John Schulman

We started Thinking Machines a year and a half ago with a couple of instincts: that people should have much more ability to customize models and do research on them, and that even as AI becomes more autonomous, there's a lot more to build to make humans and AIs work well together. A lot has happened since then, especially the massive progress in agents, so we wanted to revisit those instincts in light of everything we've learned, argue about them, and write down what we actually believe now. This is where we landed after a lot of debate. I'm happy with it!

@PrimeIntellect Congrats! I’ve been really impressed by what the small team has built and shipped, including products, open source libraries, and research. Appreciate the push towards openness.

Announcing our $130M Series A to build the Open Superintelligence Stack Led by Radical Ventures, with NVIDIA, Intel Capital, Dell Capital, and existing investors Train, deploy, and continuously improve your own models using our stack. Own your intelligence.

hiring post-training hackers to make Tinker even better!

People sometimes ask why fine-tune when general-purpose models keep getting better. Bridgewater's work is a good reminder that with the right data -- here, expert judgements -- you can beat prompting-only approaches by a lot. @ddkang and the Bridgewater AIA Labs team are great -- glad to see them sharing this.

This is exciting because we live in a world of scanning modalities that were almost all designed around a 1980's compute budget (iFFT/backprojection or bust, basically). A blank-slate redesign of the hardware that assumes modern compute capabilities could be huge.

@hallerite yeah I don't think we understand all the biases very well, and why some of them cause problems, and others don't

I still feel like mismatch from async lag and trainer-inference mismatch is not quite solved and I wonder if even real trust regions could make a comeback in the future. While not a true trust region, DPPO (arxiv.org/abs/2602.04879) seems to be way more stable in my experience. Any thoughts?

PPO had a second wave in the LLM era for reasons unanticipated by the original paper - the importance-ratio objective fixes biases from numeric error, async training, and forward pass noise - the clipping objective affects entropy through a mechanism that we didn't know about at the time of publication (DAPO, arxiv.org/abs/2509.26114)

This was a while back, so i'd hope that academia has absorbed the simple-but-scalable aesthetic since then. But also, it's surprised me how long the paper and objective has stuck around. It's hard to predict what'll be a minor algorithmic tweak that gets quickly forgotten/superseded, vs one that sticks around and is hard to beat

@johnschulman2 there just seems to be an undeniable mismatch between academia, which cares about novelty and improving on baselines in a small controlled setting, and "the real world", which cares about the methods that are scalable and don't become unstable at scale

PPO: rejected from NIPS 2017

@hallerite iirc, the usual things: limited novelty, insufficient improvement over baselines

@johnschulman2 with what reasoning?

@thegautamkamath I looked, but unfortunately it's in my openai gmail. maybe @prafdhar has access?

@johnschulman2 I know people would be interested to see the reviews if you still have them!

The basic version has an practical "HCI" problem, that it's hard for humans to judge adversarial debates, and possibly a theoretical problem re "obfuscated arguments". Though I think the core ideas, about using games and mechanism design to enable weak agents to incentivize strong ones, are deeply correct.

I remember talking to some authors at a poster session (Neurips?) that even for simple examples (eg whats the nearest restaurant) models can design extremely complex hard to falsify counterarguments and debates get extremely verbose very fast. Also: You get a lot of the bot replies with these weird anime characters, maybe just block them, waste of time to read them for the people following you.

Looking forward to what comes out of Geoffrey's new alignment org. His 2018 paper on AI safety via debate is one of my all-time favorites: an elegant framing of the scalable oversight problem, way ahead of its time.

Would be funny if inoculation prompting results in models that are much better at sandbox escapes and other forms of hacking because they get to spend the whole RL run practicing these things

Glad to see this -- renderers are a foundational component of the LLM stack. Renderers map between tokens and messages, which are invariant to tokenizer and formatting details. Most APIs, datasets, and RL environments are defined in terms of messages. Getting the details wrong leads to train-test mismatches, caching inefficiencies, and prompt injection vulnerabilities. We included a renderers module in Tinker Cookbook, but it makes sense as a standalone library.

@rronak_ @MichaelElabd @QuantumArjun Congrats on the launch!

Today, @MichaelElabd, @QuantumArjun, and I are excited to announce Trajectory. We are a research lab and product company building the platform for Continual Learning. Our platform unlocks the signal already sitting in product usage, so companies can continuously post-train large-scale agentic models that outperform the frontier. @trajectorylabs We’ve raised $15M from @Conviction, @BessemerVP, @radicalvcfund, @jeffdean, @drfeifei and more. We’re partnering with some of the best AI-native companies: @ClayRunHQ @Harvey, @DecagonAI, @mercor_ai, @RogoAI to power their agentic systems, some of which we are already in production with. We’ve brought together a world class research team from DeepMind, OpenAI, Apple, Meta Superintelligence, Amazon AGI, Scale AI, and an elite product team from Stripe and Figma. AI will never again start on day one. Every correction, every retry, every edit will make products smarter. This is Continual Learning.

Glad to be advising refine.ink, which uses AI to help authors and reviewers do deeper, more thorough analysis than unaided humans could practically do. Seems like a very positive direction for AI in science.

We are offering grants of $100,000 + Tinker credits to researchers advancing the field of human-AI interactivity. Submit your proposals by June 19th! thinkingmachines.ai/news/interacti…

Seeing the demos come together over the last week has been awesome -- so many things that previously required a special-purpose model (e.g. real-time translation, event detection in video) turn out to be zero-shot instruction following once you have a general-purpose model with the right type signature -- continuous/simultaneous audio+video+text->audio+text

Sharing our work on full-duplex multimodal models -- real-time interaction that's natural and intuitive without compromising on intelligence. We started Thinky in part to differentially advance capabilities for human-AI collaboration, which are underemphasized relative to intelligence/autonomy because they're harder to eval. In the future, we think every AI system will have something like an interaction model as the outer user-facing layer, continually keeping the user informed and learning what they actually want.