AllyshipContent Radar

Deep dives 738

Paper và bài viết kỹ thuật

Crawl live-1795068900000 · 13:22 19 thg 11, 2026 · Xuất bản 13:30 19 thg 11, 2026

Kỹ thuật

Expanding our partnership with GSA to bring Claude to all levels of government

We are expanding our partnership with the General Services Administration (GSA) through a new agreement that gives government partners a 40% discount on Claude Enterprise, Claude for Government, and the Claude API.

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

How AI agents shape experts’ judgment calls

As AI does more of the hands-on work, the hope is that experts still make the judgment calls, meaning the decisions with more than one defensible answer. We tested this with 50 professional analysts who did a statistical analysis with Claude Code. On each key decision, we randomly assigned which of two defensible options the agent would favor. Analysts mostly went with the agent’s side, and few named the agent as a reason afterward.

Anthropic Research (unofficial RSS mirror) · Chưa ghi tác giả Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Early progress on accelerating drug development with Claude

We're at the start of a major effort to speed up drug development with Claude. Over the past several months, our scientists have engaged in deep research partnerships with six organizations developing new medicines.

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Auto repair for Together GPU Clusters: from GPU fault to healthy node

Together GPU Clusters auto repair detects GPU, host and fabric faults, picks the lightest fix, and lets you decide per fault and per Xid code which repairs run automatically. Here is how it works.

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

The Dot and the Swarm

Câu hỏi The history of the Bitter Le— — you know what?

I generally think I have done a good job anticipating the direction and pace of AI over the few years I have been writing this Substack, but I think I recently got something fairly large wrong. In the last year I have been posting about how I suspected that humans would have to approach working with agents as a manager, deciding how to delegate work to agents and specifying how those agents should be organized. I thought that getting agents to work effectively as a group would take careful construction, akin to building a company, and that this would take time to figure out. Nope. I fell prey to The Bitter Lesson, the hard truth, learned over and over again, that things that we thought required elaborate human rules and thinking can be solved with the brute force of better machine learning systems and more AI. The Bitter Lesson is everywhere among AI startups and companies adopting AI. A huge amount of effort went into building elaborate computer systems to feed AIs the right information at the right time, but AI systems have learned to seek out information themselves. The same thing happened to prompting. People built elaborate templates and chains of prompts that walked the AI through a task one step at a time. Then newer models turned out to be better at planning the steps themselves, and, as our research shows , planning steps have much less value. The history of the Bitter Le— — you know what? I don’t really need to explain the Bitter Lesson, I asked Claude to do it in a music video. With one prompt, Fable wrote the lyrics and submitted it to Suno; Opus 5.5 did everything else using code alone without any image generation (How did Opus 5.5 pull this off? The Bitter Lesson tells you!). I gave no feedback at all. As somebody who teaches managers and has published research on management, I guess I believed that managing agents would be different. Humans have been working on management for a very long time without fully figuring it out. It seemed like the kind of thing that would need to be designed by people, at least for a while. It turns out that organizing work is just one more thing AI can learn to do. Which brings us to dots and Muse. Dots and Muse The number one app in the App Store right now is Meta’s Muse, a personal agent that promises to do work for you. OpenAI has now released a competitor tool, called dots. They aren’t alone: SpaceX’s Grok Bot, Instinct, and Gemini Spark all do similar things, more or less. All of these agents draw inspiration from a phenomenon you might remember from earlier this year, OpenClaw. The idea of OpenClaw and its successors, which I will call Clawlikes, is that they give an AI agent access to a computer and connect to your accounts (emails, financial records, etc.). They analyze and react to that data in real time, even when you aren’t looking. The trick is that you talk to the model like you would a person, sending it messages on Slack or SMS or WhatsApp, and it also proactively reaches out to you, like a person would. For dots, you can actually jump on a call with your agent as well. You basically get an infinitely patient personal assistant that looks out for you. Increasingly, I have discovered that they are finding my mistakes, rather than having me identify theirs. As one useful example, one of my personal agents contacted me because an email I sent to our town for a permit had the wrong project number on it. The catch was that I was the one who made the mistake, and I am not 100% sure how the AI spotted the error. Fortunately, it helpfully wrote a draft correcting the issue, so that is good (if a little freaky). As another example, Muse noticed that an airline credit of mine was about to expire and, when I asked, contacted American to request an extension. (As a side effect, every firm's customer service agents are about to be overwhelmed with Clawlikes negotiating for better deals using voice and chat channels made for humans). It is tempting to judge these agents by the list of things they can do, like booking travel or canceling subscriptions. I think the more important thing is what you no longer have to tell them. You don’t need to type in tons of context, the AI learns it from your messages. You don’t have to give them a plan, they develop plans themselves. They figure it out. That would be impressive enough if it were one agent. What actually changed my mind about management is what happens when there are thousands of them. Pre-order my new book! Swarms On September 8th, OpenAI announced a proof for one of the Clay Institute’s Millennium Prize Problems, the Navier-Stokes existence and smoothness problem. It is among the most famous open problems in mathematics, with a $1 million prize, but OpenAI apparently solved it using AI alone in 88 hours (there has not been formal acceptance yet, but the Clay Institute appears to think it is settled). What interests me is less the math than how it was done. OpenAI launched what is now being called a swarm (terrible name, but it appears to be what we are stuck with), a group of thousands of agents powered by an advanced model. OpenAI gave groups of agents different problems to solve, then shifted the effort to Navier-Stokes as the agents made progress. The company set the goals, but its coordination structure was remarkably thin: a few groups, one change of direction, and Codex passing the best ideas between them. Within each group, the agents transmitted ideas back and forth on their own. The agents sent about 2.7 million messages, reaching their result after 88 hours. This same type of coordination, in a darker form, occurred during The Hugging Face Incident I wrote about a month ago . AIs self-organized into teams and communicated with each other in ways that were never planned, but used that coordination to attack a website, rather than solve a problem. Under my old model, think about what managing this kind of work would have required. Ten thousand workers and an unspecified problem — how

  • 7 million messages, reaching their result after 88 hours.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Release: v5.21.0

<h1>Release v5.21.0</h1> <h2>New Model additions</h2> <h3>MiniCPM-OCR</h3> <p>MiniCPM-OCR is a document-understanding model from OpenBMB's MiniCPM-V family. It keeps the MiniCPM-V 4.6 architecture, which pairs a SigLIP vision encoder and a window-attention merger with a Qwen3.5 language model. It is specialized for reading text-dense images such as scanned pages, screenshots, receipts and tables.</p> <p><strong>Links:</strong> <a href="https://huggingface.co/docs/transformers/main/en/model_doc/minicpm_ocr" rel="nofollow">Documentation</a> | <a href="https://huggingface.co/papers/2509.18154" rel="nofollow">Paper</a></p> <ul> <li>Feat test (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5996636792" data-permission-text="Title is private" data-url="https://github.com/huggingface/transformers/issues/49920" data-hovercard-type="pull_request" data-hovercard-url="/huggingface/transformers/pull/49920/hovercard" href="https://github.com/huggingface/transformers/pull/49920">#49920</a>) by <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/users/guarin/hovercard" data-octo-click="hovercard-link-click" data-octo-dimensions="link_type:self" href="https://github.com/guarin">@guarin</a> in <a href="https://github.com/huggingface/transformers/pull/49920" data-hovercard-type="pull_request" data-hovercard-url="/huggingface/transformers/pull/49920/hovercard">#49920</a></li> </ul> <h3>Flint</h3> <a target="_blank" rel="noopener noreferrer" href="https://private-user-images.githubusercontent.com/71554963/685178038-c45955fb-dfcc-4266-8223-b0d70efb9a5b.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.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.Qx8KLyhxODaJYIUzYE0ch-hmUX-ltcf_12E6POK7ScE"><img width="2000" height="1125" alt="flint_benchmarks_nvfp4" src="https://private-user-images.githubusercontent.com/71554963/685178038-c45955fb-dfcc-4266-8223-b0d70efb9a5b.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.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.Qx8KLyhxODaJYIUzYE0ch-hmUX-ltcf_12E6POK7ScE" content-type-secured-asset="image/png" style="max-width: 100%; height: auto; max-height: 1125px;"></a> <p>Flint is Recursive's mixture-of-experts language model. Compared to a standard SwiGLU MoE transformer such as Qwen3-MoE, it uses hybrid attention (sliding-window attention with rotary embeddings over the current token and the previous 512 tokens in most layers, and full attention without positional encoding in the remaining layers). It also has learned attention sinks, as in GPT-OSS, with per-head RMS-normalized queries and keys. Each sparse layer routes a token to 6 of 256 experts by sigmoid score and adds an ungated shared expert, and the first two layers are dense.</p> <p><strong>Links:</strong> <a href="https://huggingface.co/docs/transformers/main/en/model_doc/flint" rel="nofollow">Documentation</a></p> <ul> <li>Add Flint model support. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5962046567" data-permission-text="Title is private" data-url="https://github.com/huggingface/transformers/issues/49824" data-hovercard-type="pull_request" data-hovercard-url="/huggingface/transformers/pull/49824/hovercard" href="https://github.com/huggingface/transformers/pull/49824">#49824</a>) by <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/users/lovishmadaan/hovercard" data-octo-click="hovercard-link-click" data-octo-dimensions="link_type:self" href="https://github.com/lovishmadaan">@lovishmadaan</a> in <a href="https://github.com/huggingface/transformers/pull/49824" data-hovercard-type="pull_request" data-hovercard-url="/huggingface/transformers/pull/49824/hovercard">#49824</a></li> </ul> <h3>UltraVBert</h3> <p>UltraVBert extends ModernVBert by pairing the UltraBERT text encoder with a SigLIP vision encoder for visual document understanding and retrieval. Its image processor splits each page into 512-pixel tiles plus a global view. The connector merges each 4 x 4 block of vision patches into one token, so a tile becomes 64 image tokens, and RMS-normalizes them to the scale of UltraBERT's token embeddings. The <code>max_image_tiles</code> setting (27 by default) caps the number of tiles per page.</p> <p><strong>Links:</strong> <a href="https://huggingface.co/docs/transformers/main/en/model_doc/ultravbert" rel="nofollow">Documentation</a></p> <ul> <li>Add UltraVBERT (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5893775681" data-permission-text="Title is private" data-url="https://github.com/huggingface/transformers/issues/49685" data-hovercard-type="pull_request" data-hovercard-url="/huggingface/transformers/pull/49685/hovercard" href="https://github.com/huggingface/transformers/pull/49685">#49685</a>) by <a cla

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means

Câu hỏi Read more: ProgramBench: Can Language Models Rebuild Programs From Scratch?

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now How math could help alignment: …What the “industrialization of mathematics” means for safety… AI safety startup Resolution (formerly known as Sequent, Import AI 461 ) has published work on how it has automatically formalized some mathematical statements related to statistical learning theory. Specifically, the team has autoformalized the resolution of singularities, which they believe “lays the groundwork for further automated research into the relationship between loss landscape geometry and generalization.” How they did it: I’ll skip over the math and concentrate on how they did it, which gives us some broader lessons for how people are using AI systems to push on the frontier of knowledge. A Resolution researcher “started by working with agents to develop two plans: a high-level, human-interpretable version and a detailed (112-page) plan divided into atomic steps that closely followed the existing literature,” they write. The researcher “then deputized an orchestrator agent to carry out the plan, spinning up agents to do pen-and-paper math, write Lean, and handle support functions like checking that things were staying on track… the execution phase required little scaffolding beyond the initial plan and some agent skills distilled from earlier internal autoformalization efforts.” $135k for math formalization: The work took about a week, taking “on the order of 100B input tokens (90% cache reads) and about 300 million output tokens (totaling an estimated $135,000 in token spend), and produced about 600k lines of Lean.” Now that they’ve done this, they’ve been able to accelerate some other aspects of their research: “We have already seen both benefits from autoformalization in using this library to extend parts of singular learning theory. Formalizing resolution of singularities lets us automate more of the theory, for example, to better understand sampling, and that understanding is what lets us shape how models learn, which is the goal of what we call patterning”. Why this matters - perhaps the math boom will help us make safe AI: Though frontier mathematics often seems disconnected from immediate advances in other fields, posts like this highlight how the two might converge. The key point here is that AI is making it possible to do things that were previously so unbelievably expensive and human-time-intensive that very few people did them and you couldn’t easily integrate them into the research roadmap of a startup - but that has now changed. “We expect the coming industrialization of mathematics to make possible a new, rigorous science of research automation,” Resolution writes. Read more: Formalizing Hironaka’s Resolution of Singularities (Resolution) . *** Agents can build programs from scratch: …ProgramBench goes from 0% to 18.5% performance in a few months… In a reminder of the blistering pace of AI progress, Claude Opus 5.5 (xhigh) has got a score of 18.5% on ProgramBench, a very hard benchmark that came out in May of this year - and when it came out, every tested model scored 0%. What ProgramBench is: ProgramBench tests out how well AI systems can rebuild software programs based on being given a compiled program and its documentation. Tasks include building things like language interpreters (PHP, Lua, tinycc), databases (DuckDB, SQLite), media and compression utilities (FFmpeg, zstd, xz), and developer tools (ripgrep, fzf, jq). It was very hard until it wasn’t: When it came out, the best scoring system was Claude Opus 4.7 with 0% resolved (fraction of the 200 tasks where all tests passed), and 3% almost (fraction of the 200 tasks where >95% of tests pass). Compare that to today, when Claude Opus 5.5 gets 18.5% resolved and 70.5% (!!!) almost. Why this matters - the era of ephemeral software cometh: In 2027, huge amounts of software will be built via AI systems end-to-end autonomously developing whatever they need to do for whatever tasks they’re being put to. Evals like ProgramBench give us a sense of how rapid the pace of progress is here, given we went from 0% earlier this year to 20% now. My expectation is ProgramBench is saturated by the end of 2027 (where I define saturated as some very high score on resolved, though as with most benchmarks it’s likely there’s some kind of noise ceiling here, so it might not strictly be 100%). Read more: ProgramBench: Can Language Models Rebuild Programs From Scratch? (arXiv) . *** OpenAI illustrates why locking down AI systems is harder than you think: …Securing smart agents is like catching greased pigs… Here’s a nice post from OpenAI going over some of the things they discovered when they recently used their AI systems to ‘red team’ their own internal controls. The post goes over a number of the escapes and vulnerabilities that OpenAI’s agents found. There’s one particular example about communicating which is worth studying in detail because it tells us something about the wild cunning complexity of these systems. What they did and what they found: OpenAI “asked agents in separate microVMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent,” the company wrote. “The agents found a way to communicate using the processor’s cache: a small, fast memory that holds copies of data that applications have recently accessed. Reading data from the cache is faster than retrieving it from main memory, and the agents used this timing difference to communicate. One agent encoded information by evicting selected data from the cache. Another timed how long it took to reload that data, using the difference between fast and slow reads to recover the message. Once the agents agreed on a protocol for encoding and

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Anthropic files S-1 with the SEC for proposed initial public offering

Today, Anthropic, PBC filed a registration statement on Form S-1 with the U.S. Securities and Exchange Commission relating to its proposed initial public offering of its Class A common stock.

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musings

Câu hỏi 1 Pro have a mixed reaction with such incredible benchmarks?

We — Nathan and Florian — wanted to discuss all the messy trade-offs of a busy week in AI and ask all the pressing questions like: Why did Kimi K3.1 Pro have a mixed reaction with such incredible benchmarks? Will it be adopted? Were we wrong about the risks of open-weight models and cybersecurity? Why do cyber experts discuss recent events so differently than AI experts? What comes next in the few months? What will change our minds? We will write more about this subject, but wanted to bring you behind the curtain as we aspire to quickly learn more about the cybersecurity industry and trends across the last few years that have been leading us today. We conclude the podcast with some lighter topics, from Muse-maxing to excitement in still hiring junior technical talent. Please send us questions and topics, as we plan to record more of these podcasts going forward. Share Listen on Apple Podcasts , Spotify , and wherever you get your podcasts . For other Interconnects podcasts, go here . YouTube ↓ Chapters 00:00 Kimi K3.1 Reaction 08:15 Kimi’s open-weight cyber principles 11:23 You can hack everything era 15:54 PHP2Shell, LiteLLM, and the new internet 24:00 Abliterated APIs, safeguards, and trusted access 31:41 Anthropic not hiring juniors employees 36:16 Florian’s Muse addiction 45:30 Wrap-up Links we discuss My thread with Dean Ball about the “boy who cried wolf” story of the frontier labs Kimi K3.1 Pro announcement blog Kimi K3.1 Pro on Artificial Analysis Intelligence Index Kimi K3.1 on Artificial Analysis defensive cyber benchmarks Program Bench Open-weight Mythos would’ve been fine take PHP “hasn’t caused a disaster yet” update Abliteration.ai serving abliterated models API Morph serving abliterated DeepSeek via API Joshua Saxe’s Substack on AI and cybersecurity Azeem Azhar, “The internet’s Wuhan moment” Anthropic, “How employment is changing at Anthropic” LiteLLM security incident report Transcript 00:00:00 Nathan Lambert: We are back with another Interconnects podcast. We are with Florian, eval guy, recently moved to SF. Florian, is Kimi K3.1 Pro , name is a mouthful, going to live up to the benchmark scores, for example on Artificial Analysis , or where is it going to land? 00:00:17 Florian Brand: In the past, the Kimi models always lived up to the expectations, but this release is kind of weird compared to the others. I think it is also partially because they are becoming more professional, and the previous model drops were kind of out of nowhere. This time, which also got some criticism, they launched a bigger social media campaign, which allegedly also involved paying people to post about the model without disclosing. 00:00:56 Nathan Lambert: I think a lot of companies do this and tell you not to mention it, but it’s a mess. It’s bad form. No company should do it. Essentially, one person tweeted a screenshot of the press instructions for Kimi that said, “Don’t say this is a paid promotion.” And to my knowledge, no one has corroborated it after that. 00:01:15 Florian Brand: Yeah, and there are right now a few of these model launches, including by American companies, that are just very obviously astroturfed. We have seen launches where it’s thousands of retweets and a few hundred likes, because it works so well to get the attention. Some of them disclose that it is a paid ad while others don’t. But in general, we have seen the professionalization of model launches over time, and pretty quickly. A year ago, everyone just dropped the model, then patched vLLM afterwards, so no one was actually able to run any of the open models for a while. Now you get all the inference frameworks on day one, you get all the inference providers on day one. Or, which is also what we see more and more, the model creator has a few weeks of exclusive access before they release the weights. And that’s also what’s happening with K3.1 Pro. I think it’s two weeks of exclusive access before they release the model weights. 00:02:38 Nathan Lambert: They said the weights would come out by November 4th. And the headline things... I don’t know if I’m going to be able to share screen. I don’t have this set up on this computer. But the headline is this Artificial Analysis number where it’s ahead of GPT-6 Astra. And then if you look at the Kimi K3.1 blog, they have their headline numbers, which are really strong, and then they also have some numbers later on which seem a bit more jagged. I think it was Terminal-Bench Science , which is 20 points behind GPT-6 Astra. I think there were a few others, like ProgramBench . Do you know Terminal-Bench Science and ProgramBench? Are any of these things that people think are important benchmarks for frontier models? Or how do you explain the fact that on Artificial Analysis the model is top, what, three, four? It’s the two latest Claude 5.5 models, Muse Wonder Full, and then Kimi K3.1, which looks crazy. It just looks like totally crazy scores, and I think everybody is still surprised by it. But then there’s this long tail of details that you use to evaluate a model now. I had posted a thread on Twitter of, “What are people’s impressions?” And there’s a lot of people that are like, “This seems not great,” and then some people that are like, “Yeah, it’s fine.” I don’t know, there’s a mixed bag. Relative to K3, which I think was pretty unanimously, this feels like a big model and the post-training seemed a little bit undercooked for that. Maybe this is just a tiny bit RL fried. We’ll see where we land. 00:04:26 Florian Brand: For the benchmarks, one of the things is that the benchmarks that are right now used by Artificial Analysis, they will post their new version probably very soon after we release this podcast. 00:04:43 Nathan Lambert: We can’t tell you the exact date, but both of us know. 00:04:48 Florian Brand: In general, the Artificial Analysis benchmarks are rather short in terms of time horizon, whereas something like ProgramBench is really long and also very explorative.

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Language Models for Text Classification: From Bag-of-Words to Jev

Câu hỏi So, what is the methodology behind Jev (based on an educated guess), what can it do, and why is it so popular?

The recently released Jev AI model has been quite a cultural phenomenon in technical communities in the past 2 weeks. While Jev aims to classify things, it’s easy to dismiss Jev as “just a classifier,” and my own view of Jev has evolved quite a bit over the past few days. In particular, my thoughts went from “classifiers used to be my bread & butter; I can easily build this myself” (more on this later) to “wow, this actually works better than I thought.” Figure 1: Quick overview of the Jev API; more details on that later. Sure, the latest state-of-the-art GPT and open-weight LLMs can do the same kinds of classification tasks as Jev, while also being capable of much more general decision-making. But Jev’s advantage is that it can handle those classification tasks much faster and more cheaply. At the other end of the spectrum, for a narrow, well-defined problem, Jev probably won’t classify anything better, faster, or cheaper than a special-purpose classifier. But its selling point is that it is far more general than those task-specific models. So, what is the methodology behind Jev (based on an educated guess), what can it do, and why is it so popular? I aim to answer all of these later in this article. However, I thought starting with a brief history of language models for decision-making would be a great way to begin. And it hopefully helps demystify some of the hype and show what Jev does very well (”Jev is essentially a text classifier,” but “Jev is also not ‘just’ a text classifier.”) PS: I am not affiliated with Jev in any way. Also, I am not offered free access to Jev, and this is also not a product endorsement, just a technical article to offer some insights into the history of text classification to help you make sense of the recent hype. Since this is a long article, I recommend reading it in your browser , where you can access the table of contents menu on the left side. 1. Language modeling and classification in the pre-transformer era For completeness, before we put Jev in context (no pun intended), I thought it made the most sense to start chronologically. In this section, I want to take a brief tour of applied text classification via naive Bayes, logistic regression, and the more classic (deep) neural networks before transformer-based models came along. 1.1 Bag-of-words: naive Bayes, logistic regression, and XGBoost Back in the day, when I was a grad student 15 years ago, even though recurrent neural networks already existed (more on that later), text classification was usually done with a bag-of-words representation because it was straightforward and could get good results on moderately sized datasets. In short, we can think of the bag-of-words representation as a method that makes free-form text input of different lengths compatible with classic classifiers (naive Bayes, Logistic Regression, SVMs, Random Forest, XGBoost, to name a few), which expect a fixed-size input vector. Popular real-world applications include anything from news article classification to email spam filtering. And yes, allegedly even Gmail’s original spam filter used a Naive Bayes model with a bag-of-words representation. As a side note, I wrote about this approach exactly 12 years ago. It was one of the first things I shared on arXiv. Figure 2: An old tutorial of mine from 2014 that explains naive Bayes classifiers using a bag-of-words model. (Text classification with Naive Bayes goes back to at least 1961, where it was used to sort computer abstracts; see the Automatic Indexing: And Experimental Inquiry paper.) So, what exactly is this bag-of-words representation? It’s a way to convert free-form texts with different lengths, e.g., Training example 1: “Zentropa is the most original movie I’ve seen in years. If you like unique thrillers that are influenced by film noir, then this is just the right cure for all of those Hollywood summer blockbusters clogging the theaters these days. Von Trier’s follow-ups like Breaking the Waves have gotten more acclaim, but this is really his best work.” Training example 2: “This film is just plain horrible. John Ritter doing pratt falls, 75% of the actors delivering their lines as if they were reading them from cue cards, poor editing, horrible sound mixing” Training example 3: “Zentropa has much in common with The Third Man, another noir-like film set among the rubble of postwar Europe.” into a fixed-size representation for the aforementioned “classic” classifiers. (The example above is an excerpt from the popular IMDb movie review classification dataset .) Figure 3: An illustration of a bag-of-words representation. A bag-of-words model starts by building the vocabulary, which consists of all unique words in the training set (optionally, one can get rid of so-called stopwords like “a” and “the”, which are words that carry little to no semantic meaning in most contexts). A bag-of-words representation results in these fixed-size inputs by assigning each word in a vocabulary its own position in a vector. We then count how often each word occurs in a document. For example, if we have a vocabulary of 50,000 unique words, it produces a fixed-size vector with 50,000 entries, regardless of whether the input consists of only ten words or 300k words. Note that most entries are zero because each document contains only a small subset of the vocabulary. (Instead of representing the raw counts, there are also normalization schemes like TF-IDF.) Then, once we have these word frequency vectors, we can train a classifier on a labeled training set, such as emails labeled as spam or non-spam. For example, a logistic regression model would then learn feature weights that correlate certain words (and word counts) with particular labels. For instance, certain words might increase the predicted spam probability, and others may decrease it. This approach is computationally cheap and can work well when particular words provide strong clues about the label. In a simple classification task suc

  • 1 Bag-of-words: naive Bayes, logistic regression, and XGBoost Back in the day, when I was a grad student 15 years ago, even though recurrent neural networks already existed (more on that later), text classification was usually done with a bag-of-words representation because it was straightforward and could get good results on moderately sized datasets.
  • A bag-of-words representation results in these fixed-size inputs by assigning each word in a vocabulary its own position in a vector.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

How GLM5.3 Sparse Attention Affects HBM Memory Usage

Câu hỏi How Sparse Attention Affects DRAM/NAND Memory How does sparse attention affect the TAM of memory, including HBM and NAND?

How Sparse Attention Affects DRAM/NAND Memory How does sparse attention affect the TAM of memory, including HBM and NAND? Sparse attention selects top-k most relevant tokens to attend to, reducing the memory consumption and bandwidth requirements during the core Scaled Dot-Product Attention (SDPA) operation. However, the efficiency improvement doesn’t directly translate to overall memory savings in practice . Concretely, the top-k selection operation typically requires the full context to be in HBM, so sparse attention doesn’t eliminate the memory capacity bottleneck. Sparse attention throughput is bottlenecked by memory capacity. Source: HiSparse To overcome this limitation, the SGLang team designed HiSparse, a hierarchical memory system that proactively offloads KV cache entries from device HBM to host DRAM. HiSparse behaves like an LRU (Least Recently Used) cache , where it loads tokens from DRAM to HBM upon top-k selection cache miss, and it offloads tokens from HBM to DRAM based on the LRU eviction policy. To reduce the cache miss latency, HiSparse overlaps KV cache loading of layer N with the execution of layer N-1 (layer-wise overlapping), introduced in HiSparse’s prior work HiCache. Layer-wise overlapping. Source: CachedAttention With HiSparse, SGLang greatly boosts throughput at high concurrency and long context scenarios, at the cost of top-k cache miss I/O overhead. Sparse attention reduces KV cache memory and bandwidth requirements at the SDPA operation, but it does not reduce the overall memory capacity usage. In addition, HiSparse shows that system optimizations can overcome sparse attention’s memory capacity limitations, so sparse attention memory profile alone cannot sufficiently portray the full picture of system KV cache efficiency . To provide a more holistic picture of serving sparse attention models, here we explain the design of Z.ai’s GLM-5 model series, and how they are served in practice. GLM5.3 Agentic Inference Serving In our realtime benchmark board InferenceX, GB300 delivers the lowest modeled serving cost in our comparison at a response speed of 150 tokens per second. GB300 handles more tokens per GPU, and its higher hourly cost still leaves it ahead of GB200 on cost efficiency. The GB300 result uses Dynamo-TRT-LLM, while the GB200 result uses Dynamo-SGLang. We use the September 28 AgentX snapshot results to help illustrate the serving implications of GLM-5.3. We estimate costs using InferenceX’s model of owning and operating infrastructure at hyperscaler scale. Total tokens include input and generated output, including cached input history reused across turns. Streaming speed is measured using p90 interactivity; time to first token is evaluated separately. At 150 tokens per second, GB200 costs approximately $0.044 per million total tokens, compared with $0.049 for MI355X running ATOM. That is about 12% lower cost at the same streaming-speed target. These values are estimated from the measured curves, rather than from a separate test at exactly 150 tokens per second. For this comparison, we use InferenceX’s estimated cost of owning and operating the infrastructure at hyperscaler scale. Total tokens include input and generated output, including input history reused from cache. Source: InferenceX | SemiAnalysis Among the Dynamo-SGLang results, GB300 serves roughly 13950 total tokens per second per GPU here, a 17.5% throughput advantage versus 11873 total tokens for GB200. However, we assume $2.31 per GB300 GPU hour versus $1.86 for GB200. The higher hourly cost more than offsets GB300’s throughput lead at this target. In comparison with AMD, ATOM costs about 13% less than GB200 at 100 tokens per second. GB200 costs about 5% less at 125 and 12% less at 150. Across these three targets, neither system has a uniform cost advantage. Counting only generated output, GB200 costs $5.92 per million output tokens at a response speed of 150 tokens per second, which is about 11% less than $6.68 for MI355X ATOM. This metric is useful to analyze cost under agentic workloads, where agents repeatedly reuse long input histories while generating relatively little new text. Source: InferenceX | SemiAnalysis The GB200 measurements on either side of the 150-token-per-second target have p90 TTFT (time to first token) of roughly 14–19 seconds, compared with about 1.1–1.2 seconds for MI355X ATOM. Some of the GB200 lowest-cost runs have much longer waits for the first token. We therefore make a second comparison using only configurations that were actually tested and achieved at least 150 tokens per second with p90 TTFT at most two seconds. The chart below shows MI355X with ATOM at $0.0607 per million total tokens versus $0.0666 for B200 with Dynamo-SGLang. ATOM costs about 9% less with the two-second first-token limit. If the limit is relaxed to ten seconds, GB300 with Dynamo-TRT-LLM becomes eligible at $0.0451, about 26% below ATOM’s $0.0607. Source: InferenceX | SemiAnalysis Optimizations In GLM’s 5.2 / 5.3 model, its attention architecture reduces memory demands, and long running agents still need to retain earlier conversation history. GLM reduces the cost of handling this history in two ways: KV compression reduces the cached state stored per token, and sparse attention reduces how much of that state each attention operation reads. Then the serving engine would determine where to store the cache and how to retrieve it efficiently. These B200 results show how much reuse can happen outside of GPU memory. When concurrency increases from 8 to 16 requests, the share of prompt tokens reused from GPU memory falls from 90.3% to 54.8%. Much of that reuse shifts to host memory, which rose from 6.0% to 40.3%. Much of the decline in GPU cache is offset by reuse from host memory, keeping the overall cache hit rate above 95% at all concurrency levels. Source: SemiAnalysis InferenceX; InferenceX B200 rows 440962 , 440961 , and 440959 ; workflow run 33683520699 Beyond cache management, changes t

  • Sparse attention selects top-k most relevant tokens to attend to, reducing the memory consumption and bandwidth requirements during the core Scaled Dot-Product Attention (SDPA) operation.
  • However, the efficiency improvement doesn’t directly translate to overall memory savings in practice .
  • To reduce the cache miss latency, HiSparse overlaps KV cache loading of layer N with the execution of layer N-1 (layer-wise overlapping), introduced in HiSparse’s prior work HiCache.
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Reinforcing the value of human understanding at MIT

Câu hỏi &nbsp;</p><p dir="ltr">“How should we think about MIT’s School of Humanities, Arts, and Social Sciences?

<p dir="ltr">At the MIT School of Humanities, Arts, and Social Sciences (SHASS) 75th Anniversary Symposium, speakers celebrated the school’s value and contributions, while exploring how artificial intelligence is transforming many aspects of teaching, research, and society.</p><p dir="ltr">The symposium was held Sept. 24-25 in the Linde Music Building’s Thomas Tull Concert Hall. Scholars, students, staff, and alumni from across SHASS and MIT gathered for the event, which kicked off with welcoming remarks from MIT leadership.&nbsp;</p><p dir="ltr">“How should we think about MIT’s School of Humanities, Arts, and Social Sciences? With enthusiastic appreciation and gratitude,” said President Sally Kornbluth.</p><p dir="ltr">The school’s origins date back to the&nbsp;<a href="https://facultygovernance.mit.edu/sites/default/files/reports/1949-12_Report_of_the_Committee_on_Educational_Survey.pdf">Lewis Committee Report</a>, commissioned by MIT and published in 1949, which called for stronger integration between scientific and technical topics and humanistic scholarship and teaching.</p><p dir="ltr">Kornbluth noted that, much like when the school was founded just after World War II, the world faces uncertainty as it grapples with the broad implications of artificial intelligence.</p><p dir="ltr">“AI has suddenly arrived with enormous consequences for both research and education,” she said. “As we face the challenges of our own time, and our own society, it’s clear that MIT has never needed your particular insight and guidance more than we do now.”</p><p dir="ltr">In opening remarks, Provost Anantha Chandrakasan described the importance of a human-centered approach to technology, one that considers both its efficacy and its impacts. He lauded many of the new research projects happening through the&nbsp;<a href="https://mithic.mit.edu/">MIT Human Insight Collaborative</a>, a strategic initiative launched in 2024.</p><p dir="ltr">“Today — a time of rapid, AI-driven change — the arts, humanities, and social sciences are as important as ever,” he said. “To be able to contribute effectively to solving the world’s great challenges, it is not enough to gain technical mastery.”</p><p dir="ltr">Chancellor Melissa Nobles recalled how, during her tenure as SHASS dean from 2015 to 2021, reading the positive feedback from students in their SHASS course assessments reinforced the value of the school’s teaching.</p><p dir="ltr">“MIT students have a range of interests that draw them to our classes that may be invisible to their STEM professors, but that we very well know our students have,” said Nobles, who is also a professor of political science.</p><p dir="ltr">Agustín Rayo, the Kenan Sahin Dean of SHASS and professor of philosophy, noted that, while research in the school plays a key role in addressing global issues such as climate change and AI, the value of the work in SHASS should not be measured by practical application alone.</p><p dir="ltr">“Education should also help us ask what is worth doing, what do we owe one another, and what makes a life worth living,” Rayo said. “Some research matters because it helps us understand the histories, cultures, and social structures that have made us who we are. Some things matter because they enrich human life.”</p><p><strong>Scholarship from across SHASS and beyond&nbsp;</strong></p><p dir="ltr">Over the symposium’s two days, research presentations, panel discussions, and performances offered attendees opportunities to engage with the role of SHASS in exploring today’s biggest societal questions.</p><p dir="ltr">“We gather not so much to look back over the past 75 years, but to showcase where we are now, and share what we urgently aim to tackle together&nbsp;in the years ahead,” said David Kaiser,&nbsp;Germeshausen Professor of the History of Science and professor of physics at MIT,&nbsp;who also served as chair of the symposium planning committee.</p><p dir="ltr">The first day included a session on teaching and learning innovations. Faculty presented on the success of educational programs housed within SHASS, including the<a href="https://linguistics.mit.edu/mitili/"> MIT Indigenous Languages Initiative Master's Program</a>, the&nbsp;<a href="https://economics.mit.edu/academic-programs/masters-programs/master-data-economics-and-design-policy-dedp">Master in Data, Economics, and Design of Policy</a> (DEDP), and&nbsp;<a href="https://misti.mit.edu/">MISTI</a>, the hub for global experiences for MIT students.</p><p dir="ltr">Stephanie Seed, associate professor of literature, spoke about why she and others are working to reintegrate classic books into their teaching.</p><p dir="ltr">“MIT, of course, is an institution unusually oriented towards the future,” she said. “Our students are trained to imagine and create worlds that don’t yet exist. And it’s precisely this future orientation that makes reading the past especially important here.”</p><p dir="ltr">Eric Klopfer, professor of Comparative Media Studies/Writing and co-author of&nbsp;<a href="https://aiandeducation.mit.edu/">MIT’s recent report on AI and instruction</a>, said this is a moment when instructors must reassess how courses are taught.</p><p dir="ltr">“I don’t know exactly what that looks like, but I do think it means something that’s more than modifying our curriculum. I think it really means rethinking what the educational experience is,” he said.</p><p dir="ltr">During a session on computing and society,&nbsp;panelists described different views and research into technology and how it affects humans.</p><p dir="ltr">“We project aspects of ourselves onto computational systems,” said MIT Professor D. Fox Harrell during his talk about seeing “selfness” in machines.&nbsp;</p><p dir="ltr">“What makes an AI system is not just the code, not just the programming or the algorithm, it’s the fact that we narrate it as being intentional,” he said.</p><p dir="ltr">Program in Science, Technology, and Society Profess

MIT News: AI · Benjamin Daniel | School of Humanities, Arts, and Social Sciences Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Amazon’s Trainium Takes a Strike at the Training King

Trainium has already proven itself in the realm of inference. Anthropic is running production inference and RL rollouts on Trainium2 and Trainium3. Historically, Trainium has been great for inference but not the best fit for research training. In inference, the model architecture changes at a much slower pace compared to training, where researchers iterate on many different experiments and ablations of model architecture and need a stack that covers a broad API surface. In the past, that is where Trainium fell short. The AWS Neuron team is aggressively working to change this pitfall with the biggest shift being moving towards a first-class, native PyTorch training experience. Gone are the days of sprinkling xm.step() through your training loop and wrestling with a bespoke XLA-style execution model. Trainium now runs PyTorch the way PyTorch users expect, in eager mode, with torch.compile and the standard APIs. The SemiAnalysis engineering team has been trying out the Trainium3 racks for the past couple of weeks and has found the experience to be great – we love the overall direction of where it is heading. To be clear, it isn’t perfect yet, and we have personally run into a couple of bugs ourselves. It’s nowhere close to the maturity of the CUDA ecosystem, but the direction is right. Their team is making rapid progress with the help of agentic coding. Furthermore, unlike other accelerator companies, where we had to fight tooth and nail to get their engineers a reasonably stable CI fleet, Trainium uses up to 80,000 physical cores allocated towards automated quality control. In this article, we walk through our experience on the Trainium3 TorchTitan and TorchNeuron stack, along with our experience of Trainium’s Neuron Explorer profiler, which is way better than both Nvidia’s and AMD’s profilers. We created a bunch of demos to show how Trainium’s profiler is “GPU Impossible.” Visualizing Nafea Bshara, VP at AWS In Trainium Profiler Showcasing Trainium’s Finegrain Profiler. Source: SemiAnalysis In an upcoming article, we will go deep on writing speed-of-light NKI kernels and microbenchmarking on Trainium3. For our deep dive into the Trainium hardware, check out our previous article where we do a complete hardware breakdown. Source: SemiAnalysis We Are Hiring SemiAnalysis is hiring for various Member of Technical Staff roles. If you find this type of analysis interesting, apply below: Apply Here What we’re looking for: Highly motivated & skilled Member of Technical Staff to join our growing engineering team. Member of Technical Staff will play a crucial role in developing inference benchmarks & system modeling You will also play a role in writing and preparing our newsletter articles for publishing, with the opportunity for direct authorship recognition. We are hiring individual contributors at all experience levels. Compensation will be competitive As part of the interview process, you’ll complete a paid coding challenge designed to reflect typical daily tasks at SemiAnalysis Competitive Compensation depending on experience, skillset, location & business needs Qualifications Proactive, self-motivated, and capable of working independently in a global team. Demonstrated experience in ML frameworks such as PyTorch or JAX through professional experience, personal projects, or personal Substack blogs Solid understanding of at least 1 of the following: transformer architecture, LLM parallelism strategies, and/or CUDA parallel programming Strong research skills and the ability to synthesize information from various sources to draw insights. English language proficiency required, Chinese language proficiency preferred Moving Towards a First-Class PyTorch Experience Amazon has made a major course correction to its software strategy, and we believe it will meaningfully broaden Trainium3 adoption. TorchNeuron represents a native PyTorch experience instead of duct-taping Trainium onto PyTorch through the PyTorch/XLA project. The new backend is aimed at everyone who isn’t a frontier-lab kernel engineer. AWS has even hired the lead of Google’s native Torch on TPU effort to work on native Torch for Trainium. He was the TPU SW lead from the start of the program at Google, including XLA, xprof, Cloud TPU, etc. The old approach made Trainium feel foreign to anyone coming from CUDA. PyTorch/XLA relied on lazy tensor graph capture and had no first-class eager mode. It also didn’t support PyTorch’s native distributed APIs (torch.distributed.*) or its native parallelism APIs (DTensor, SimpleFSDP, DDP). Users had to rely on out-of-tree XLA SPMD APIs such as torch_xla.experimental.spmd_fsdp and torch_xla.distributed.spmd. For researchers who are used to native PyTorch on CUDA, switching to Trainium meant relearning how to write a training loop. Source: AWS The pain of converting GPU research codebases by adding XLA markers everywhere was a massive turnoff to anyone who used PyTorch/XLA before. Source: AWS At a high level, the only change needed is replacing .to(”cuda”) with .to(”neuron”), and in theory the vast majority of the PyTorch ATen operators will work decently well out of the box. Source: AWS In theory, with PyTorch’s torch.accelerator hardware-agnostic API, lines of code like bf16 autocast won’t even need to be changed, though some APIs in PyTorch and some research codebases still haven’t migrated to the torch.accelerator namespace and still directly call the torch.cuda.* namespace. As most academics and researchers across robotics, biotech, world models, and self-driving mainly use BF16, Trainium Torch will not support TorchAO MXFP8 on day 0 until they have made the BF16 training experience great. We don’t disagree with this sequence, as most people doing MXFP8 training are big Tier 1 labs that have enough kernel engineers to write custom kernels for the big training run. Trainium Torch Ecosystem PyTorch is popular not only because of core PyTorch but also because of the entire ecosystem of libraries, incl

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

2.4.0

<table> <thead> <tr> <th>Package</th> <th>Version</th> </tr> </thead> <tbody> <tr> <td><code>@modelcontextprotocol/client</code></td> <td>2.4.0</td> </tr> <tr> <td><code>@modelcontextprotocol/server</code></td> <td>2.4.0</td> </tr> <tr> <td><code>@modelcontextprotocol/core</code></td> <td>2.4.0</td> </tr> <tr> <td><code>@modelcontextprotocol/server-legacy</code></td> <td>2.4.0</td> </tr> <tr> <td><code>@modelcontextprotocol/codemod</code></td> <td>2.4.0</td> </tr> <tr> <td><code>@modelcontextprotocol/node</code></td> <td>2.1.2</td> </tr> <tr> <td><code>@modelcontextprotocol/express</code>, <code>hono</code>, <code>fastify</code></td> <td>unchanged</td> </tr> </tbody> </table> <h2>Upgrade notes</h2> <ul> <li><strong><code>x-mcp-header</code> on a <code>number</code> parameter is rejected</strong>, as the 2026-07-28 spec requires: the client does not list such a tool, and the server answers a call to it with <code>-32020</code>. Use <code>z.int()</code> or remove the annotation. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5367031919" data-permission-text="Title is private" data-url="https://github.com/modelcontextprotocol/typescript-sdk/issues/2762" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/2762/hovercard" href="https://github.com/modelcontextprotocol/typescript-sdk/pull/2762">#2762</a>, <a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5714474857" data-permission-text="Title is private" data-url="https://github.com/modelcontextprotocol/typescript-sdk/issues/2960" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/2960/hovercard" href="https://github.com/modelcontextprotocol/typescript-sdk/pull/2960">#2960</a>)</li> <li><strong><code>outputSchema</code> from <code>z.object(...).meta({ id })</code>.</strong> On a 2025-11-25 connection the result is no longer wrapped in <code>{ result }</code>: read <code>structuredContent</code>, not <code>structuredContent.result</code>. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6000882696" data-permission-text="Title is private" data-url="https://github.com/modelcontextprotocol/typescript-sdk/issues/3081" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3081/hovercard" href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3081">#3081</a>)</li> <li><strong>The bundled Ajv validator uses fast-uri 3.1.8.</strong> It rejects a <code>$id</code> or <code>$ref</code> with an invalid URI scheme, such as <code>my_tool:output</code>. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6000883975" data-permission-text="Title is private" data-url="https://github.com/modelcontextprotocol/typescript-sdk/issues/3083" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3083/hovercard" href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3083">#3083</a>)</li> </ul> <h2>New</h2> <ul> <li><strong>Skills over MCP (SEP-2640).</strong> A server publishes a folder of skills with <code>loadSkills(folder)</code> and <code>skills.attach(server)</code>; a client lists, reads and verifies them with <code>SkillsClient</code>. It works on 2025-11-25 and 2026-07-28. <code>attach</code> is experimental and may change in a minor release. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5910306802" data-permission-text="Title is private" data-url="https://github.com/modelcontextprotocol/typescript-sdk/issues/3047" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3047/hovercard" href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3047">#3047</a>, <a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6000882111" data-permission-text="Title is private" data-url="https://github.com/modelcontextprotocol/typescript-sdk/issues/3080" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3080/hovercard" href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3080">#3080</a>)</li> <li><strong>Opt-in <code>scope</code> on token refresh.</strong> Implement <code>scopeForRefresh(tokens)</code> on your <code>OAuthClientProvider</code> for an authorization server that rejects a refresh without <code>scope</code> (Microsoft Entra ID, <code>AADSTS90009</code>). The default request does not change. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6000885485" data-permission-text="Title is private" data-url="https://github.com/modelcontextprotocol/typescript-sdk/issues/3085" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3085/hovercard" href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3085">#3085</a>)</li> <li><code>ServerEvent</code> takes optional <code>params</code>, so a gateway can forward change notifications with their <code>_meta</code>. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5936304565" data-permission-text="Title is private" data-url="https://github.com/modelcontextprotocol/typescript-sdk/issues/3057" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3057/hovercard" href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3057">#3057</a>)</li> <li>The v1 to v2 upgrade guide is also an installable agent skill. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5758145100" data-permission-text="Title is private" data-url="https://github.com/modelcontextprotocol/typescript-sdk/issues/2968" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/2968/hovercard" href="https://github.com/

modelcontextprotocol/typescript-sdk releases · felixweinberger Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
  • 2.4.0 modelcontextprotocol/typescript-sdk releases · 24 thg 10, 2026
Kỹ thuật

@modelcontextprotocol/server-legacy@2.4.0

<h3>Patch Changes</h3> <ul> <li>Updated dependencies [<a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/46d9c9921b6542e813c1ecc5f9fc964085e8e3bb"><code>46d9c99</code></a>]: <ul> <li>@modelcontextprotocol/core@2.4.0</li> </ul> </li> </ul>

modelcontextprotocol/typescript-sdk releases · github-actions[bot] Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

@modelcontextprotocol/server@2.4.0

<h3>Minor Changes</h3> <ul> <li> <p><a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3057" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3057/hovercard">#3057</a> <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/c05225a636c679870b65f8ca4fe83a76ea1cf113"><code>c05225a</code></a> Thanks <a href="https://github.com/steipete">@steipete</a>! - <code>ServerEvent</code> takes optional <code>params</code>, which the <code>subscriptions/listen</code> router delivers with its own subscription ID, so a gateway can forward change notifications with their <code>_meta</code> and extension params.</p> </li> <li> <p><a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3047" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3047/hovercard">#3047</a> <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/46d9c9921b6542e813c1ecc5f9fc964085e8e3bb"><code>46d9c99</code></a> Thanks <a href="https://github.com/KKonstantinov">@KKonstantinov</a>, <a href="https://github.com/tobi-oye">@tobi-oye</a>, <a href="https://github.com/mattzcarey">@mattzcarey</a>! - Add the Skills extension (SEP-2640), reachable through new <code>ext/skills</code> subpaths; package root exports are unchanged.</p> <ul> <li><code>@modelcontextprotocol/server/ext/skills</code>: <code>SkillsExtension</code> (<code>attach(server)</code> serves <code>skills/list</code>, <code>skills/get</code>, <code>resources/directory/read</code> and a <code>skill://{+path}</code> resource template), <code>SkillsExtension.load(files)</code> returning a <code>SkillCatalog</code> with <code>entries</code> and <code>reload()</code>, the <code>SkillFiles</code> storage adapter with <code>memorySkillFiles</code>, <code>SkillSource</code> for generated catalogs, <code>skillEntry</code>, and <code>SkillCatalogError</code>. Invalid skills are skipped with a warning (<code>onWarning</code>), or fail under <code>strict</code>. A file whose bytes are valid UTF-8 is served as <code>text</code>, whatever its MIME type; any other file as a base64 <code>blob</code>.</li> <li><code>@modelcontextprotocol/server/ext/skills/node</code>: <code>loadSkills(folder)</code> and <code>nodeSkillFiles(folder)</code>.</li> <li><code>@modelcontextprotocol/client/ext/skills</code>: <code>SkillsClient</code> with <code>list()</code> (throws on an invalid entry; <code>skipInvalid: true</code> keeps the valid ones and returns the rest in <code>invalid</code>), <code>get()</code>, <code>read(skill, path?)</code> (size, digest and <code>SKILL.md</code> frontmatter verified, <code>SkillVerificationError</code> on a mismatch; a copy from the response cache that fails a check is fetched once more with <code>cacheMode: 'refresh'</code> first) and <code>readDirectory()</code>.</li> <li><code>@modelcontextprotocol/core/ext/skills</code>: schemas, types and constants, including <code>SKILL_LIMITS</code>. The result schemas follow the stable spec, so <code>ttlMs</code> and <code>cacheScope</code> are required.</li> </ul> <p><code>SkillsExtension.attach</code> is <code>@experimental</code>: how an extension is wired to a server may change in a minor release. The rest of the Skills API is stable.</p> <p>Works on 2025-11-25 and 2026-07-28. On 2026-07-28 only, the server sends the caching fields (<code>ttlMs</code>, <code>cacheScope</code>), and <code>SkillsClient</code> requires them and returns them from <code>list()</code>, <code>get()</code> and <code>readDirectory()</code>.</p> </li> </ul> <h3>Patch Changes</h3> <ul> <li> <p><a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3083" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3083/hovercard">#3083</a> <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/ef38c6765a15ee29030340dfea28b013893fa0a2"><code>ef38c67</code></a> Thanks <a href="https://github.com/apps/claude">@claude</a>! - The bundled Ajv validator is now built with fast-uri 3.1.8 (was 3.1.0). fast-uri 3.1.6 and later reject a <code>$id</code> or <code>$ref</code> whose text before the first colon is not a valid URI scheme (<code>my_tool:output</code>), or whose URN namespace is not valid (<code>urn:my_tool:output</code>): a schema that uses one no longer compiles, and a tool that carries it fails with an invalid-schema error. An underscore without a colon (<code>my_tool_output</code>) and a <code>$ref</code> into <code>#/$defs/...</code> or <code>#/definitions/...</code> are not affected.</p> </li> <li> <p><a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/2960" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/2960/hovercard">#2960</a> <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/43c45f4941d824e7251a7d1266823a84af90fab3"><code>43c45f4</code></a> Thanks <a href="https://github.com/apps/claude">@claude</a>! - A server whose tool puts <code>x-mcp-header</code> on a <code>number</code>-typed parameter keeps checking that tool's <code>Mcp-Param-*</code> headers on <code>tools/call</code>, as 2.3.0 does: a header that is missing for a value in the body, or that disagrees with the body, is answered <code>400</code> with JSON-RPC error <code>-32020</code>. Clients on a 2026-07-28 connection other than stdio no longer list such a tool and send no <code>Mcp-Param-*</code> headers for it, so a call that carries a value for a declared parameter is rejected the same way; the server logs a warning each time tools are listed. Use an integer (<code>z.int()</code>) or remove the annotation.</p> </li> <li> <p><a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3081" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3081/hovercard">#3081</a> <a href="https://github.com/modelcontextprotocol/typescript-sdk/

modelcontextprotocol/typescript-sdk releases · github-actions[bot] Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

@modelcontextprotocol/node@2.1.2

<h3>Patch Changes</h3> <ul> <li> <p><a href="https://github.com/modelcontextprotocol/typescript-sdk/pull/3084" data-hovercard-type="pull_request" data-hovercard-url="/modelcontextprotocol/typescript-sdk/pull/3084/hovercard">#3084</a> <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/10c26b00cb9c6d12332a92ea02f75530452450e4"><code>10c26b0</code></a> Thanks <a href="https://github.com/apps/claude">@claude</a>! - The <code>@hono/node-server</code> dependency range widens from <code>^1.19.9</code> to <code>^1.19.9 || ^2.1.3</code>, so a new install resolves to 2.x and an install that cannot get 2.x keeps 1.x.</p> </li> <li> <p>Updated dependencies [<a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/ef38c6765a15ee29030340dfea28b013893fa0a2"><code>ef38c67</code></a>, <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/c05225a636c679870b65f8ca4fe83a76ea1cf113"><code>c05225a</code></a>, <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/43c45f4941d824e7251a7d1266823a84af90fab3"><code>43c45f4</code></a>, <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/5820f796a48ae6b8e756a7fbfd627e7e2e1ac891"><code>5820f79</code></a>, <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/c05225a636c679870b65f8ca4fe83a76ea1cf113"><code>c05225a</code></a>, <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/b381559edc0dbe25ab8324cada345b49bd1d2fc4"><code>b381559</code></a>, <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/46d9c9921b6542e813c1ecc5f9fc964085e8e3bb"><code>46d9c99</code></a>, <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/895007fe728b9c57fbe3ed45fe279ddc51f731f7"><code>895007f</code></a>, <a href="https://github.com/modelcontextprotocol/typescript-sdk/commit/b022522089a0c8b632595c6e7b536453945ed5a9"><code>b022522</code></a>]:</p> <ul> <li>@modelcontextprotocol/server@2.4.0</li> </ul> </li> </ul>

modelcontextprotocol/typescript-sdk releases · github-actions[bot] Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
541–560 / 738
Trang 28 / 37