🔬 The Coolest Diffusion Research Isn't in LLMs — Evan Feinberg & Sergey Edunov, Genesis Molecular AI

**Evan Feinberg** (0:00)
I remember very clearly in like 2017, 2018, talking about GANs and how generative adversarial networks and how they're clearly the future of image generation. Obviously, they didn't work very well for proteins or protein-legged systems. And we sort of had to wait for the right primitive to get created. And that turned out to be diffusion, which turned out to be a much more useful primitive for the space. What's kind of cool is right now, for people that are interested in really core fundamental AI research, actually some of the most innovative diffusion research is happening in our field. It's happening in 3D structure prediction right now. No one would have predicted that then, but now that's a pillar of diffusion, I'd say.

**RJ Haneki** (0:48)
Hi there. Welcome to the Latent Space AI for Science Podcast. I'm RJ. Haneki, CTO of Miraomics. I'm joined by my co-host Brandon Anderson.
We're privileged to have Evan Feinberg, a founder and CEO of Genesis Molecular AI, and Sergey Edunov, who led Llama 2 and Llama 3 pre-training before he joined Genesis as CTO.

**Sergey Edunov** (1:12)
Hi, I'm Sergey. I studied physics at school. It was a long time ago.
And after graduation, I happened to work in software engineering, but I will never need physics again. But then ML came up and became a thing, and turns out that a lot of things you're doing in ML were actually very similar to what you would do in physics. So I jumped in on ML bandwagon and I did a lot of AI research, being a part of their team in Facebook for quite a while.
I later on led Llama team on Llama 2 and Llama 3 models. And then recently I decided to pivot my career again and recover my roots in physics a little bit, and I joined Genesis as the CTO.

**Evan Feinberg** (1:57)
Hey, I'm Evan. I'm the founder and CEO of Genesis Molecular AI. And like Sergey, also a physics major. I was a bit different from everyone in my family growing up. Almost everyone was a medical professional of some kind, and my sister became an accomplished TV writer, playwright, novelist, and my love was physics and computer science.
But my mental model of an adult was, well, you should help people and help patients ideally. So I was always searching for the right way to do that. And after you're arriving at Stanford, doing my PhD in Vijay Pandey's lab, we were excited in the mid 2010s of everything amazing going on in machine learning, to use a data term for images and for language.
And around the same time that Sergey was at fair working on a lot of big graphs, I was at Stanford working on many small graphs. Molecules are really networks of atoms and bonds and spatial interactions. And if you're at the right place, the right time to bring to bear our backgrounds in physics, to improve AI algorithms for looking at molecules.
We published a few papers in the area of graph machine learning. And sort of like Sergey, I thought, well, I won't need this physics again because there's machine learning. But it turns out the massive amounts of simulations on GPUs of proteins that we ran also came in quite handy as Genesis evolved. And we've been really excited in the past few years to figure out how to build foundation models for a totally new domain and make them useful for patients.

**Brandon Anderson** (3:33)
Yeah, that's a really nice lead in to my first question, which is, we've both been in this domain of sort of machine learning for molecules and bio for roughly 10 years. It's kind of an entire generation of tech bio has come and gone since then.
Well, a lot of machine learning for molecules has been, I think, quite effective. One domain where it's been not effective has, where historically really resisted machine learning modeling has been the world of protein small molecule interactions. And with some of recent advances that Genesis has put out, it seems like you might have actually started to make real improvement on this in a way that we haven't seen for a long time. Can you talk about what you have done, what Genesis has done, the developments which led to improvement and why you think this is actually a real improvement over some of the traditional machine learning strategies which were ambiguously helpful?

**Evan Feinberg** (4:30)
I totally agree, Brandon. And the amazing thing is, one of the really remarkable things, is that when we were founding the company, initially as a spin out of the research we were doing in AI at Stanford, we were afraid that we could be too late.