ChatGPT Co-Creator Liam Fedus: After OpenAI's 722 Manuscripts, AI Should Move Into the Lab

OpenAI has posted 722 math manuscripts on GitHub in one go, covering 17 fields, from number theory and geometry to theoretical computer science and mathematical physics. Some of the papers come with formal proofs in Lean, which a computer checks mechanically. According to the company, an internal frontier model wrote all of the texts, and that model has not been released yet. A few days later, OpenAI withdrew several of the manuscripts. Against this backdrop, Liam Fedus, one of the creators of ChatGPT, makes his case: after mathematics, AI needs to move into the lab.

What exactly was published

The scale is unusual. A typical mathematician writes dozens of papers over a career, while here hundreds of texts across different branches of science appeared at once. What unites them is their author, which is not a human but an unreleased model.

Here is what is known from the announcement:

• 722 manuscripts were posted as a single package on GitHub;
• they cover 17 mathematical and adjacent fields;
• some of the papers have formal Lean proofs, which are checked without human involvement;
• the source is an internal OpenAI model with no public release;
• several manuscripts were later removed from the repository.

The excerpt of the material available to us does not give the reasons for the withdrawals, so it is too early to draw conclusions about the quality of the withdrawn papers. The fact of the withdrawal itself points to something else: at this volume, selection and verification become a separate task.

Why mathematics went first

Mathematics suits AI for one reason: answers are cheap to check. A proof either passes in Lean or it does not. For reinforcement learning, that is a nearly ideal signal, because the model gets an honest score with no experts and none of the costs of the real world. That is why systems quickly reached results here that previously required professional-level expertise.

This convenient setting has a limit. A manuscript without formal verification remains a text that a specialist has to read. Hundreds of papers at once mean hundreds of hours of peer review. Lean proofs for some of the papers solve the problem only for those results that could be formalized.

The lab as the next proving ground

Fedus's thesis makes sense given his trajectory. According to public information, he left OpenAI and founded Periodic Labs, a startup that builds AI systems for physics, chemistry and materials science. The bet is that a model should get feedback from real experiments, not only from text and formal verifiers.

In the lab, verification is expensive. Synthesizing a material takes hours or days, measurement requires equipment, and noise in the data does not go away just because the model is smart. But the result cannot be disputed: a sample either has the desired property or it does not. If mathematics showed that AI can produce proofs in a steady stream, the lab tests whether it can produce verifiable discoveries in a steady stream.

The excerpt of the material has no details of the interview itself, so we do not retell Fedus's specific arguments. The headline, however, sets the direction: the bottleneck is shifting from generating ideas to the speed and cost of verifying them.

What to watch next

Three markers are useful for anyone following the industry.

1. The fate of the manuscripts. How many papers independent mathematicians confirm, and how many more get withdrawn. This is the best measure of the internal model's reliability. 2. The model's release. If OpenAI releases it, it will become clear how well the results reproduce outside OpenAI's internal infrastructure. 3. Automated labs. Repeated "hypothesis, experiment, model update" cycles in materials science will show whether success carries over from a formal environment to a physical one.

For teams building products on LLMs, the takeaway is practical: a model's value grows where a fast, cheap verifier exists. That can be Lean, tests in code, a simulator, or a robot in a lab. Build the pipeline starting from verification, not from generation.