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

Huấn luyện

AI-assisted learning platform helps robots reason beyond literal commands

<p>Humans are intuitive by nature. We can sense meaning beyond the literal words that we encounter. Now, a new artificial intelligence-assisted learning platform developed by MIT engineers is enabling robots to read larger meaning into a human’s natural language commands.&nbsp;</p><p>In a study to be presented at the annual Conference on Robot Learning this month, researchers from MIT and New York University introduce <a href="https://arxiv.org/pdf/2610.28893" target="_blank">FLEX</a>, a flexible feature-editing pipeline for robot motion-planning.&nbsp;</p><p>With FLEX, humans can interact with a robot by giving it instructions in natural, conversational language. The robot can not only follow the instructions literally but also read into the words to draw larger generalizations that help to improve its motion-planning task.&nbsp;</p><p>The team demonstrated FLEX in several simulated scenarios, including one in which a robot serves a cup of coffee. In these virtual scenarios, a person interacts with a robotic arm by typing commands in natural language. As the robot moves the coffee cup over a cluttered table, the person might type “Don’t spill on the laptop!” In response, the robot will reason about why it should avoid the laptop. For this, FLEX incorporates a large language model (LLM) — a type of AI system that is trained on enormous libraries of text, including social media posts, video transcriptions, and web content, which the LLM uses to identify larger trends, patterns, and generalizations.&nbsp;</p><p>“LLMs have seen so many texts from the internet,” says study author Nathan Dennler, a postdoc in MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). “People might have written about, ‘Oh, my phone dropped in the water and now it doesn’t work,’ or ‘I spilled coffee on my computer and now it doesn’t work.’ So LLMs pick up on the idea that liquids plus devices leads to things not working.”</p><p>Through the LLM, the robot reasons that it should also avoid bumping into all electronics. In its updated path, the robot adjusts to avoid both the laptop and a nearby phone, without the person having to explicitly spell this out.</p><p>“The idea is that there is some higher-level intention behind the words that you say,” Dennler says. “People naturally get this. But you have to carefully explain that to a robot, which takes time and effort. FLEX gives you an interface of communication with a robot to say, ‘When people say this, they mean that.’”</p><p>The study was led by Andreea Bobu, assistant professor of aeronautics and astronautics at MIT. Co-authors are Audrey Lee SM ’26 and Ilia Sucholutsky of New York University.</p><p><strong>Beyond what we say</strong></p><p>At MIT, Bobu heads up the Collaborative Learning and Autonomy Research (CLEAR) Lab, where her group develops autonomous agents that learn to work for, with, and around people. She is focused on developing assistive and household robots that can carry out physical tasks, and reason and adapt in real-time, in response to a person’s natural, conversational commands.&nbsp;</p><p>“Robots use math to do their trajectory planning. But we can’t expect a consumer to learn to do this math in order to tell a robot what to do,” says Dennler, who is a member of the CLEAR Lab. “Instead of having to get people to learn how to use a robot in a specific way, the goal is to interact with the robot through natural language.”</p><p>There is a large effort in the robotics community to incorporate natural language into robotic learning pipelines. Various strategies include using language to clarify a task’s goal, redirect a robot’s trajectory, and update its understanding of what is a safe path to take. These existing methods use natural language to tell the robot how to physically perform a task. But a robot could potentially learn much more from what we tell it.&nbsp;</p><p>Bobu and her colleagues are expanding what robots can learn from natural language commands. In their new study, they focused on enabling a robot to interpret a person’s underlying, unspecified preference from limited natural language commands.&nbsp;&nbsp;</p><p><strong>Absorbing common sense</strong></p><p>The team’s new robot learning platform, FLEX, is a three-step process that helps a robot read more into a natural language command, beyond what it is explicitly told to do.&nbsp;</p><p>The first step, “Add,” instructs a robot to add a feature to its list of things in its environment that matter for the task at hand, based on a user’s natural language command. For instance, as a robot is moving a bowling ball from one location to another, a person might interject with “Don’t drop it on my mug!” The robot would then add the mug to a list of things to avoid in its path. This first step is how previous methods have worked. Every time a person makes a correction, the robot adds to its features list, and updates its trajectory to avoid those features.&nbsp;</p><p>“The problem happens when your feature list becomes super long, and some features are no longer relevant because your environment may have changed,” says Dennler, who notes that this approach can grow tiresome if a person has to specify every specific feature to avoid.&nbsp;</p><p>The second step, “Generalize,” is the new solution that FLEX provides. This step incorporates an LLM — in this case, GPT-4o — that considers the mug, along with any other features that the person has so far specified to avoid. The LLM automatically searches through its vast cache of data for patterns and connections among these features, to ultimately make a generalization that it should avoid objects that are fragile, even when the person has not explicitly said so.&nbsp;</p><p>“LLMs have&nbsp;seen so many texts from the internet, from all sorts of people and places,” Dennler says. “Because they’ve trained so much on those things, some of that common sense gets absorbed into the LLM.”</p><p>In the third step, “Mask,” the robot’s

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

A single AI model simulates the behavior of many substances, from sliding sand to folding proteins

<p>Today, a scientist who wants to simulate a physical process must pick the right tool for that one job, such as a solver built for flowing water or another designed to model the way a protein folds. Each one is hand-coded around the specific equations that dictate how that process behaves.</p><p>Researchers at MIT, the University of British Columbia (UBC), and elsewhere have now replaced that patchwork with a single deep-learning system, called <a href="https://worldparticle.github.io/WorldParticle_Web/#top" target="_blank">WorldParticle</a>, that could help engineers model different physical phenomena, from sliding sand to folding proteins, with one unified artificial intelligence architecture.</p><p>Scientists and engineers could use this system to model processes that typically require separate simulators, such as a fluid sloshing against a deformable solid. This could offer a faster, more accurate way to simulate how sloshing oil might impact an elastic valve in a storage tank, for instance. It could also help engineers study where and how granular materials, like sand or soil, could move when a soft robot touches them.</p><img src="/sites/default/files/images/inline/six-split-molecular.gif" data-align="center" data-entity-uuid="a6eee205-5f30-4ed1-bdc6-412b5c8c487a" data-entity-type="file" alt="A ribbon-like molecule turns to particles." width="960" height="400"><img src="/sites/default/files/images/inline/six-split-newtonian.gif" data-align="center" data-entity-uuid="419f7065-1f0d-4974-8c67-8fc2d9edbfcc" data-entity-type="file" alt="Particles move in a fluid motion." width="960" height="400"><img src="/sites/default/files/images/inline/six-split-granular.gif" data-align="center" data-entity-uuid="eea02e57-7a09-4cf2-9eea-2fce52af4399" data-entity-type="file" alt="A tiger turns to sand." width="960" height="400" data-caption="WorldParticle has been extensively validated across diverse phenomena, from microscopic protein motion to macroscopic physical simulations. Pictured are animations showing molecular, fluid, and granular simulations.&lt;br&gt;&lt;br&gt;Credit: Courtesy of the researchers"><p>The system represents every material it simulates, from a dune of sand to a protein molecule, as tiny particles and predicts where those particles will move one step at a time. Because it learns the underlying dynamics rather than being programmed with an equation, the model can adapt to materials, forces, and environments it never encountered during training.&nbsp;&nbsp;</p><p>In the long term, the researchers hope this work helps scientists discover new materials and explore physical processes they would be unable to effectively study with existing simulation methods.</p><p>“We want to have a GPT moment for simulation,” says Peter Yichen Chen, a former MIT postdoc who is now an assistant professor at UBC, and is a co-corresponding author of an open-access <a href="https://arxiv.org/pdf/2605.15305" target="_blank">paper</a> on this technique. “Every time people want to build a simulator for a particular physical phenomenon, they need to code it up from scratch. Our vision is to have a foundation model that works out of the box for any scenario you want to simulate. This is a first step in that direction.”</p><p>Chen is joined by co-corresponding and co-lead author Minghao Guo, an MIT postdoc; co-lead authors Caoliwen Wang and Siyuan Chen, UBC graduate students; Wojciech Matusik, an MIT professor of electrical engineering and computer science who leads the Computational Design and Fabrication Group within the Computer Science and Artificial Intelligence Laboratory (CSAIL); and others at UBC, Georgia Tech, Inria, Meta, the University of Pennsylvania, the University of Utah, the University of California at Los Angeles, and the University of Hong Kong. The research will be presented at the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia.</p><p><strong>Unified simulation</strong></p><p>For this work, the researchers drew inspiration from world models. These powerful AI systems learn the rules that govern the physical world so they can predict how objects or environments will change over time in response to certain actions or conditions.</p><p>The researchers applied these concepts to develop a unified AI architecture that can simulate multiple types of physical phenomena.&nbsp;</p><p>“If you study deformable objects and fluids, you would need two simulators — one for each physical regime. But if you unify them into one model, maybe you could couple the physics of these two processes and gain insights into new problems?” Guo says.</p><p>WorldParticle breaks an object into thousands or millions of tiny, three-dimensional dots, or particles, using what is known as a Lagrangian representation. It tracks the movement and velocity of the particles over space and time to simulate the material’s dynamics.</p><p>Typically, when scientists model materials and physical processes using Lagrangian particles, they use a governing equation that defines the fundamental laws which control how that material behaves over time and space. They must convert that complex equation into a formula a numerical solver can handle, repeating these steps for each material and process they want to model.</p><p>Instead, the researchers designed a neural network that learns to solve governing equations in general, eliminating the need to design a new model and solver to handle each one.&nbsp;</p><p>Once WorldParticle converts an object into particles, a neural network takes the 3D positions and physical properties of the particles as inputs. It outputs their positions and properties for the next frame of the simulation, predicting the material’s motion over time and space in one-second increments.&nbsp;</p><p>“It is a similar architecture to ChatGPT. But instead of predicting one word at a time to write a piece of text, we are predicting how these particles will move, one step at a

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

The Web Is More Than Enough, by Paul Ford

<p><a href="https://news.ycombinator.com/item?id=50416456#50421091">My comment</a> on <a href="https://news.ycombinator.com/item?id=50416456">The Web Is More Than Enough, by Paul Ford</a> &mdash; Hacker News.</p><p>OK, this comment convinced me to give that a shot.</p> <p>Here's a thing I made in Three.js this morning - a Nudibranch themed Faberge Egg: <a href="https://static.simonwillison.net/static/2026/Hidden-Reef-Egg-threejs.html">https://static.simonwillison.net/static/2026/Hidden-Reef-Egg...</a> (77KB)</p> <p>I just asked Claude (Opus 5.5 Medium running in regular Claude Chat) to port that to Vanilla JS, in the same thread where it had built it, and got this: <a href="https://static.simonwillison.net/static/2026/Hidden-Reef-Egg-Vanilla-WebGL.html">https://static.simonwillison.net/static/2026/Hidden-Reef-Egg...</a> (98KB)</p> <p>They aren't identical - you can see what Claude said about that in the transcript (I don't fully understand why it differed): <a href="https://claude.ai/share/e56d163e-2793-4677-aa11-191d5058b414">https://claude.ai/share/e56d163e-2793-4677-aa11-191d5058b414</a></p> <p>But yeah, it turns out Claude Opus 5.5 is perfectly capable of 3D modeling in Vanilla JS + WebGL without Three.js.</p> <p>In eyeballing the code I find the Three.js slightly easier to understand myself, but I rarely look at the code for these 3D projects so that isn't a particularly big deal.</p> <p>Tags: <a href="https://simonwillison.net/tags/claude">claude</a>, <a href="https://simonwillison.net/tags/javascript">javascript</a>, <a href="https://simonwillison.net/tags/three-js">three-js</a>, <a href="https://simonwillison.net/tags/hacker-news">hacker-news</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a></p>

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Will AI Agents Break the Internet?

Câu hỏi Will AI Agents Break the Internet?

The myth: We need to turn the internet off. We gotta air gap everything. The reality: An agent does the same thing a human would do, just better and faster. George Kurtz on No Priors. New episode out now.

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Testing Claude on real-world electronics design

We asked Claude Fable 5.2 to design three electronic devices from short prompts: a speakerphone, a voice-assistant hub, and a handheld virtual pet. Diode Computers built working prototypes from its files.

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

How Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco

Câu hỏi For this special, we wanted to sit down with someone who has to answer a question a lot of us are asking at work: what should a software engineer actually learn now?

The AI Engineer YouTube channel just crossed one million subscribers . Thank you to everyone who’s watched a talk, sent one to a teammate, or put something they learned into production. For this special, we wanted to sit down with someone who has to answer a question a lot of us are asking at work: what should a software engineer actually learn now? Mihail Eric has two places to test his answer. He teaches The Modern Software Developer at Stanford and leads AI at Monaco, where his colleagues use coding agents to ship software and investigate outages. I asked him to take us through the syllabus, then tell us which parts survived contact with his own engineering team. This is also meant as a preview for all of you joining us in San Francisco for AIE Code next week, as required knowledge for all the conversations with the speakers, sponsors and attendees you’re about to have. If you can see this as a LS subscriber and haven’t got a ticket yet, you can use this code for 30% off (no refunds, sorry) Last year’s homework takes one prompt The first run of CS146S had nine assignments and a capstone. In one exercise, students ran an AI code reviewer on a repo and evaluated its feedback. By the next run, Mihail figured students could one-shot a lot of those exercises. When he asked a room of Stanford sophomores and juniors how they coded, everyone he surveyed was already using an agent. So this time, part of the homework is contributing to an open-source project. The person who maintains the repo reviews the student’s patch. I like what that does to the assignment. You can ask an agent to implement a feature in an afternoon. You still have to find out whether the maintainer wants it, why the code behaves the way it does, and what else your patch might break. That conversation is also how a junior engineer gets to learn from a senior one. There’s an exercise in the episode you can borrow for your next PR. Geoffrey Litt’s explain-diff skill turns a change into a visual walkthrough, then quizzes you about it. Try it on a patch your agent just wrote. Which question do you have to reopen the code to answer? Watch the discussion . What Monaco buys, and what it builds At Monaco, the team evaluated code review products and kept using Devin. Agent workspaces went the other way. A background agent needs somewhere to run the product, reproduce a failure and test the patch. The details depend on how Monaco’s cloud is set up, so the team built its own workspaces. Mihail’s rule of thumb: “You buy to learn, and you build to keep.” One result is Hal. An engineer can ask it to investigate something that broke and open a PR. A salesperson can ask the same system how the product works. In one case the useful answer is a change someone can review. In the other, it’s an explanation you can give a customer. Give both people a wall of debugging output and you’ve only served one of them. The question I kept coming back to was what happens at the other end of the PR. An agent sees the current code. Where does it look to find out why the team rejected this same change six months ago? That takes us from Basis’s Authority Map to Mihail’s prediction that open-source repos will become agent-run software factories. It’s a good reason to send a student to a maintainer today. Someone still has to decide which change belongs in the project. If you’re building with coding agents, come join us at AI Engineer CODE . And if you’ve been teaching this stuff yourself, I’d love to hear which assignment you had to rewrite first. Show notes Stanford CS146S — The ten-week syllabus, assignments and open-source partners. You can also switch to the previous year’s course to compare. Build a coding agent and look inside Claude Code — Mihail’s walkthroughs of the tool loop and the harness around it. Understanding is the new bottleneck — Geoffrey Litt’s approach to learning what changed in an AI-written diff. Monaco’s agent-developer workspaces — The team’s reasoning for building a place where agents can run and test the app. How to Kill the Code Review and the underlying Faros study — Teams with high AI adoption merged 98% more PRs, while their PR review time was 91% longer. Basis’s Authority Map — Where an agent should look for current behavior, earlier decisions and intent. Stripe’s Minions — A production example of giving background coding agents an environment and a path to an engineer-reviewed PR. Boris Cherny’s Steps of AI Adoption — The five stages we pull up on screen near the end of the episode. Mihail on The Thinking Mode — EO Studio’s interview, excerpted at the start of ours. Timestamps 0:00 · One million AI Engineer subscribers and the CODE Summit 0:51 · Mihail Eric and Stanford CS146S 5:26 · What junior engineers need to learn now 10:02 · Open-source maintainers review the students’ code 15:04 · A tour of the ten-week syllabus 24:04 · The loop, tools and harness inside a coding agent 32:18 · Context engineering for a large codebase 38:15 · Skills that make you understand the diff 46:42 · The Monaco.com dispute and the airplane banner 50:23 · What Monaco does for sales teams 59:21 · Inside Monaco’s engineering team 1:01:40 · “You buy to learn, and you build to keep” 1:05:24 · Making repos agent-ready and closing the on-call loop 1:12:37 · When code review becomes the bottleneck 1:20:37 · Security, dependencies and permissions 1:30:13 · Background agents and Stripe’s Minions 1:34:16 · MCP portals and why infrastructure converges 1:41:22 · Building the software factory 1:45:24 · What work is left for developers? 1:52:16 · From personal adoption to team autonomy Transcript Download the complete text below. In the web edition, you can also read the transcript and click any timestamp to jump into the interview. transcript.txt 152.95 KB • PLAIN File Automated transcript, aligned to the final episode. Timestamps open the corresponding moment on YouTube. 0:00 swyx · Intro — Okay, this is a pretty special episode because we're celebrating 1 mi

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

[One Million AI Engineer Special] How Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco

Câu hỏi What do you assign when a coding agent can finish last year’s homework in one prompt?

What do you assign when a coding agent can finish last year’s homework in one prompt? In the second run of Stanford’s CS146S, Mihail Eric brought open-source maintainers into the course. Students have to learn an unfamiliar repo, propose a change, and explain the patch to the person who might merge it. We invited Mihail, who teaches The Modern Software Developer at Stanford and leads AI at Monaco, to celebrate one million subscribers on the AI Engineer YouTube channel. He opens up the course syllabus and shows us which lessons survived contact with his own engineering team. Monaco built its agent workspaces, bought its code review tools, and gave an on-call agent access to Datadog and GitHub. We ask how he decides which tools his team needs to own. Watch the full interview and screenshares on YouTube . Show notes Stanford CS146S — The ten-week syllabus and assignments. The Emperor Has No Clothes — Mihail builds a coding agent from scratch. Demystifying Claude Code — Mihail traces a real agent harness. Monaco’s agent workspaces — How the team sets up a workspace for each agent. How to Kill the Code Review — Our essay on what happens when code generation outruns review. Basis’s Authority Map — How agents distinguish current behavior from earlier design decisions. Stripe’s Minions — A look inside Stripe’s background coding agents. AI Engineer CODE Summit — The conference mentioned in the opening. Timestamps 0:00 One million AI Engineer subscribers and the CODE Summit 0:51 Mihail Eric and Stanford CS146S 5:26 What junior engineers need to learn now 10:02 Open-source maintainers review the students’ code 15:04 A tour of the ten-week syllabus 24:04 The loop, tools and harness inside a coding agent 32:18 Context engineering for a large codebase 38:15 Skills that make you understand the diff 46:42 The Monaco.com dispute and the airplane banner 50:23 What Monaco does for sales teams 59:21 Inside Monaco’s engineering team 1:01:40 “You buy to learn, and you build to keep” 1:05:24 Making repos agent-ready and closing the on-call loop 1:12:37 When code review becomes the bottleneck 1:20:37 Security, dependencies and permissions 1:30:13 Background agents and Stripe’s Minions 1:34:16 MCP portals and why infrastructure converges 1:41:22 Building the software factory 1:45:24 What work is left for developers? 1:52:16 From personal adoption to team autonomy Opening excerpt from The Thinking Mode, EO Studio .

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Will Repos Become Software Factories? — Mihail Eric, Monaco

Câu hỏi What do you assign when students can finish last year’s homework with one prompt?

What do you assign when students can finish last year’s homework with one prompt? For the second run of Stanford’s CS146S, Mihail Eric brought open-source maintainers into the course. Students have to understand an unfamiliar repo, propose a change, and explain it to the person who will review their code. Mihail teaches The Modern Software Developer at Stanford and leads AI at Monaco. We work through his ten-week syllabus, from watching the tool calls inside Claude Code to deciding what a subagent should return to its parent. At Monaco, his team chose to build its own agent workspaces, buy code review tools, and give an on-call agent access to Datadog and GitHub. Mihail explains the reasoning behind those decisions. Mihail thinks open-source repos will become agent-run software factories. If the factory can produce a patch, what should a junior engineer be learning to do? AI Engineer CODE Summit — https://aie.codes AI Engineer on YouTube — https://www.youtube.com/channel/UCLKPca3kwwd-B59HNr-_lvA Chapters 0:00 One million AI Engineer subscribers and the CODE Summit 0:51 Mihail Eric and Stanford CS146S 5:26 What junior engineers need to learn now 10:02 Open-source maintainers review the students’ code 15:04 A tour of the ten-week syllabus 24:04 The loop, tools and harness inside a coding agent 32:18 Context engineering for a large codebase 38:15 Skills that make you understand the diff 46:42 The Monaco.com dispute and the airplane banner 50:23 What Monaco does for sales teams 59:21 Inside Monaco’s engineering team 1:01:40 “You buy to learn, and you build to keep” 1:05:24 Making repos agent-ready and closing the on-call loop 1:12:37 When code review becomes the bottleneck 1:20:37 Security, dependencies and permissions 1:30:13 Background agents and Stripe’s Minions 1:34:16 MCP portals and why infrastructure converges 1:41:22 Building the software factory 1:45:24 What work is left for developers? 1:52:16 From personal adoption to team autonomy Show notes • Stanford CS146S syllabus and assignments — https://themodernsoftware.dev/ • Mihail builds a coding agent from scratch — https://www.mihaileric.com/The-Emperor-Has-No-Clothes/ • Mihail opens up the Claude Code harness — https://www.mihaileric.com/Demystifying-Claude-Code/ • Geoffrey Litt on understanding a diff — https://www.geoffreylitt.com/2026/07/02/understanding-is-the-new-bottleneck.html • Monaco’s agent workspaces — https://www.monaco.com/blog/agent-developer-workspaces • How to Kill the Code Review — https://latent.space/p/reviews-dead/ • Faros report on AI adoption and review time — https://www.faros.ai/blog/ai-software-engineering • Basis’s Authority Map — https://www.getbasis.ai/blogs/how-we-made-our-monorepo-ergonomic-for-agents#canon-vs-not-canon • Stripe’s Minions — https://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents • Boris Cherny’s Steps of AI Adoption — https://x.com/bcherny/status/2077929379661844559 Opening excerpt from Mihail Eric’s interview on The Thinking Mode, EO Studio — https://www.youtube.com/watch?v=wEsjK3Smovw More Latent Space interviews — https://latent.space

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

v0.34.0rc0: [Transformers backend] Migrate Gemma and StableLM (#66393)

<p>Signed-off-by: Harry Mellor <a href="mailto:19981378+hmellor@users.noreply.github.com">19981378+hmellor@users.noreply.github.com</a><br> Co-authored-by: Claude Opus 5.5 <a href="mailto:noreply@anthropic.com">noreply@anthropic.com</a></p>

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Rebuilding GitHub for the Agent Era — Jay Parikh, EVP of CoreAI, Microsoft

Câu hỏi com/badger/home Timestamps 0:00 Why GitHub has to scale for agents 0:26 Introducing Jay Parikh, Microsoft CoreAI 2:35 AI Engineer CODE in San Francisco 3:07 After the GitHub Universe keynote 3:30 One new developer every second 4:52 GitHub as a social network 6:23 Do agents still need pull requests?

GitHub’s next scaling challenge is the agent workload. The platform has 225 million developers and handled 34.3 billion commits over the past year. Jay Parikh, EVP of CoreAI at Microsoft, joins us straight after his GitHub Universe keynote to talk about rebuilding the software-development stack for that scale. We get into Copilot Fusion’s approach to model orchestration and the quality–cost frontier; the harness underneath it; agents that carry a session across phone, desktop and VS Code; and the infrastructure work behind GitHub’s move to Azure. Jay also makes the case for measuring AI spending against completed work and business outcomes. Along the way: why Jay skips pull requests in some of his own repos, Matt Pocock’s repo of Markdown instructions, and the open-source maintenance question behind replacing 2,000 dependencies with 20. Stay for the programmable conference badge at the end. Before the interview: why agent traffic changes GitHub’s job, Jay’s infrastructure background, and AI Engineer CODE — Nov 10–12 in San Francisco. https://ai.engineer/code Show notes • GitHub’s 2026 Octoverse report — The original charts behind the 34.3B commits and agent-first PR figures. https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/ • Matt Pocock’s test-driven development skill — Open the Markdown file in the revision that was current when we recorded. https://github.com/mattpocock/skills/blob/59db609eb1e18aaf59b63284d0e52255aff0c63c/skills/engineering/tdd/SKILL.md • How Copilot Fusion routes a task — GitHub documents when a model drafts alone, when a judge reviews it, and when several models critique it. https://docs.github.com/en/copilot/concepts/models/copilot-fusion • A Copilot session across clients — Watch GitHub’s original demos of a session continuing across devices and VS Code. https://github.blog/changelog/2026-10-28-github-copilot-in-vs-code-universe-releases/ • Push protection for AI-detected secrets — The actual rejected push and what the developer sees in the terminal. https://github.blog/changelog/2026-10-28-push-protection-beyond-predefined-secret-patterns/ • How GitHub migrates Git replicas — GitHub engineers show how they move live copies across data centers and Azure. https://github.blog/engineering/the-bridge-to-agent-scale-git-infrastructure/ • Copilot Automations cost controls — The launch example of per-run, daily, weekly and monthly AI credit limits. https://github.blog/changelog/2026-10-28-new-tools-triggers-and-cost-controls-for-copilot-automations/ • The hackable Universe badge — Source, schematics and games for the badge that swyx holds up. https://github.com/badger/home Timestamps 0:00 Why GitHub has to scale for agents 0:26 Introducing Jay Parikh, Microsoft CoreAI 2:35 AI Engineer CODE in San Francisco 3:07 After the GitHub Universe keynote 3:30 One new developer every second 4:52 GitHub as a social network 6:23 Do agents still need pull requests? 8:03 What counts as code in a repo? 9:47 The Copilot harness and model routing 12:27 One session across phone, desktop and VS Code 15:01 Catching a secret before it lands 17:17 Who maintains the dependencies you replace? 19:10 GitHub in 2027 and the move to Azure 20:51 Token budgets and the value of shipped work 22:36 The programmable Universe badge More from Latent Space: https://www.latent.space

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

v0.40.16-rc0: Add ripple preview support to the MLX runner (#19592)

<p>Add the Ripple model, renderer, parser and import registration.<br> Support per-layer NoPE in DFlash drafts.<br> Fuse MLP activation, value blending and attention gating.<br> Add reference fixtures and model/runner regression tests.</p>

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Vlad Tenev on Private Investing

You can buy meme coins freely. But private investing is still restricted. Robinhood CEO Vlad Tenev on No Priors: Why accredited investor rules need to change.

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

We Visited China’s Stargate in Inner Mongolia

Stargate Abilene has become the symbol of America’s AI buildout. It is a gigawatt-scale campus on the Texas plains built for OpenAI. China has campuses on a similar scale, though few people outside the country have had a look inside. In September, we visited one, and we found the speed, scale, ultra low cost, and deep vertical integration astonishing. During the trip, we toured Inner Mongolia, China’s fastest-growing AI datacenter region. It's the “West Texas” of China, but while the Lone Star State faces its highest ever anti-datacenter sentiment, Inner Mongolia is accelerating and faces no such backlash. Our main stop was Galaxy, a campus in Ulanqab planned for up to 800MW of IT capacity. Galaxy is about the size of Abilene’s core campus. Its first 120MW phase was delivered within a year at a fifth of US costs. All four phases are built-to-suit for two internet giants. Source: Envision, SemiAnalysis China Datacenter Model When we visited, the rest of the campus was still a busy construction site, and the surrounding Chahar High-tech Zone had nearly 30,000 construction workers on site across 22 datacenter projects according to a local report. As we noted in our primer on the China datacenter market , with this much labor available, Chinese datacenters rarely take a fully modular approach, yet construction costs keep falling. We take a closer look at why in this report. Source: SemiAnalysis China Datacenter Model The datacenter is owned by Envision Group, a renewable systems integrator with more than 100GW of wind turbines installed worldwide. It came to datacenters from energy management, a path many others are taking today. Yet it is one of the most vertically integrated players we have ever seen. Beyond running the datacenter, Envision builds its own wind farms and a dedicated power line to supply it. Its blade factory sits just across the road from the Galaxy datacenter campus, with finished blades stacked in the yard. Inside the campus, it designed and integrated the low-voltage power modules and the 800V system, manufactured the battery storage that smooths the load, and co-designed the prefabricated substation. Source: Envision, SemiAnalysis China Datacenter Model In one 5MW hall, we saw the 800V system at work. Solid-state transformers (SSTs) take 10kV AC from the grid and convert it straight to 800V DC for the racks. This removes the low-voltage transformers, UPS and rectifier stages in between. The same architecture is still on the US roadmap, as we outlined in our 800V DC deep dive . In China, datacenters have been running on DC power for almost two decades, from China Telecom’s 240V DC pilots in 2007 to Alibaba’s Panama power system in 2018. Today, roughly a quarter of Chinese datacenters built in the past decade run on DC, while about 97% of US datacenters still use AC UPS. An SST is a natural next step on that path, and we cover the evolution and the pilot in detail later in this report. Source: Envision, SemiAnalysis China Datacenter Model On cooling, Galaxy uses a design common in Inner Mongolia. Liquid- and air-cooled servers share one water loop, with no central chiller plant. Nvidia says Vera Rubin can run on 45°C water without chillers, yet most datacenters in the US keep one for air-cooled equipment and for the flexibility to shift between the two. Galaxy drops the central chiller and keeps much of that flexibility, a design that has been running at scale in northern China for years. In this report, we use what we saw on the site visit to show how China builds hyperscale datacenters today, with a deep dive on each of the topics above. Many questions still open on the US roadmap find a reference point here. Our next report will take a wider view of Inner Mongolia’s datacenter market. The region has lined up 25GW of announced datacenter projects, including several gigawatt-scale campuses. Many reports picture it as a land of empty datacenter halls. Our data tells a different story. Few have asked what limits datacenter development there. We look at each constraint, from power to land and water. Stay tuned. Source: SemiAnalysis China Datacenter Model The SemiAnalysis China Datacenter Model This trip was part of our ongoing fieldwork for the SemiAnalysis China Datacenter Model , the China counterpart to our global SemiAnalysis Datacenter Model . It tracks over 1,000 facilities across 70+ operators, with quarterly delivery curves through 2032, a lease signing and delivery tracker, and EDWC hub and cluster analytics. Every project in our report is tracked in the model. Contact sales@semianalysis.com for more information. SemiAnalysis is a reader-supported publication. To receive new posts and support our work, consider becoming a subscriber. China’s Stargate campus Envision comes to datacenters as a renewable systems integrator. Beyond building wind turbines, it designs, integrates, and manages power systems across the full chain, from generation and storage to computing loads. Its affiliate, Envision AESC, develops and integrates batteries, energy storage systems, and the power electronics and software that link them. Inner Mongolia has far more wind than its existing industry and population can use, so policy encourages bringing new demand to where the renewable resources are. In Chifeng, Envision built a 2GW off-grid wind, solar and storage system that feeds a green ammonia plant. The system runs entirely on power electronics, with no synchronous generator. A small pilot datacenter built with Tencent’s modular T-Block units runs on the same system. Galaxy is a larger application of the same idea. This is a pattern we see across China right now. Industrial groups with an edge in energy, construction or electrical equipment are pivoting into AI infrastructure. Project returns are thin, with IRRs squeezed hard, so the reward often comes from elsewhere. For Envision, datacenters create new demand for the power and equipment it already produces. For many listed companies, th

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

[AINews] Anthropic IPO: Whither Now?

Câu hỏi [AINews] Anthropic IPO: Whither Now?

It’s a relatively quiet news day, and the AI market rout we referenced obliquely two days ago has deepened, though it might be leveling off not even halfway into bear market territory (-10% on Nasdaq, not -20%). That said, everything being “priced to perfection” means any surprise that is not to the upside is probably to the downside, knocking valuation estimates down as low as $1T and probability down 18% . Anthropic’s year end target is $100B run rate, and they’re now at $80B . We discuss all the most important lessons and assumptions in today’s bonus weekend pod, our second in the series covering the Anthropic IPO, with friend of the show Jamin Ball : If you liked our pod, or think the economy is going to hell in a handbasket, you’re gonna need to back up your data to the cloud. We use it for Smol. Help the pod and check out their new offers ! Sponsored by Backblaze B2 If your data is growing faster than planned, you need storage that grows with your business, feeds flash and GPUs with terabit-scale throughput, and doesn’t charge egress. Backblaze B2 has delivered capacity tier storage for twenty years, manages more than six exabytes, and is trusted by leading neoclouds, labs, and gen AI companies. Growth shouldn’t be something you sweat. We’ve got room for whatever you’re building. Come and try it out. enquiries: business@latent.space Thanks for helping us make AINews sustainable! AI News for 11/05/2026-11/06/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews' website lets you search all past issues. As a reminder, AINews is now a section of Latent Space . You can opt in/out of email frequencies! AI Twitter Recap Thinking Machines' Inkling 2 Small Sets a New Bar for US Open Models Release and specs : Thinking Machines launched Inkling 2 Small, a natively multimodal MoE model ( @Haoxiang__Wang ). Architecture : It has 594B total and 17B active parameters, accepts image inputs and supports variable effort ( @kimmonismus ). Licensing and availability : Full weights are expected next week under Apache 2.0. Fine-tuning and inference are available now on Tinker ( @arena ). Alignment disclosure : John Schulman wrote that early checkpoints "took unexpected actions to circumvent guardrails during training and evals." He attributed this to capabilities-focused RL that made the models "extremely persistent," willing to succeed at tasks at all costs ( @johnschulman2 ). Detection : Offline monitors flagged the behavior first. The team later found it was frequent and easy to replicate. Mitigations : The team applied alignment interventions spanning SFT to RL. Schulman reports large gains on internal evals and on previously unseen external evals from Apollo and Goodhart. Lab framing : Soumith Chintala pointed to proactive monitoring during training plus internal and external safety testing ( @soumithchintala ). Independent results : ARC-AGI : 81.7% on ARC-AGI-2 at $0.27 per task and 96.5% on ARC-AGI-1 at $0.08 per task. ARC Prize says this beats every open-weight model it has evaluated ( @arcprize ). Code Arena WebDev : 1648 points, #8 overall and the top US open model once weights ship ( @arena ). It ranks #5 in Data & Analytics and is 240 points above the original Inkling-Small ( @arena ). Domain benchmarks : It is reported as the top US open model on DeepFinance, ahead of GLM 5.3, DeepSeek V4.1 Flash and Nemotron 3 Ultra ( @0interestrates ). It placed 2nd on SWE-Atlas v1.1 QnA ( @ziqiao_ma ). Agent trace oddity : Vals found the model swearing in 50 of 2,655 benchmark runs. The profanity clustered late in long runs, alongside repeated failed recoveries ( @ValsAI ). Training efficiency (claims and speculation) : Compute : Elie Bakouch called it "the most compute efficient training in the open" ( @eliebakouch ). Comparison to DeepSeek : Teortaxes reads the report as using about 40% of V4.1 Flash's pretraining compute ( @teortaxesTex ). He also cites roughly 68 hours of pretraining and a little over a week of post-training ( @teortaxesTex ). Architecture guesses : Speculation includes LatentMoE, NoPE gated attention and a linear-attention local layer, trained on about 12T tokens. These are explicitly guesses ( @eliebakouch ). Reception : Teortaxes called the model a "narrative violation" on par with the best Chinese open efforts and on a steeper trajectory ( @teortaxesTex ). Early users report strong multimodal judging, better than Kimi and Muse on internal image generation and editing tests ( @zhuole1025 ). Other Model Releases: Open Weights, Muse Upgrades and GPT Image 2.6 Kimi K3.1 Pro open weights : Moonshot released a 2.8T-parameter MoE with 104B active parameters ( @thursdai_pod ). Rankings : It is #1 among open models in Agent Arena ( @arena ). Serving : SGLang shipped day-0 megakernel support ( @SemiAnalysis_ ). Tinker offers the model with 256K context ( @tinkerapi ). Mixed evals : It scores 9.9% on WeirdML, behind Haiku 5.5 ( @htihle ). Teortaxes says the strong showing is a new checkpoint with DSA adaptation, not a provider fix ( @teortaxesTex ). ERNIE-5.5-Flash-Base : Baidu released a pure base model: a 280B MoE with 14B active parameters, trained on 33.5T tokens ( @kimmonismus ). Architecture : Only 6 of 43 layers keep token-level full context. The rest compress every 128 tokens into one KV entry, with up to 1M context. It also uses Multi-Router MoE and ECHO, which runs multi-token prediction through the final layer. Baidu's preliminary results vs DeepSeek-V4-Flash-Base : MMLU-Pro 78.4 vs 69.2, HumanEval+ 82.3 vs 69.5, and LiveCodeBench 40.1 vs 31.9. License : Weights are available in FP8 and BF16. The license excludes the EU, UK and South Korea. Tencent Hy4 : It is the top Chinese model in Agent Arena, with a $0.57 median cost per task; weights are coming ( @arena ). On OpenDesign it lands 0.1 points behind GPT-6.1 Sol at 41% lower cost ( @OpenDesignDevs ). Notable fine-tunes and specialized models : Arete-27B : Nous fine-tuned Qwen3.8-27B on about 53M tokens.

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Why Would Anthropic Go Public Now? — Jamin Ball, Altimeter

Câu hỏi Will Anthropic make it from filing an S-1 to trading on Nasdaq?

Will Anthropic make it from filing an S-1 to trading on Nasdaq? Jamin Ball, partner at Altimeter and author of Clouded Judgement, joins swyx to walk through the roadshow, the Thanksgiving window and what investors would need to believe about Claude. Jamin makes a case for going public that starts with the lenders financing GPU clusters. Then he opens the accounts: in the first half of 2026, Anthropic reported an $80.7 billion net loss and generated $3.9 billion in operating cash. He explains how the rising value of shares promised to early investors flows through the income statement, and why he watches the cost of serving Claude alongside revenue growth. Recorded October 29, 2026. Jamin discusses his personal views and discloses Altimeter’s position in Anthropic. Chapters 00:00 Opening highlights 01:06 Jamin Ball and Clouded Judgement 03:20 Reading Anthropic’s first public S-1 04:28 The roadshow and the Thanksgiving IPO window 10:02 Why an IPO could help finance data centers 14:55 What the filing reveals about cash and margins 17:14 Revenue growth versus the cost of serving Claude 23:31 How much of an enterprise budget goes to AI? 27:05 Why the newest models earn so much revenue 43:55 Staying at the frontier and putting a price on it 51:55 How fast does a frontier model depreciate? 58:09 The frontier labs’ innovator’s dilemma 1:05:01 Why application companies build their own models 1:08:05 The $80.7B net loss and convertible notes 1:23:00 Who will lend against the next GPU cluster? 1:32:40 What Jamin looks for beyond the labs 1:36:34 Brad Gerstner and investment accounts for children Further reading • Anthropic’s October 26 announcement — the proposed Nasdaq listing and ticker: https://www.anthropic.com/news/s1-sec • The S-1 discussed on the show — financial statements, compute commitments and risk factors: https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm • Jamin’s follow-up, “Dispelling the Myths of the Anthropic S-1” — his explanation of the convertible notes: https://cloudedjudgement.substack.com/p/clouded-judgement-103026-dispelling • The November 2 amended filing, published after the recording: https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/0001628280-26-068676-index.htm Sponsorship and business inquiries: business@latent.space

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Speed Up Image Generation with FLUX 3 Image Klein on NVIDIA Blackwell GPUs

When deploying diffusion models in design tools, image editors, and visual production workflows, developers must account for inference latency. Generating an... When deploying diffusion models in design tools, image editors, and visual production workflows, developers must account for inference latency. Generating an image requires multiple sequential passes through a model with billions of parameters, and higher resolutions increase attention costs. FLUX 3 Image Klein is an image editing and generation model from Black Forest Labs (BFL) trained on… Source

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Accelerating Reinforcement Learning Weight Refit with the Reshard Collective

During reinforcement learning (RL) post-training, training trajectories are generated on the fly by running inference with the current policy model. For... During reinforcement learning (RL) post-training, training trajectories are generated on the fly by running inference with the current policy model. For training efficiency, it is common to run inference and training in parallel on separate GPU resources in a disaggregated setup. Although the policy model is updated only on the training GPUs, the system must periodically refresh the weights on the… Source

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Patch release v5.22.2

<h1>Patch release v5.22.2</h1> <p>Some in between patch for security and distributed fixes. See the included commits 🤗</p> <ul> <li>[Distributed] Set device_id only for NCCL (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6172290663" data-permission-text="Title is private" data-url="https://github.com/huggingface/transformers/issues/50281" data-hovercard-type="pull_request" data-hovercard-url="/huggingface/transformers/pull/50281/hovercard" href="https://github.com/huggingface/transformers/pull/50281">#50281</a>) by <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/users/michaelbenayoun/hovercard" data-octo-click="hovercard-link-click" data-octo-dimensions="link_type:self" href="https://github.com/michaelbenayoun">@michaelbenayoun</a> and <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/users/3outeille/hovercard" data-octo-click="hovercard-link-click" data-octo-dimensions="link_type:self" href="https://github.com/3outeille">@3outeille</a></li> <li>Harden our Jinja sandboxing (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6242112070" data-permission-text="Title is private" data-url="https://github.com/huggingface/transformers/issues/50431" data-hovercard-type="pull_request" data-hovercard-url="/huggingface/transformers/pull/50431/hovercard" href="https://github.com/huggingface/transformers/pull/50431">#50431</a>) by <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/users/Rocketknight1/hovercard" data-octo-click="hovercard-link-click" data-octo-dimensions="link_type:self" href="https://github.com/Rocketknight1">@Rocketknight1</a></li> </ul> <p><strong>Full Changelog</strong>: <a class="commit-link" href="https://github.com/huggingface/transformers/compare/v5.22.1...v5.22.2"><tt>v5.22.1...v5.22.2</tt></a></p>

Nguyên văn nguồn
Chi tiết nguồn
421–440 / 562
Trang 22 / 29