Markus J. Buehler

@vllm_project @nvidia 👏

Excited to see vLLM as the rollout engine in Molt, the new agentic-first RL framework from the @NVIDIA NeMo team. 🎉 vLLM (over Ray) carries the rollout here: fast async serving up to 1T-class MoE scale, simple to drop in. That lets the RL core above stay small and hackable. Can't wait to see what people train with it. 🚀

The pain of becoming is the cost of turning intelligence into science. An AI system becomes scientific by exposing its hypotheses to consequences, discovering when they fail, revising the representations that produced them, and preserving what it learns as executable knowledge. This article captures key insights through my talks and discussions at the Humanity & AGI Summit 2026 at @Stanford and the @NSF 2026 Summer School on carbon nanomaterials over the past few days.

I can’t say enough good things about John Carmack @ID_AA_Carmack and his Keen Technologies. But now Khurram Javed @kjaved_ and I have broken away to start our own startup and pursue a slightly different path toward understanding intelligence. Like Keen (and like Ineffable) we at Oak Lab @oaklab_ai believe in reinforcement learning and that intelligence is created and maintained from run-time experience. But we think current deep learning methods are weak and inefficient, and need not more tweaks, but fundamentally new ideas and a thorough reworking before they can provide a solid foundation for achieving the more ambitious goals of AI.

@omarsar0 💯

I don't think Anthropic realizes how disruptive these changes are to users. I appreciate the extension, but please stop playing games. Either keep it under the subscriptions or put it under the API already.

@SebastienBubeck 👏

from Sparks to The Real Thing in 3 years, pretty good ;-)

@lhyTHU Thank you!

@ProfBuehlerMIT super cool work!

J-Space introduced by Anthropic is pretty cool. Our paper in ACL 2019 (aclanthology.org/P19-1144.pdf), however, showed that the hidden states of intermediate layers in language models contain the information of not only the next token, but the following sub-sequence based on semantic dependency.

quick reminder to join us at @MIT tonight for an awesome talk by @jamesproney from @sokrypton's lab!

"All things are made of atoms”: @ProfFeynman called it a foundational seed of truth from which, if all of science were lost, the world could be rebuilt. This is the generative grammar of the universe, compressed into singular form - compression driven to its limit, the point where all redundancy is gone, measurement has nothing left to resolve, and what remains is the incompressible residue. This grammar defines the shape of AI that generates knowledge from first principles by building, breaking, and validating in the physical world, from molecules to machines. I’ll join the Humanity & AGI Summit 2026 at @Stanford on July 12 to discuss this and other emerging topics, including how AI discovers, validates, and creates, together with @lx2015, @james_y_zou, and @SnyderShot. Looking forward to the conversation!

@NeriOxman Nice work @NeriOxman!

Introducing our new work, Vigils. oxman.com/projects/vigils For a century, the color of what we wear has come from petrochemicals. Synthetic dyes shade nearly every fiber that touches our skin. Even natural fibers like cotton and silk, marketed as clean, reach our bodies by way of chemical-heavy dye baths and industrial fixatives, leaving behind effluent that lingers in ecosystems long after the garment is discarded. Vigils is a series of capes whose color is not applied but grown. We engineer bacteria to produce indigo and melanin pigments, then let them work directly on the surface of a 3D-knitted silk textile. When the process is complete, the cells are washed away, leaving only the pigment they have grown into the fiber. Vigils questions how a textile comes into being. In place of the industrial dye pipeline, it introduces a manufacturing process that unites synthetic biology with computational fabrication: engineered microorganisms working in concert with CNC knitting machines, in a workflow where biology and robotics operate together at the resolution of a single thread. We believe the future of fashion is not more sustainable versions of the same industrial paradigm. It is a fundamentally different relationship with the living world; one in which growth replaces application, biology replaces petrochemistry, and every textile becomes a singular expression of life. Vigils is our first step, and its logic extends far beyond color. More to come. Time to grow ..

@NotebookLM Check this example @NotebookLM ⤵️

Short Video Overviews are officially 💯 rolled out on mobile and web across all users in English! As always, your opinion means everything to us. Share your favorite creations below and tell us what features we need to add next! ❤️

@GabrielAsher02 @jm_alexia Interesting connection

@ProfBuehlerMIT This reminds me a lot of @jm_alexia 's brilliant TLM paper arxiv.org/pdf/2510.04871 which one could argue is a formalization/application of this finding. Functionally they behave similarly though Anthropic's research does not involve architecture modification.

Fascinating result…intelligence may feature selective broadcasting of intermediate representations into a shared workspace where they can be recombined, critiqued, and exposed. This is an interesting perspective especially for science and discovery.

@AnthropicAI Fascinating result! Intelligence may feature selective broadcasting of intermediate representations into a shared workspace where they can be recombined, critiqued, and exposed. This is an interesting perspective especially for discovery.

New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.

@OpenAIDevs Boston/Cambridge in the plans @OpenAIDevs ?

DevDay Exchanges are also coming to Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo, and Mexico City. Stay tuned for when the application opens.

Applications to attend OpenAI DevDay 2026 are open. Join us in San Francisco on September 29 to: • Tinker with what’s new • Swap build notes with other builders • Go deep in technical sessions • Bring your sharpest questions Apply by July 10: devday.openai.com

@DeryaTR_ @romainhuet @OpenAI 👀

Something really cool with @romainhuet at @OpenAI dropping soon!

@nvidia 👏

Open models are becoming the foundation for modern AI research. At #ICML2026, 145 accepted papers cite NVIDIA Nemotron models and datasets, 74 NVIDIA papers were accepted, and ~2,000 accepted papers cite NVIDIA GPUs.

‼️

@thomasahle 😂

new model currently being pretrained by sama's spinoff lab 👀