Mm-hmm.

Uh, our data comes from our physical labs, and this is sort of our ultimate truth.

It's not enough just to do optimization against, um, some answers that were known in some papers or textbooks, because we're going beyond that, and I think that's one of the biggest differences.

But along with that, there's issues of, say, decision-making under uncertainty.

So when you're doing optimization against math, there's a, a high precision to it.

You're not really dealing with, you know, kinda like variance or uncertainties or, you know, aberrant measurements, whereas that's very key to, to our process.

Like, um, for example, when we are doing materials discovery loop, things don't come off the, out of the furnace labeled, right? Even the labeling process can be stochastic and noisy.

Yeah

And talk a little bit more about why you chose to start Periodic and what you guys are building.

Yeah.

One of the biggest objectives, I think, for this technology is to start changing the course and trajectory and speed of science.

AI for science is a really popular, uh, area now.

I think that's kinda the end game of this technology.

If we can start building up new knowledge, not just sort of, um, resurfacing and, um, recombining existing knowledge, but really like pushing the frontier, this would lead to just a huge acceleration and overall productivity, um, gives us agency to kinda begin to construct and do things in our physical world.

But I felt in order to do this, it was necessary to connect to the physical world.

Hmm.

What is your origin story?

Yeah, origin.

So I grew up in Maine and also had like a really early love for physics.

And so in college, I was actually a bit between physics, chemistry, economics, and I was part of a dark matter detection program.

And one of the key pieces for that was reconstructing the particle tracks, which became like a physics model.

But it's really a physics-informed machine learning model.

And so I started doing machine learning in physics in a very early stage.

And on the LM side as well, which...

Mm.

Mid-training allows you to inject some of your, um, company's knowledge or know-how that may not be in the existing model.

And so basically, how do you internalize that? And this is something you can do in addition to just giving a search or rag.

It's basically the difference of an employee showing up and having to figure things out on the fly versus they've already kind of internalized some of the knowledge.

And then we do reinforcement learning.

So given those three pieces, we now have agents where we can start with a, a model.

Uh, it could be extremely bad on our benchmarks.

Uh, from your perspective, what are these AI systems really good at?

talk about how... you created ChatGPT.

So ChatGPT originally, the technology evolved very rapidly over the last few years.

When we were first creating it, it was a very standard RLHF pipeline.

So you have a pre-trained model, And it's sort of like this raw substrate.

And what you're trying to do is take this auto-completion model and turn it into something useful.

The way we did it at that point was we would have supervised data.

So given some input, we would say this is a desired output.

What else?