Energy-Based Models can now scale better than Diffusion models. Introducing TEM: Transformer Energy-based Models TLDR: • TEM outscales Diffusion without guidance • First scalable contrastive Energy-Based Models • Better generalization, test-time scaling, and OOD detection than Diffusion 🧵

Can a JEPA world model provably avoid contraction and merging of latent states over long rollouts? Unconstrained autoregressive predictors may contract the latent distribution. Our Symplectic Hamiltonian JEPA (LeHami) replaces the predictor with an action-conditioned Hamiltonian flow over a latent state (q, p), integrated with an exactly symplectic leapfrog step. Volume is preserved at every rollout step, independently of the learned parameters or action sequence. With @randall_balestriero @ylecun Paper: https://t.co/JqJ7rrOXmF Website: https://t.co/I5pmtHM9VB Code: https://t.co/D9EIFGS87U 1/6... 🧵👇

Last night, @higgsfield sent @ComfyUI a cease and desist (attached) They want the Open Source Higgsfield challenge taken down within 48 hours. Their argument is that our use of the Higgsfield name and that using their CEO's likeness is an infringement violation (man, the irony). Claiming that it makes the event look like Higgsfield is associated with it. This couldn't be farther from the truth, which is that Higgsfield made a promise to open source its product and never did. For the past two weeks, builders have been recreating popular closed apps as open workflows anyone can run. So we asked our community to deliver where Higgsfield failed: apps in the open, built by the community, for the community. We've made it clear that the challenge is organized by ComfyUI and not affiliated with or endorsed by Higgsfield. When we started Comfy, we made the community a promise. Our product stays open, free, and yours. We’ve kept it every day since, and one letter isn’t going to change that. The challenge stays up. Comfy and open source will win!

+1 more

I improved NanoGPT Speedrun from 39.9s → 6.7s (-83.2%), more than every record since 2025 combined (-81.4%)! The web is full of rare, repeated content that the LM can’t memorize in time. I solve this with retrieval-augmented n-gram hints🧵 https://t.co/PKQdSpWdxP

Quick quality life of upgrade if you have lots of tasks in Codex Tell codex to rename all the tasks in your project according to this article You’ll be able to find stuff again and understand what’s happening!

Evals built by @PrimeIntellect residents, fully open: - MazeBench by @patience_cave - PMPP-Hard by @myainotez - PretrainGym by @HuNtEr_J1324 - OpenTTD-Bench by @thesophiaxu [and even more :) ] Super happy to support such amazing work, check it out!

they built a modular robotics platform that can transform into 5 different robots a biped, quadruped, wheeled robot, or a 7dof robotic arm, all powered by the same actuators this means researchers can train entirely different robotic capabilities for under $6k USD https://t.co/MRRIHAlnOb

Today, we’re sharing a look at our expanded system for monitoring frontier models at OpenAI, which in September helped us catch a model’s attempt to access the live internet during training. Since the Hugging Face incident—when models broke out of an internal cybersecurity investigation where we weren’t running our misalignment monitors—we’ve substantially expanded our monitoring of frontier agents. We now monitor all frontier reinforcement learning runs and evaluations, as well as all internal Codex use. When our system finds evidence of a problem that might require intervention, it rapidly alerts human responders who can investigate and take action. To detect when a model’s behavior departs from what we intended, we use several layers of monitoring. Activation probes look for patterns in a model’s internal activity. Reviewer models examine its reasoning and tool use, and investigative agents can read back through the model’s entire transcript to understand how a sequence of actions fits together and whether it warrants human review. To test whether our current system would have detected the Hugging Face incident, we replayed the recorded activity with our new monitors in place. In the replay, our monitors alerted our security team about 30 minutes after the agents escaped their sandbox. That was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident. In September, a model exploited a DNS misconfiguration to reach the live internet. Our monitors flagged the breach within 15 minutes, and a responder began reviewing the alert three minutes later. The alert enabled responders to investigate while the run was still underway. As we give models more autonomy, we need to know when they cross a boundary we expected to hold. We believe that monitoring is essential to catching those failures in time for people to intervene. https://t.co/dIYP7B1GlD

Today we’re sharing an experimental release of Spark Radix, a WebGPU-powered Gaussian splat renderer built on three.js. Render and explore detailed 3D worlds directly in your browser. Available today for developers to try and test 👇 https://t.co/ecw63Ljen0

Control flow just got more expressive in Mojo! The new match/case construct goes beyond simple value matching. Combine exact matches, conditional guards, and catch-all cases to handle complex decisions with clear, readable code. More flexibility. Less branching. More Mojo. 🔥 Continue exploring Mojo's superpowers with our cheat sheets: https://t.co/UiEhSYlR7T

Evals built by @PrimeIntellect residents, fully open: - MazeBench by @patience_cave - PMPP-Hard by @myainotez - PretrainGym by @HuNtEr_J1324 - OpenTTD-Bench by @thesophiaxu [and even more :) ] Super happy to support such amazing work, check it out!

Before he logs off, @mitchellh sends Codex agents to compare tools and prototype features he’s still figuring out. He picks up from there the next morning. https://t.co/AXVctw8Es5

We’ve updated the Model Spec, our public description of how we want our models to behave. This revision makes clearer what we want from models as they take on more autonomous work in the world and become capable of helping us in new ways. 🧵 https://t.co/EdBbpGwyGS

3/ The results: - LlamaParse Agentic and Agentic Plus: 179/179, zero Boos - Best frontier VLMs (GPT-6 Luna, GPT-6.1 Terra): 99.4% - Open VLMs: 61% to 93%

2/ The pages are ordinary and from real documents: dot leaders, tables full of 0.00, near-identical rows, scans in rare scripts. A parser gets a “Boo” if it repeats past what’s on the page, returns nothing, or is still going at the 10-minute timeout.

I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data. Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less? Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.

sam: local inference is a terrible allocation of scarce resources “a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment somehow many people's intuition is the opposite of that”

Sakana AI won 1st place in the company category at the “Recursive Self Improvement in Mathematics” Hack (OpenMath 2026), held with @stephen_wolfram at Harvard Innovation Labs. Competition: https://t.co/GIe4OW7GYq Results: https://t.co/uTHtSIQVtH Drawing mainly on our Sakana Fugu models and internal agentic tools, our team (Rachel Teo, Abderrahmene Hamdi, Wenyi Wang) contributed 14 mathematical families, including 6 complete resolutions, plus new bounds and structural advances. Thanks to the organizers and everyone involved. We plan to share more about our approach and learnings in a future report.

I do not think robots will replace people. For decades they will work together: few jobs a robot can do alone, and people are nearby. Knowing how to cooperate matters. HALO (ICML 2026 Spotlight): a humanoid stays coordinated with an unpredictable human. https://t.co/C4cw8Zgmca https://t.co/40mLExoj0Y

@GuillaumeGonano Fun setup... I went fully local for the voice part and talk to Claude Code over ACP instead of claude -p. https://t.co/h0kfQWJEMA

# Frontier is rebooted in 2026 I don't think I've announced this here, but I was able to teach Claude Code how to port Frontier from a backup taken in 2014. It's an Electron app, but also can run headless. It's incredibly reliable for something that was coded almost from scratch so recently. The one big piece we already had was the outliner, but everything else is new code in JavaScript. Also the object database is imp in SQLite, so there will be substantial things it can do in SQL, and Frontier never had SQL integrated like this. (Note, the odb works exactly as it used to work, and we have guest databases.) We experimented with new tables of glue scripts, so we are able to program Mastodon and Bluesky, as easily as we used to program WordPress or Twitter. I went on to work on other projects, but I use Frontier in everything I do, it's where I write and maintain my JS code for the other projects, so while I haven't done any development on it in the last few months I've been using it constantly every day. I would love to get a new community going as long as it isn't hostile, but I think we're better at controlling the trolls in 2026 than we were in the early days of the web.

happy @githubuniverse ! check out all the cool things VS Code is shipping in our 1.144 release notes https://t.co/FM6F0R0BlW my favorite highlight? image generation via GPT Image 2.5 Flare is now available in @code, @github app, and the CLI 🧑‍🎨🖼️ had a lot of fun working on this so please give it a try :D

L6N5401-0.75 seems to be the 'DJ neuron' - it puts everything into the context of a DJ set https://t.co/hTZjOnBhBe

great time with @matthewridenour and @kaggle today at build summit evals are getting way more complicated! And with models being used in so many domains, it was great to see the diversity of creative approaches used to generate good eval datasets :) https://t.co/B1i7XlsGOE

@Scobleizer said today that Specs couldn't augment robots...I couldn't let that stand. Shoutout @JohannesTscharn for open sourcing the foundation of this. https://t.co/zeNBqr4dlp https://t.co/XIvjg49fbp

What are some inference tasks that you’re still bottlenecked on intelligence (and not compute or $$) ?

Excited to share what we’ve been building: Carbon is here! Hands down the smartest model I’ve worked with, already my daily work partner, and casually helped me snag hard-to-get tickets🤣🥳 Enjoyed the process of building it to tackle open frontier problems. Also so proud of what our team did for Gemini's fast iteration. Let’s keep pushing! 🫡

🚀 Unlock frontier-level AI training performance with DeepSpeed! https://t.co/2fx0CqQIzw Join Masahiro Tanaka of DeepSpeed at PyTorch Day Japan 2026 as he presents “DeepSpeed at the Frontier: Community-Driven Optimizations for State-of-the-Art Training Performance,” exploring the latest innovations advancing the performance and scalability of distributed AI training. Register by November 11 and save ¥3,000. https://t.co/FmzLlNZAwh

Introducing Bivia. A multimodal AI decision model built to complete entire workflows, not just answer individual questions. We've spent 3+ years building and deploying AI in production across semiconductors, energy, manufacturing, and other complex industries. Today, we're bringing that experience to everyone.

The @github engineering team built a best-in-class, truly multi-model harness, which we've implemented across Microsoft products. Introducing Microsoft Copilot harness in Foundry Agent Service, a simpler way to build agents without rebuilding the execution layer. It handles orchestration, tools, context, and controls so you can focus on your agent’s skills and outcomes. Watch @jeffhollan's demo to see it in action.

Dyna-2 trained on more than a million hours of human video, and we were its largest data partner. We've delivered 2M hours since, 30+ teams are now under contract, and at 450,000 hours a month of capacity the next million takes about ten weeks. https://t.co/cSLAzElFOy

We gave robots eyes. Now we're giving them nerves. Introducing the world's most sensitive robot, this is unreal: