ChatGPT Co‑Creator Liam Fedus: After Math, AI Should Move Into the Lab

ChatGPT Co‑creator Liam Fedus: After Mathematics, AI Should Head to the Lab

A couple of days ago, OpenAI released 722 math manuscripts at once (later with a few withdrawn), covering 17 fields including number theory, geometry, theoretical computer science, and mathematical physics, with some accompanied by Lean formal proofs that can be mechanically checked by computers. According to OpenAI, they all come from an as‑yet‑unreleased internal frontier model, and each result on average consumed computational power roughly equivalent to three hours of ChatGPT Pro’s thinking mode.

This is not OpenAI’s first major move in mathematics. On September 8, a month ago, it announced that it had internally scheduled tens of thousands of agents for about 88 hours, producing a breakthrough proof related to the Navier‑Stokes millennium problem, and immediately became embroiled in disputes over authorship and priority. After the release of 722 manuscripts, skepticism followed suit.

Beyond the controversy, there is a fact that is rarely mentioned: mathematics has become the fastest scientific field for AI to make inroads largely because it comes with its own judge—a Lean 证明—whether it’s correct or not, the compiler decides.

But once the problem moves from paper to the material world—whether a new material can achieve superconductivity at liquid‑nitrogen temperatures, or whether an alloy will crack at high temperature—no formal system can replace a real experiment.

This is exactly the point emphasized by Liam Fedus in the latest episode of the podcast The Frontier. This co‑creator of ChatGPT, former OpenAI vice president and head of post‑training, is now building a physical lab in Menlo Park, trying to provide AI scientists with the data that the internet can’t supply.

In a 50‑minute conversation with the host and Mercor founder Brendan Foody, he reviewed the internal details before and after ChatGPT’s birth, explained why ChatGPT and AI programming were the first breakthroughs, and clarified why building a lab is essential for Periodic Labs.

https://x.com/BrendanFoody/status/2108663727197462818

From Rural Maine to ChatGPT: A Physicist’s Detour

Fedus grew up in rural Maine. He recalls that school had an assignment called “Science in the News,” which required students to write about local science news, but the countryside newspaper had hardly any technological developments to cover, so he had to look online for material and ended up being deeply fascinated by a physics website. In college, he wavered between physics, chemistry, and economics, and ultimately chose physics because it was “fundamentally solid.”

While studying physics at MIT, he took part in a dark‑matter detection project called DMTPC. According to one theory, dark‑matter signals would exhibit directional changes over time, and the team hoped to use this to build a more sensitive detector.

It was also here that he first ran into machine learning problems: inferring a particle’s direction and energy from raw detector data is essentially a machine‑learning task. After a brief stint in finance, he pursued particle physics at UC San Diego, worked on CERN’s particle‑track reconstruction, and the same problem resurfaced. “I realized that to push the frontier of machine learning, I had to stay at the frontier institutions.”

The turning point came during the hype around generative adversarial networks (GANs). Fedus tried to use GANs for language generation and discovered a double dilemma: GANs are inherently unstable and prone to mode collapse, and the generator outputs discrete tokens, so the discriminator’s gradients cannot be back‑propagated directly. In other words, “you get all the problems of GANs and all the problems of reinforcement learning at once.”

He wrote an email about this to Ian Goodfellow, who was then a researcher at Google, and the reply convinced him to join that group.

In 2016, he joined the Google Accelerated Sciences team as an intern, using convolutional networks to study yeast cell replication and division. He later moved to Google Brain and completed his Ph.D. in computer science at Mila under Yoshua Bengio and Hugo Larochelle.

At Google Brain, Fedus collaborated with Barret Zoph and Noam Shazeer to propose the Switch Transformer, scaling the Mixture-of-Experts (MoE) architecture to the trillion-parameter level. This work later became a key reference for open-source MoE models.

He joined OpenAI in 2022, helping create ChatGPT, and in October 2024 succeeded Barret Zoph as head of the post‑training team. In March 2025 he left to start his own venture.

He founded Periodic Labs, with co‑founder Ekin Dogus Cubuk. The two have known each other for nearly a decade at Google Brain; Cubuk later led the chemistry and materials team at DeepMind and is one of the authors of the GNoME paper, which in 2023 predicted a large number of new stable crystal structures. In 2025, Periodic Labs emerged from stealth with a $300 million seed round, backed by investors including a16z, DST, NVIDIA’s NVentures, Accel, as well as individual investors such as Jeff Bezos, Eric Schmidt, Jeff Dean, and others.

In mid‑September this year, Periodic released its first model Periodic Neon, a roughly 1‑trillion‑parameter model fine‑tuned on the company’s own laboratory data for interpreting X‑ray diffraction spectra; the company claims it outperforms GPT‑6 Astra on its internal benchmark, though this result is currently self‑reported.

ChatGPT Almost Became a Meeting Bot

On Fedus’s first day at OpenAI, Greg Brockman showed the newcomers the freshly trained GPT‑4. “It’s about a year and a half ahead of any system I’ve seen.” But he also admitted that GPT‑4 still makes the usual mistakes of language models: it hallucinates and can’t solve math problems. It was clearly a huge leap, yet no one knew how to turn it into something valuable for the world.

With about six months left before the planned release of GPT‑4, the team discussed many product forms, including writing bots, coding bots, and a few members who were especially fond of a meeting bot that would sit inside conference software to take notes and generate to‑do items. The host joked that if they had actually built a meeting bot back then, the world would be very different. In the end, John Schulman called the shots, insisting they should keep it general and make a chat bot.

Interestingly, the early internal signals were not optimistic. Fedus estimated that among friends and family who got early access to GPT‑4, about 90% didn’t really care. As a researcher whose main focus is architecture research, he took this as a negative sign at the time. “But in hindsight, there’s actually a very positive signal hidden in there: 5% to 10% of people love it and use it every day. Every time we push an update, they complain, ‘I liked the previous version better.’”

Even more surprising is that the globally explosive ChatGPT didn’t even use GPT‑4, but rather GPT‑3.5, which had been available as an API for several months. The real variable was the product form:the shift from a “completion” interface to a “conversation” interface.

Fedus gave an example: in completion mode, if you want the model to write a poem in a certain style, you have to first fabricate a “recently discovered lost poem by a famous author, starting as follows:”, whereas in chat mode you can simply say “please help me write a poem in this style.” “It’s a small thing, but it feels very natural.” Apart from that, it’s about doing a lot of tedious detailed work around supervised fine‑tuning and RLHF – “everything ultimately comes back to the data, we keep staring at the data.”

When an internal poll tried to predict ChatGPT’s user count, people guessed 10 k, 50 k, and the boldest guessed 1 million. “We originally just wanted to do a low‑key research preview, publish a blog, maybe 10 k researchers would try it.” The result was chaos, “all the dashboards were red.” In the summer of 2023 usage dropped, the team fell into a “survival crisis,” fearing the hype around ChatGPT would end, only to later discover that students were on summer break.

When you’re too close to technology, you only see flaws

In the later stage of OpenAI, Fedus leads the post‑training team, and one of the projects he’s most proud of isbringing reasoning models into chat scenarios.

Reasoning models excel at math and programming, but making them suitable for everyday conversation isn’t easy. “If you just ask ‘How are you?’, and the model spends three minutes thinking, that’s obviously a poor user experience.” The team must manage the dynamic boundaries of thinking and design different reasoning strategies for different domains—the reasoning approach for solving a math problem is completely different from that for drafting an article.

Another aspect of the job that satisfies him isthe data flywheel: how to use user feedback to continuously improve the system. He strongly embraces OpenAI’s consistent iterative deployment philosophy—engage with reality as quickly as possible, close the loop quickly, “let the world gradually adapt to what’s coming while figuring out where the system will fail.” In his view, establishing predictable scaling patterns based on user feedback is a major achievement for the team.

When asked about his first impression of GPT-4, Fedus said something with a distinctly researcher’s tone: “When you’re too close to the technology, all you see are its flaws.” Researchers don’t celebrate a model that gets 80% of questions right; they focus on evaluations, asking why it fails on a particular math problem or why it can’t follow instructions to summarize. Early on, everyone kept some private “feel‑test” questions; he had a quantum mechanics problem that all previous models performed poorly on until a reasoning model appeared and got it right for the first time. “I was truly blown away; it displayed very interesting traits, like self‑correction.”

Why Chat and Programming Broke Through First

The host raised a key question: In the past few years there have been two real breakthroughs, one is ChatGPT, and the other is AI programming.

Why specifically these two fields? What is missing for the physical world to see a similar breakthrough?

Fedus answered that ChatGPT relies on a sufficiently strong pre‑training foundation and RLHF.

He believes that ChatGPT cannot succeed with GPT‑3‑level technology; GPT‑3.5 compresses the distribution of internet data much more effectively, providing a better foundation for subsequent fine‑tuning and reinforcement learning. RLHF distills human preferences into the neural network, offering a good optimization signal. However, it has a fundamental limitation: the reward model trained by “do you prefer answer A or B” is not sensitive to correctness, whether mathematical correctness, code correctness, or factual accuracy. “What they actually encode is style, intent, and form.” ChatGPT doesn’t need to be perfect at everything; it only needs to cross the product‑market fit threshold.

Programming is different. Modern programming agents must spend a lot of compute on long‑chain reasoning, interact with the environment, and continuously adjust based on failures and new evidence, which requires strong verifiability and error‑correction capabilities. During testing, computation plus verifiable rewards is the real key to unlocking programming ability. The host added that programming also benefits from the massive human knowledge base on GitHub, which naturally contains abundant unit tests, whereas most scientific and knowledge‑work domains lack an equivalent.

Fedus uses AlphaFold as an example: it succeeded because it is backed by decades of accumulated protein structure data. “Machine learning systems are not magic; whatever data you train them on and whatever task you give them, that’s what they’ll excel at.” Materials science, on the contrary, suffers from a severe lack of good data. Online experimental data are very noisy, and two labs performing the same experiment may reach different conclusions; papers often omit crucial details, and a simple sentence like “we synthesized this material” hides millions of degrees of freedom; even worse, there are almost no negative results. “Everything you see online succeeded, which clearly reflects a highly biased data distribution.”

“To build AI scientists, you must actually do science.”

This is precisely the fundamental reason why Periodic Labs wants to build a lab.Fedus envisions a scenario: even if you have a superintelligent system that can overview the frontiers of all human knowledge, it judges which knowledge is coherent, discards certain claims, runs some simulations, and derives several hypotheses about the world, “but if you lack the ability to put ideas into practice, you remain only in thought.”To collapse multiple possibilities into a single truth, you must conduct experiments.

Periodic’s approach is to combine three technical pathways:the reasoning ability of language models, computational physics simulations, and high‑throughput laboratories.

Fedus recalled that when he discussed GNoME with Cubuk, he found that the system used active learning but did not employ language models at all. The two instantly clicked, believing that these components should be combined. They chose to start with materials discovery because materials are the foundation of all advanced industries such as semiconductors, aerospace, and energy; the iteration cycle is relatively fast, the signal‑to‑noise ratio is acceptable, and mature simulators already exist for many materials.

“If you want to create an AI scientist, you have to actually do science.” Fedus said, AI systems can learn from the internet, run physical simulations, and thereby construct a hypothesis space about which atomic configurations are stable and what properties they might have, but ultimately they must be validated in a Menlo Park lab, obtain negative results, calibrate the feedback data, and cross‑verify with simulations and published literature. “We believe there is no room for negotiation on this.”

The data collection methods also went through some trial and error. What they cared most about was the full context: what the hypothesis was, how the experiment was performed, and what the results were. A physics simulation dataset from twenty years ago is almost useless if handed to you in isolation, because you don’t know its intent or whether the result is positive or negative. Early on the team even tried having scientists wear GoPros while walking around the lab, “but they didn’t like it,” and later switched to simpler voice recordings and uploading lab notebooks.

An unexpected finding is that as soon as a small‑scale lab is set up, the team quickly hits an intelligence bottleneck. “Scientists and engineers tell us: we can’t possibly understand all this data without help.” In Fedus’s view, AI makes a research paradigm that is closer to the scale of machine learning possible, and the data collection itself must also be smart enough. “If your data collection is clumsy and incomplete, the next‑generation AI will simply ignore that part.”

When asked how many labs there will be in the future, his answer was “easily dozens.” Current labs are still unable to conduct serious alloy or polymer research, and different physical domains require different experimental conditions, but how to orchestrate labs, how to identify bad data, and how to deploy intelligence across machines are experiences that can be transferred.

Superconductivity, Nuclear Fusion, and Moore’s Law

Periodic’s long‑term goal, in Fedus’s words, is “to intentionally understand and design matter,” and new superconductors are one of the most important targets, which will benefit medical devices, maglev trains, and even loss‑less power transmission.

He specifically clarified that whether room‑temperature superconductivity is physically possible is still undecided, “We are looking for it, and I think this is one of humanity’s best chances ever on this issue.” However, it doesn’t have to reach room temperature to be valuable: as long as the critical temperature is significantly above the liquid‑nitrogen range, combined with good ductility, easy manufacturability, and lower reliance on rare earths, it will have a huge impact on nuclear‑fusion reactors, because they need superconductors to create powerful magnets.

When askedhow far commercial fusion is, he said based on the evidence he’s seen, “it seems very likely within the next seven years or so.” He added, “Capital and technology are intertwined; the speed of technological progress is not a fixed constant—it largely depends on how many resources society invests.”

If energy becomes cheap enough, the nature of many problems will change. Fedus uses climate change as an example: from a purely engineering perspective, excess CO₂ in the atmosphere is essentially an energy problem. When energy is extremely cheap, past public externalities can be addressed within the economic system.

However, the most exciting impact for him iscomputation itself. “Computation is essentially a physical process.” As society pours more and more capital into data centers, if we can eliminate the material bottleneck for the next five years and keep Moore’s Law going, AI progress can continue to advance. He also mentioned that Musk often talks about material challenges such as rocket heat shields. Looking around, “we are constrained everywhere by existing materials.”

Data, compute, and the future of work

Regarding the common external question “should money be invested in data or compute,” Fedus returned to the Scaling Law: when other variables are unconstrained, loss predictably decreases as compute, data, and parameter count increase. The key is which regime you are in—whether you are compute‑constrained or data‑constrained. If certain domains that are valuable for downstream tasks are severely under‑represented in pre‑training corpora and existing reinforcement‑learning task distributions, then investing in data yields a much larger benefit. He believes every company should strive to make performance gains from data as predictable as those from compute.

The host naturally asked the expert about the future of data. Fedus thinks demand will only keep growing because many critical, economically valuable data remain severely under‑represented, and “work across industries is constantly evolving, so taking a snapshot of a profession in 2026 doesn’t solve the problem once and for all.” He predicts that, at least for the next decade, industries will heavily rely on this type of data.

When discussing AI’s impact on employment, Fedus cited Geoffrey Hinton’s old prediction that radiologists would be dramatically reduced, yet in reality radiologists are still being hired in large numbers because the job involves far more than just looking at images.

He used software engineers as an analogy: if you compare a 2020 engineer’s time allocation chart to today’s models, the model might already be able to accomplish 80% to 90% of that work, and by that standard “we have achieved AGI.” However, the nature of the work is constantly evolving, and engineers now spend more time on cross‑functional collaboration and ensuring designs meet expectations.

He predicts that as intelligence becomes increasingly commoditized, “the things between people” will become ever more important.

As for what the world will be like in ten years, this researcher who lived through the ChatGPT explosion refuses to give specific predictions. “Ten years is a very long period. Looking back to 2016, the Transformer hadn’t even been invented yet.” He only says he is generally optimistic about the future and agrees with Dario Amodei’s notion of a “genius nation in the data center.”

Conclusion

Viewing Fedus’s interview against the backdrop of 722 mathematical manuscripts yields an interesting contrast. Mathematics has Lean, peer review, and proofs that can be checked line‑by‑line by a compiler; yet even so, the outside world still demands “reproduce first, then acknowledge” for OpenAI’s results. In the realms of materials, chemistry, and physics, there isn’t even such a ready‑made arbiter—the only thing that can deliver a final verdict is a real experiment.

In a research blog, Periodic summed up this idea as “Nature is our learning environment.” This company is betting that, after internet corpora become nearly exhausted, it can tap a continuously flowing data well for AI. Whether it can deliver will depend on whether the lab can turn massive capital into reproducible physical discoveries. But one thing is clear, as Fedus put it:On the road to becoming an AI scientist, thinking alone is not enough.

© THE END 

For reprints, please contact this public account for permission

Submit articles or request coverage: liyazhou@jiqizhixin.com