Training an AI Scientist with Feedback from Reality, w- Liam Fedus & Ekin Dogus Cubuk (from a16z)

**Erik Torenberg** (0:00)
Hello, and welcome back to The Cognitive Revolution. Today, I'm pleased to share a special cross post from the A16z podcast, featuring A16z General Partner, Anjney Midha, who also recently joined me on The Cognitive Revolution to discuss sovereign AI. Today, in conversation with Liam Fedus, former VP of Post-Training Research and co-creator of ChatGPT at OpenAI, and Ekin Doguş Cubuk, former head of Material Science and Chemistry Research at Google DeepMind. Together, they've co-founded Periodic Labs and just announced a $300 million seed investment led by Andreessen Horowitz. Before diving in, a quick note. While Turpentine was recently acquired by A16z, my editorial independence remains unchanged. And I'm sharing this episode simply because I think it does offer a really valuable perspective on the future of AI-powered science. Regular listeners will no doubt notice some overlap between this conversation and our recent episode with Radical AI. Both companies believe that there simply isn't enough high-quality experimental data in the existing scientific literature to train foundation models for physics and chemistry. And both have raised serious capital to build automated physical laboratories that are meant to connect AI-generated hypotheses directly to real-world experiments, using feedback from physical reality as the reinforcement learning signal, with the goal of teaching AI models a form of scientific intuition and, thereby, accelerating scientific progress itself. Of course, there are still many possible ways to focus such an ambitious project. And while Radical AI has recently announced a contract with the US. Air Force to develop high-entropy alloys for use in hypersonic aviation, Periodic Labs has set the goal of discovering a high-temperature superconductor as their North Star, with the expectation that to get there, they'll need to achieve countless sub-goals along the way, including autonomous synthesis and autonomous characterization.
Importantly, while the science and macro strategies are similar, the conversations are actually quite different. Whereas I tend to explore the technical details in arguably tedious depth, Anjney focuses much more on the human and organizational dimensions of building an AI for science company. As you'll hear, because no human comes close to holding all of the scientific knowledge and intuition that Periodic Labs hopes to train into their AI systems, they prioritize people with intense curiosity and mission alignment, and they don't require advanced degrees. They take pride in their no-stupid-questions culture, and they host weekly teaching sessions in which ML researchers, physicists, and chemists can all learn from one another. And recognizing that even $300 million won't be enough to achieve their ultimate goals, and that even a wildly successful company is only one part of the broader scientific ecosystem, they have thoughtful plans to commercialize their progress in the form of an intelligence layer for advanced manufacturing companies, while also starting a grant program, even at this early stage, meant to elicit key contributions from academia. Overall, I love the vision and ambition on display here, and I admire the conviction with which a16z and others are backing it. And while this doesn't come up in the episode, I've long believed that long-term AI safety might best be achieved by creating domain-specific superintelligences. Which would mean that the AIs that advance fundamental science don't need to have advanced theory of mind or persuasion skills. And in any case, as much fun as I'm having playing around with Sora 2, it does seem quite clear that a future of truly radical abundance requires AI systems that go beyond the digital world and iterate directly against nature's own ground truth. With that, I hope you enjoy this conversation about building an AI research company meant to develop systems that autonomously explore and deeply understand the physical world. From the a16z podcast with hosts Anjney Midha and Liam Fedus and Ekin Dogus Cubuk, founders of Periodic Labs.

**Liam Fedus** (3:55)
Ultimately, science is driven against experiment in the real world. And so that's what we're doing with Periodic Labs. We're taking these precursor technologies and we're saying, okay, if you care about advancing science, we need to have experiment in the loop.

**Anjney Midha** (4:09)
The applications of building an AI physicist, for lack of a better word, that can design the real world are so broad. You can apply them to advance manufacturing, you can apply the material science to chemistry. Any process where there's R&D with the physical world required, it seems like will benefit from breakthroughs that Periodic is working on.

**Ekin Doguş Cubuk** (4:28)
For example, if we could find a 200 Kelvin superconductor, even before we make any product with it, to be able to see such quantum effects on such high temperatures, I think would be such an update to people's view of how they see the universe.

**Anjney Midha** (4:44)
So, Liam, you were the co-creator of ChatGPT. Doğuş, you were running some of the physics teams at DeepMind. Let's talk about how you guys met and what was the moment where you realized that you guys had to leave both of those labs to start Periodic.