The creator of ChatGPT teams up with the mastermind from DeepMind to tackle high-temperature superconductivity with AI, and half of Silicon Valley is vying to invest.

Liam Fedus, the former vice - president of research at OpenAI, and Ekin Cubuk, a leading figure in materials science at DeepMind, co - founded Periodic Labs and emerged from stealth mode with a staggering $300 million in seed funding, shocking Silicon Valley. However, OpenAI, their former employer that once offered blessings, did not participate in this round of investment.

The former vice - president of research at OpenAI who helped create ChatGPT, and the leader of materials science and chemistry research at Google DeepMind, have joined forces to start a new venture!

Their new company, Periodic Labs, made a splash by securing $300 million in seed funding right off the bat. The investment lineup is star - studded, with a16z leading the round, followed by DST, NVIDIA's NVentures, Accel, Felicis, etc., as well as tech titans like Jeff Bezos, Eric Schmidt, and Jeff Dean.

Such an exaggerated funding amount and investor lineup are extremely rare among startups, causing quite a stir in the industry.

What exactly does this company aim to do? Why has it attracted so much attention?

Leaving Top Labs to "Truly Do Science"

One of the co - founders is William Liam Fedus, the former vice - president of research at OpenAI and a core team member who participated in the creation of the groundbreaking ChatGPT.

The other is Ekin Dogus Cubuk (nicknamed "Doge"), who previously led materials science and chemistry research at Google DeepMind.

He is also one of the leaders of the GNoME project, which shocked the academic community by using AI to discover over 2 million new crystal materials in 2023.

Presumably, they have reached the pinnacle in their respective fields and have a bright future ahead.

William Liam Fedus

As the head of the post - training department at OpenAI, Liam Fedus was mainly involved in the research and development of the underlying models for ChatGPT, APIs, and AI agents.

Previously, he worked at Google Brain, focusing on optimizing the efficiency of neural networks through MoE technology.

In 2022, he officially joined OpenAI. Initially, as a core developer, he joined the reinforcement learning team and was one of the co - creators of ChatGPT, mainly responsible for data processing and model evaluation.

During this period, he led the post - training R & D of several important models (including 4o, o1 - mini, o1 - preview, etc.).

In October 2024, Fedus succeeded Barret Zoph and was promoted to the head of the post - training team.

At that time, CTO Mira Murati and Chief Research Officer Bob McGrew also left the company.

Fedus obtained a bachelor's degree in physics from MIT (during which he participated in a directional dark matter detector project: DMTPC) and another bachelor's degree in physics from the University of Cambridge.

In 2016, he received a master's degree in elementary particle physics from the University of California, San Diego, under the supervision of David Meyer and Gary Cottrell.

Subsequently, he earned a Ph.D. in computer science from the University of Montreal, studying under Yoshua Bengio and Hugo Larochelle.

Ekin Dogus Cubuk

The other co - founder, Ekin Dogus Cubuk, was previously a research scientist at Google DeepMind.

He joined Google Brain in 2017, participated in the flagship project GNoME in the field of materials science discovery, and built several automated synthesis experiment platforms within the company, focusing on using AI to find new materials.

He holds a Ph.D. in condensed matter and materials physics and computational science from Harvard University.

However, in March this year, Fedus resolutely resigned from OpenAI, and Cubuk also chose to leave DeepMind to embark on entrepreneurship.

The Origin: Flipping Tires Together at Google

The two first crossed paths at Google. An interesting incident of flipping tires together became the catalyst for their acquaintance - but what truly brought them together was their clear understanding of the limitations of the current AI research path and their shared pursuit as AI scientists.

Currently, training AI mainly relies on internet text. Although the internet seems boundless, it is actually limited.

It is estimated that there are about 10 trillion tokens of valuable text data on the internet (about 1 - 2 tokens per English word). The top large models have almost exhausted this data in recent years.

Without fresh data, it is difficult to achieve qualitative breakthroughs by simply increasing the parameter scale indefinitely.

As Fedus bluntly stated in an interview:

The main goal of AI is not to automate office work for white - collar workers. The main goal of AI is to accelerate science.

In his view, the hype around large - model applications in Silicon Valley these days is somewhat "intellectually lazy." AI should truly focus on accelerating the speed of scientific discovery.

Cubuk also pointed out that it is impossible to come up with earth - shattering scientific discoveries by simply having large models reason in text for days and nights. Real scientific breakthroughs require a large number of experiments and countless failures.

What current AI models lack is precisely the "hands - on experiment" part.

So, the two hit it off at the beginning of this year: Instead of being limited by existing data, they decided to let AI "walk into" the laboratory and create data from scratch.

They aim to create an "AI scientist" that can propose hypotheses and conduct repeated experiments in the real world, learning from the experimental results, whether they are successful or not.

As Fedus said when communicating with investors:

To make AI truly do science, it must be allowed to do real science hands - on.

Peter Deng, a former colleague of Fedus at OpenAI and now an investment partner at Felicis, even stopped in his tracks on a San Francisco hillside when he first heard this statement and decided to invest immediately.

In his view, large models only master the "normal distribution" of training data, that is, existing human knowledge;

To achieve original innovation, AI must step out of its comfort zone, propose new hypotheses like a scientist, and verify them.

This concept became the origin of Periodic Labs' entrepreneurship.

Autonomous Lab: Making Nature an Environment for Reinforcement Learning

Cubuk summarized that three major technological advancements in recent years have made all this possible.

First, robotic arms capable of handling powder synthesis have become reliable, meaning machines can automatically mix raw materials and sinter new materials;

Second, machine - learning - driven physical simulations are more efficient and accurate, sufficient to simulate complex materials and chemical systems;

Third, the reasoning ability of large language models (LLMs) has improved significantly and can handle more complex planning and analysis.

The leaps in these three fields together paint a picture: AI can make assumptions and calculations in the virtual world, conduct hands - on experiments in the real world, and then analyze the experimental results and adjust its thinking.

Now is the right time to build an automated closed - loop laboratory for materials science.

In fact, Cubuk was one of the participants in the relevant pioneering work.