Michael Kratsios

And that's where you kind of had this first phase of large language models.

And the second one was coding.

And if you think about how do you get a really good coding model, you have to train it on existing code.

And that was, again, something that is relatively easier to acquire than other types of data.

And you saw great progress and jumps in the coding models.

I think the third big sort of shift that hasn't really been touched on yet, which the government itself is trying to push on, is the AI for science question.

And why it's so challenging for scientific discovery to tie in with the way that LLMs are traditionally trained is that the science data is extraordinarily fragmented and it's not done in a way or formatted in a way that can easily be applied to a large language model sort of like training run.

And if you think about scientific discovery, it's spread out across so many different disciplines.

You have chemistry data, you have math data, you have material science data.

And all of that is all types of different formats.

And our effort in administration, we launched something called the Genesis Mission, which is our attempt to sort of make these big, bold leaps in AI for scientific discovery.

And our national labs at the Department of Energy have been doing incredible research over the last 50, 60 years to

All of that is sitting and is ready to be used to be trained for these models.

My hope is that over the next year, we're going to see a lot more work in scientific discovery to be able to actually accelerate how quickly we can choose which experiments to run, run those experiments quickly,

go back and figure out what we did wrong and run them again.

And this ties in with lots of interesting ideas that people have around some of these AI labs where you essentially have, you can put in the thesis or the hypothesis and ultimately these labs can do lab experiment itself and move forward.

So that's kind of the dream that I have that ultimately we as a country can almost double our R&D output over the next 10 years because of AI.

Yeah, I think they're the ones that I think can make a big impact are first the the the the experimentation and training runs around fusion technology.

are extraordinarily computation heavy.

And they themselves, if we can have a faster feedback loop on how we do these simulations for fusion, we can move the timelines in for fusion.