dependency-name: nanoid dependency-type: indirect ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 9714a48c2daf0b33ebdbba66b7190395a7bcdeb9 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 10 08:44:31 2024 +0000 chore(deps-dev): bump @vercel/nft from 0.27.7 to 0.27.9 (#17851) Bumps [@vercel/nft](https://github.com/vercel/nft) from 0.27.7 to 0.27.9. - [Release notes](https://github.com/vercel/nft/releases) - [Commits](https://github.com/vercel/nft/compare/0.27.7...0.27.9) --- updated-dependencies: - dependency-name: "@vercel/nft" dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit b17d3605c89eec1caab406ec8ffceee2d5ac9b39 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Dec 10 08:44:18 2024 +0000 chore(deps-dev): bump discord-api-types from 0.37.110 to 0.37.111 (#17849) Bumps [discord-api-types](https://github.com/discordjs/discord-api-types) from 0.37.110 to 0.37.111. - [Release notes](https://github.com/discordjs/discord-api-types/releases) - [Changelog](https://github.com/discordjs/discord-api-types/blob/main/CHANGELOG.md) - [Commits](https://github.com/discordjs/discord-api-types/compare/0.37.110...0.37.111) --- updated-dependencies: - dependency-name: discord-api-types dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 844e6580083202341b1795cf6158458c37852160 Author: pseudoyu <[email protected]> Date: Tue Dec 10 11:29:59 2024 +0800 feat(route/twitter): only apply third-party api supported routes commit 17e2fc729d52d17ba8b07988ae168cc3c956fdc7 Author: CaoMeiYouRen <[email protected]> Date: Tue Dec 10 10:30:42 2024 +0800 feat(route/github): enhance /repos route with optional parameters for type, sort, and direction (#17827) * feat(route/github): enhance /repos route with optional parameters for type, sort, and direction - Updated the route to accept optional parameters: type, sort, and direction. - Added filtering logic for repository types: all, owner, member, public, private, forks, and sources. - Improved documentation for parameters to clarify usage. * refactor(github): 优化仓库路由参数配置 - 移除 `direction` 参数，简化排序逻辑 - 更新路由路径以匹配简化后的参数 commit 8434009b734149ba85875b3626737b48763c7055 Author: Tony <[email protected]> Date: Mon Dec 9 13:33:28 2024 -0800 fix(route/sohu): add image URL decryption for fetched articles (#17837) commit 74f50c3154ad39eb3e6f3ba1cc517a6a3c496b20 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 9 21:04:32 2024 +0800 chore(deps): bump pac-proxy-agent from 7.0.2 to 7.1.0 (#17835) Bumps [pac-proxy-agent](https://github.com/TooTallNate/proxy-agents/tree/HEAD/packages/pac-proxy-agent) from 7.0.2 to 7.1.0. - [Release notes](https://github.com/TooTallNate/proxy-agents/releases) - [Changelog](https://github.com/TooTallNate/proxy-agents/blob/main/packages/pac-proxy-agent/CHANGELOG.md) - [Commits](https://github.com/TooTallNate/proxy-agents/commits/[email protected]/packages/pac-proxy-agent) --- updated-dependencies: - dependency-name: pac-proxy-agent dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 134d114ffaf60ce54301f8e80d3d1311c86742f1 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 9 04:30:50 2024 -0800 chore(deps-dev): bump vite-tsconfig-paths from 5.1.3 to 5.1.4 (#17830) Bumps [vite-tsconfig-paths](https://github.com/aleclarson/vite-tsconfig-paths) from 5.1.3 to 5.1.4. - [Release notes](https://github.com/aleclarson/vite-tsconfig-paths/releases) - [Commits](https://github.com/aleclarson/vite-tsconfig-paths/compare/v5.1.3...v5.1.4) --- updated-dependencies: - dependency-name: vite-tsconfig-paths dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 43f635d389af8b3710672498f62e020feddef2c8 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 9 04:29:17 2024 -0800 chore(deps-dev): bump @stylistic/eslint-plugin from 2.11.0 to 2.12.0 (#17831) Bumps [@stylistic/eslint-plugin](https://github.com/eslint-stylistic/eslint-stylistic/tree/HEAD/packages/eslint-plugin) from 2.11.0 to 2.12.0. - [Release notes](https://github.com/eslint-stylistic/eslint-stylistic/releases) - [Changelog](https://github.com/eslint-stylistic/eslint-stylistic/blob/main/CHANGELOG.md) - [Commits](https://github.com/eslint-stylistic/eslint-stylistic/commits/v2.12.0/packages/eslint-plugin) --- updated-dependencies: - dependency-name: "@stylistic/eslint-plugin" dependency-type: direct:development update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit c70fda5b99e6b6fee098dfac0c9d49dd581e30cb Author: pseudoyu <[email protected]> Date: Mon Dec 9 18:17:11 2024 +0800 feat(route/twitter): support tweets with replies commit 503a779091b36f4fa2fda3b6c4333d238aeb5e18 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 9 08:34:19 2024 +0000 chore(deps): bump https-proxy-agent from 7.0.5 to 7.0.6 (#17834) Bumps [https-proxy-agent](https://github.com/TooTallNate/proxy-agents/tree/HEAD/packages/https-proxy-agent) from 7.0.5 to 7.0.6. - [Release notes](https://github.com/TooTallNate/proxy-agents/releases) - [Changelog](https://github.com/TooTallNate/proxy-agents/blob/main/packages/https-proxy-agent/CHANGELOG.md) - [Commits](https://github.com/TooTallNate/proxy-agents/commits/[email protected]/packages/https-proxy-agent) --- updated-dependencies: - dependency-name: https-proxy-agent dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 95c71c5988759535fb51d5e9e221843419a9486c Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 9 08:28:05 2024 +0000 chore(deps): bump socks-proxy-agent from 8.0.4 to 8.0.5 (#17833) Bumps [socks-proxy-agent](https://github.com/TooTallNate/proxy-agents/tree/HEAD/packages/socks-proxy-agent) from 8.0.4 to 8.0.5. - [Release notes](https://github.com/TooTallNate/proxy-agents/releases) - [Changelog](https://github.com/TooTallNate/proxy-agents/blob/main/packages/socks-proxy-agent/CHANGELOG.md) - [Commits](https://github.com/TooTallNate/proxy-agents/commits/[email protected]/packages/socks-proxy-agent) --- updated-dependencies: - dependency-name: socks-proxy-agent dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 4af0c75d0594e91ad9b0cc7714d3697a7fc3faea Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Mon Dec 9 08:27:48 2024 +0000 chore(deps): bump tldts from 6.1.65 to 6.1.66 (#17829) Bumps [tldts](https://github.com/remusao/tldts) from 6.1.65 to 6.1.66. - [Release notes](https://github.com/remusao/tldts/releases) - [Changelog](https://github.com/remusao/tldts/blob/master/CHANGELOG.md) - [Commits](https://github.com/remusao/tldts/compare/v6.1.65...v6.1.66) --- updated-dependencies: - dependency-name: tldts dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 3b1de0dc9e91ba2109642b64b8ca5842366830a9 Author: Ethan Shen <[email protected]> Date: Mon Dec 9 09:48:29 2024 +0800 feat(route): add 游侠网资讯 (#17815) * feat(route): add 游侠网资讯 * fix typo commit 5c3d6fa348907c340a8d49fa3ee24cf393054441 Author: PangBo <[email protected]> Date: Sun Dec 8 00:21:19 2024 +0800 feat(route): 添加 上海交通大学教务处 面向学生的通知 路由 (#17823) * feat(route): add `sjtu/jwc/students` * feat(route/sjtu/jwc): add description for /sjtu/jwc/students commit 0d07768955d577e1df7ffb2a34ad2c34f03c758d Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri Dec 6 08:28:12 2024 +0000 chore(deps): bump hono from 4.6.12 to 4.6.13 (#17818) Bumps [hono](https://github.com/honojs/hono) from 4.6.12 to 4.6.13. - [Release notes](https://github.com/honojs/hono/releases) - [Commits](https://github.com/honojs/hono/compare/v4.6.12...v4.6.13) --- updated-dependencies: - dependency-name: hono dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 8eb7abb8e2f0c9f61bc7077d6c54f7a6958338b4 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri Dec 6 08:27:52 2024 +0000 chore(deps): bump imapflow from 1.0.169 to 1.0.171 (#17817) Bumps [imapflow](https://github.com/postalsys/imapflow) from 1.0.169 to 1.0.171. - [Release notes](https://github.com/postalsys/imapflow/releases) - [Changelog](https://github.com/postalsys/imapflow/blob/master/CHANGELOG.md) - [Commits](https://github.com/postalsys/imapflow/compare/v1.0.169...v1.0.171) --- updated-dependencies: - dependency-name: imapflow dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 60dd8f6ba6263c6ed63f77a280b072c2d31cb05b Author: pseudoyu <[email protected]> Date: Fri Dec 6 16:12:32 2024 +0800 feat(route/twitter): support getUserTweets thirdparty api commit db18c561cbe6923bef7ae4b0b46ac659fa589902 Author: pseudoyu <[email protected]> Date: Fri Dec 6 07:54:20 2024 +0800 fix(route/joneslanglasalle): radix to 10 and set default to 12 commit d7447b25cff4830a4818e8c7b6abb4fde0d98530 Author: pseudoyu <[email protected]> Date: Fri Dec 6 07:14:05 2024 +0800 chore(route/joneslanglasalle): fetch 12 items by default commit ce2b89a6428979a7ca91861c88a5fc41283581fe Author: Yuhang <[email protected]> Date: Fri Dec 6 00:52:06 2024 +0800 fix: qweather route example inaccuracy. (#17805) * fix: qweather route example inaccuracy. qweather route example inaccuracy * fix(qweather): throw on config not found --------- commit db98160eed66ff11e754990880c8b017482de867 Author: CaoMeiYouRen <[email protected]> Date: Thu Dec 5 23:53:45 2024 +0800 fix(mihoyo/bbs): 修复 timeline API URL (#17806) - 更新 timeline API URL 以匹配最新的 API 路径 commit 166e69bfaa6d5143c1785677660574fbdf578d3a Author: pseudoyu <[email protected]> Date: Thu Dec 5 22:48:13 2024 +0800 fix(api): decrease maxHeaderSize commit b35d47d68519f55be0cc0fa739aa3212940d3d19 Author: Tony <[email protected]> Date: Thu Dec 5 22:19:39 2024 +0800 fix(36kr): add checks commit 9b684c2183c01973e1048bea2122763586533ff8 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Thu Dec 5 21:47:54 2024 +0800 chore(deps): bump @opentelemetry/exporter-prometheus (#17811) Bumps [@opentelemetry/exporter-prometheus](https://github.com/open-telemetry/opentelemetry-js) from 0.55.0 to 0.56.0. - [Release notes](https://github.com/open-telemetry/opentelemetry-js/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-js/blob/main/CHANGELOG.md) - [Commits](https://github.com/open-telemetry/opentelemetry-js/compare/experimental/v0.55.0...experimental/v0.56.0) --- updated-dependencies: - dependency-name: "@opentelemetry/exporter-prometheus" dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <[email protected]> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> commit 96416bda3b8329b746556cd376009a3cb40317ef Author: dependabot[bot] <4969933…
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Involved Issue / 该 PR 相关 Issue
Close #17959
Example for the Proposed Route(s) / 路由地址示例
New RSS Route Checklist / 新 RSS 路由检查表
PuppeteerNote / 说明
This RSS is the latest news related to Anthropic, a leading LLM developer.