William "Liam" Fedus

ref · May 28, 2026, 8:52am

Physicist-turned-AI researcher who co-created ChatGPT at OpenAI, introduced the Switch Transformer, and co-founded Periodic Labs to build AI scientists for physical discovery.

Profile


Nationality	American
Current Institution(s)	Periodic Labs (Co-founder)
Research Areas	Large Language Models, Post-Training, Mixture of Experts, Reinforcement Learning, AI for Science
Doctoral Advisor	Yoshua Bengio, Hugo Larochelle
Doctoral Thesis	(title unconfirmed) (Université de Montréal, c. 2021)
Website	acsweb.ucsd.edu/~wfedus
X / Twitter	@LiamFedus
GitHub	liamb315
Google Scholar	William Fedus — 51,900+ citations

Overview

William Fedus — universally known as Liam — is an American AI researcher whose career traces a straight line from theoretical physics to some of the most consequential model development of the current era. He is best known for co-authoring the Switch Transformer (2021), which demonstrated that sparse Mixture-of-Experts (MoE) routing could scale language models to trillion-parameter regimes at constant computational cost, a result that has since shaped the architecture of every major frontier model family. At OpenAI, where he joined in 2022, he was part of the small founding team that built ChatGPT and ultimately rose to Vice President of Post-Training Research before leaving in March 2025 to co-found Periodic Labs. The startup raised a $300 million seed round led by Andreessen Horowitz and is building autonomous physical laboratories pairing AI-generated hypotheses with robotic experiments, targeting materials discovery — most ambitiously, room-temperature superconductors — as its first domain.

Early Life & Education

Fedus is a physicist by training. He completed a B.S. in Physics at MIT, where he contributed to the DMTPC (Dark Matter Time Projection Chamber) collaboration, a directional dark matter detector experiment. He then moved to UC San Diego for an M.S. in Physics, co-advised by David Meyer and Gary Cottrell, where his interests drifted toward the computational end of the discipline. In 2017 he enrolled in a PhD in Computer Science at the Université de Montréal, joining the Mila ecosystem and working under the co-supervision of Yoshua Bengio and Hugo Larochelle. Throughout his doctorate he held concurrent positions as a Student Research Scholar at Google Brain, completing internships there in 2016 and 2017 before transitioning to a full research role.

Career

Google Brain (2016–2022)

Fedus began his Google Brain relationship as an intern in Summer/Fall 2016 working on Generative Adversarial Networks, returning for a second internship in Winter/Spring 2017. After starting his PhD, he continued as a Student Research Scholar embedded at Google Brain from 2018 onwards — a common arrangement that allowed him to pursue academic research while contributing to Google-scale projects. His most significant output from this period is the Switch Transformer (2021), co-authored with Barret Zoph and Noam Shazeer, which showed that replacing the standard dense feed-forward layer in a Transformer with a learned routing mechanism selecting a single expert per token — the “switch” — achieved up to 7× pre-training speedups over T5-equivalent baselines with the same compute budget. The paper was published in the Journal of Machine Learning Research and has accumulated over 4,000 citations. The architecture laid the MoE groundwork subsequently adopted in models such as Mixtral, GPT-4, and Gemini.

OpenAI (2022–March 2025)

Fedus joined OpenAI from Google in 2022. He was part of the small team that built and launched ChatGPT in November 2022 — described by multiple sources as one of the most consequential product launches in AI history. He subsequently ran OpenAI’s post-training research function, which encompasses the reinforcement learning from human feedback (RLHF), instruction-tuning, and alignment fine-tuning work that converts a pre-trained base model into a deployed assistant. His team’s work contributed to GPT-4, GPT-4o, and the Operator (now Agent) product. In Fall 2024 he was formally promoted to Vice President of Post-Training Research, a recognition of his team’s centrality to OpenAI’s product pipeline. He announced his departure to colleagues in March 2025, framing it as an amicable transition: OpenAI initially signaled plans to invest in and partner with his new venture, though that investment did not ultimately materialise.

Periodic Labs (March 2025–present)

After leaving OpenAI, Fedus co-founded Periodic Labs with Ekin Doğuş Çubuk, former head of materials science and chemistry research at Google DeepMind. The company came out of stealth in September 2025 with a $300 million seed round — one of the largest in history — led by a16z, with Felicis writing the first cheque. Additional backers include DST, NVentures (NVIDIA’s venture arm), Accel, and angel investors including Jeff Bezos, Elad Gil, Eric Schmidt, and Jeff Dean. Periodic’s thesis is that autonomous physical laboratories, in which AI models propose experiments and robotic systems execute them, can generate the high-quality out-of-distribution training data that internet-scraped corpora cannot provide. The company has assembled a founding team of roughly 26 researchers spanning AI, materials science, and experimental physics, including Alexandre Passos (a creator of o1 and o3), Dzmitry Bahdanau (a creator of the neural attention mechanism), and Rishabh Agarwal. Its initial scientific target is the discovery of high-temperature superconductors; it is simultaneously working with an unnamed semiconductor manufacturer on heat-dissipation challenges.

Key Contributions

Switch Transformer (arXiv 2101.03961, JMLR 2022; with Zoph and Shazeer) — Demonstrated that simplifying MoE routing to a single-expert “switch” per token enables trillion-parameter models at constant FLOPs, with up to 7× pre-training speedup vs. dense T5 baselines; a cornerstone of all subsequent sparse LLM architectures with 4,000+ citations.
Co-creation of ChatGPT — Was part of the founding team that built and launched ChatGPT in November 2022, shaping the post-training methodology (RLHF + instruction fine-tuning) that became the template for aligned conversational AI.
OpenAI’s Operator / Agent — Credited by Periodic’s own materials as part of the team that created OpenAI’s agentic product, which enables models to take autonomous actions in browsers and software environments.
Revisiting Fundamentals of Experience Replay (ICML 2020; with Ramachandran, Agarwal, Bengio, Larochelle, Rowland, Dabney) — Systematic empirical analysis showing that larger replay buffers and higher replay ratios substantially improve Q-learning agents in ways previously underappreciated; a foundational RL systems contribution.
Periodic Labs — Co-founded and capitalized at $300M seed to build autonomous scientific laboratories that use physical reality as the RL reward signal for AI scientists, targeting materials discovery beginning with superconductors.

Awards & Recognition

$300M seed round at Periodic Labs (2025) — Among the largest seed rounds in AI history, reflecting the strength of the founding team’s track record across ChatGPT, Switch Transformers, GNoME, and neural attention.
Google Scholar citation count — Over 51,900 cumulative citations, with Switch Transformer alone accounting for more than 4,000.
VP of Post-Training, OpenAI (promoted Fall 2024) — One of the highest research leadership roles at OpenAI, overseeing the fine-tuning work directly responsible for the quality of deployed models.

Key Relationships

Yoshua Bengio & Hugo Larochelle — PhD co-advisors at Mila / Université de Montréal; Bengio in particular represents Fedus’s entry into the deep learning mainstream from a physics background.
Barret Zoph & Noam Shazeer — Co-authors of the Switch Transformer at Google Brain; Shazeer is now CEO of Character.AI and a foundational figure in Transformer architecture.
Ekin Doğuş Çubuk — Periodic Labs co-founder; former Google DeepMind materials scientist and lead author of the GNoME paper; the pairing of Fedus’s LLM post-training expertise with Çubuk’s experimental materials science background is the core intellectual thesis of the company.
Dzmitry Bahdanau — Periodic Labs team member and creator of the original neural attention mechanism (2014 Bahdanau et al. paper); his presence signals the depth of scientific talent Fedus assembled.
Rishabh Agarwal — Periodic Labs team member and co-author with Fedus on the Experience Replay paper; a prominent RL researcher from Google Brain.
Alexandre Passos — Periodic Labs team member; credited as a creator of OpenAI’s o1 and o3 reasoning models, bringing frontier reasoning expertise into the materials discovery mission.
Sam Altman / OpenAI leadership — Fedus’s departure was publicly described as amicable, with Altman’s organisation viewing AI for science as “strategically important” for AGI, underscoring the alignment between Periodic’s mission and OpenAI’s long-term goals.

Personal Style

Fedus’s intellectual identity is built around a genuine physicist’s disposition: a preference for verifiable ground truth, suspicion of benchmark saturation, and an instinct to ask whether a model’s capability has been tested against reality rather than the training distribution. This shows up in the Switch Transformer’s emphasis on measuring actual wall-clock speedups rather than theoretical FLOPs, in his post-training work’s focus on what deployed models actually do for users, and most directly in Periodic’s core premise that AI will not learn genuine scientific reasoning until it runs real experiments and fails in real laboratories. His public communication is characteristically understated for someone with his track record — his departure tweet from OpenAI was notable for its warmth rather than self-promotion.