Jeff Klune

So the AI scientist is a system that basically has AI doing the entire arc of what a human scientist does when they're doing scientific work.

So it comes up with its own idea for how to improve a system or its own hypothesis to test.

It comes up with a plan of experiments that will test that hypothesis or if that idea works.

It programs or creates the software to run those experiments.

It looks at the data.

It iterates a few times if it thinks it's necessary.

It visualizes the data like charts and plots.

It looks at those plots and says, hmm, do I actually have enough data yet to decide if this idea or this hypothesis is true or false?

And then eventually it writes up the entire paper describing the results from the first word in the title to the last period in the references section.

And then it does one final step.

It actually conducts its own peer review where it studies how good is this science?

Is it good enough to be published?

And if it is good enough, it kind of returns to the human and says, hey, I've done a bunch of experiments and some scientific work.

Here are my findings in a human readable form in the form of this scientific paper.

That's correct.

Yeah, the stuff of science fiction is becoming reality at a remarkable pace.

So in this case, you know, my whole career, I've been doing neural nets for over 20 years.

We've always made the joke, every time we automate a little piece of the scientific process with AI, we say, you know, one day we'll eventually have the system write the full paper.

And a few years ago, we decided...

We think it's now possible.