William "Liam" Fedus
Last reviewed
Jun 7, 2026
Sources
10 citations
Review status
Source-backed
Revision
v1 · 1,402 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
10 citations
Review status
Source-backed
Revision
v1 · 1,402 words
Add missing citations, update stale details, or suggest a clearer explanation.
William Fedus, who goes professionally by Liam Fedus, is a machine learning researcher and entrepreneur known for his work on large-scale neural networks and for helping build some of the most widely used AI systems of the early 2020s. At Google Brain he was the lead author of the Switch Transformer, a 2021 paper that pushed sparse mixture of experts language models past a trillion parameters. He later joined OpenAI, where he was one of the researchers who built ChatGPT and rose to vice president of research for post-training before leaving in 2025 to co-found Periodic Labs, a startup that pairs AI models with automated laboratories to accelerate discovery in the physical sciences. [1][2][3]
Fedus studied physics as an undergraduate at the Massachusetts Institute of Technology, then earned a Master of Science in physics at the University of California, San Diego, where he was co-advised by David Meyer and Gary Cottrell. He completed a PhD in computer science at the Universite de Montreal and the affiliated Mila institute, co-advised by Yoshua Bengio and Hugo Larochelle. His physics background, which he has repeatedly cited as the motivation for his later turn toward AI for science, threads through his entire career. [9][1]
During and after his doctorate, Fedus worked at Google Brain, first as a student researcher and later as a research scientist. His early publications spanned generative modeling and reinforcement learning. In 2018 he co-wrote MaskGAN, a method for text generation built on a generative adversarial network, with Ian Goodfellow and Andrew M. Dai; in 2020 he co-authored "Revisiting Fundamentals of Experience Replay," a study of replay buffers in deep reinforcement learning. [7]
His best known contribution came in January 2021, when he, Barret Zoph and Noam Shazeer published "Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity." The paper simplified the mixture-of-experts routing scheme by sending each token to a single expert rather than several, which cut communication and computation costs while preserving model quality. The largest Switch Transformer reached 1.6 trillion parameters and trained up to seven times faster than a comparable dense model, demonstrating that sparsely activated transformer networks could scale efficiently. Published in the Journal of Machine Learning Research, the work became one of the most cited references in the scaling laws and sparse-model literature. Fedus and Zoph followed it in 2022 with ST-MoE, which tackled the training instabilities that had limited earlier sparse expert models and produced models that transferred well to downstream tasks. [5][6][7]
Fedus also contributed to large collaborative efforts at Google. His publication record lists him among the authors of the 2022 paper "Emergent Abilities of Large Language Models," which documented capabilities that appear only once models cross certain scale thresholds, and he is credited as a contributor on work surrounding Google's PaLM effort. By the time he left Google, his research had accumulated tens of thousands of citations. [7]
Fedus left Google and joined OpenAI in 2022. He was part of the small team that built ChatGPT, which launched in November 2022 and became one of the fastest-growing consumer applications on record; several accounts describe him as a co-creator of the product. Within OpenAI's research organization he focused on post-training, the stage in which a base model is refined through techniques such as instruction tuning and reinforcement learning from human feedback so that it follows instructions and behaves reliably for users. He led post-training research and development for a string of flagship releases, including GPT-4o and the first o1 reasoning models, o1-preview and o1-mini. [8][1]
In October 2024, during a period of senior departures at the company, Fedus was named to lead OpenAI's post-training organization, the role previously held by Barret Zoph, his co-author on the Switch Transformer. He held the title vice president of research for post-training, placing him at the center of OpenAI's model development pipeline. [8]
On March 17, 2025, Fedus announced that he was leaving OpenAI as an employee to start his own company. In a note to colleagues that he also posted publicly, he wrote that he had "gotten really excited about AI for science" and that his "undergrad was in physics and I'm keen to apply this technology there." OpenAI characterized AI for science as one of its most strategically important areas and said it planned to invest in and partner with his new venture. He framed his exit as a move from employee to partner rather than a clean break. [1]
Fedus co-founded Periodic Labs in 2025 with Ekin Dogus Cubuk, a former colleague from his Google years. Cubuk earned his PhD at Harvard and completed a postdoctoral fellowship at Stanford before leading materials science and chemistry research at Google DeepMind, where his team produced GNoME, a graph network that predicted millions of new candidate crystal structures. The two founders argued that frontier AI models had largely exhausted the public internet as a training resource and were missing the real-world experimental data needed for genuine scientific discovery. [2][3][4]
Periodic Labs is building what it calls "AI scientists" together with the autonomous laboratories for them to operate. The approach couples large models that can form hypotheses, run simulations and plan syntheses with robotic powder-synthesis labs that mix and heat precursors, measure the outcomes and feed the results back into the models. In the founders' framing, nature itself becomes the reinforcement learning environment, closing the loop between hypothesis and physical reality and generating proprietary data that no web-scraped corpus contains. The company's early targets include new superconductors, magnets and heat shields, with a longer-term ambition of room-temperature superconductivity and applications across advanced manufacturing, semiconductors, energy and aerospace. [3][4]
The startup emerged from stealth on September 30, 2025, with a $300 million seed round led by Andreessen Horowitz. Felicis, whose partner Peter Deng committed the first check, also invested, alongside DST Global, Accel and Nvidia's venture arm NVentures, plus a roster of angels that included Jeff Bezos, Eric Schmidt, Jeff Dean and Elad Gil. Reports placed the round's valuation between roughly $1 billion and $1.3 billion, an unusually large seed that TechCrunch described as setting off a venture capital frenzy. The founding team drew other prominent researchers, among them Alexandre Passos, who had worked on OpenAI's o1 and o3 models, the materials scientist Eric Toberer, and Matt Horton, a developer of Microsoft's materials science AI tools. In March 2026, Bloomberg reported that Periodic Labs was in talks to raise additional funding at a valuation of about $7 billion, underscoring how quickly investor interest in AI for the physical sciences had grown. [2][3][4][10]
| Year | Work | Selected co-authors | Venue |
|---|---|---|---|
| 2018 | MaskGAN: Better Text Generation via Filling in the ____ | Ian Goodfellow, Andrew M. Dai | ICLR |
| 2020 | Revisiting Fundamentals of Experience Replay | P. Castro, M. Bellemare and others | ICML |
| 2021 | Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity | Barret Zoph, Noam Shazeer | JMLR (2022) |
| 2022 | ST-MoE: Designing Stable and Transferable Sparse Expert Models | Barret Zoph, Jeff Dean, Noam Shazeer | arXiv |
| 2022 | Emergent Abilities of Large Language Models | Jason Wei and others | TMLR |
| 2023 | GPT-4 Technical Report | OpenAI | arXiv |