Terence Tao

Yeah, this decade, I can see it like making a conjecture between two unrelated, two things that people thought was unrelated.

Oh, interesting.

Yeah, and actually has a real chance of being correct and meaningful.

No, that would be truly amazing.

Current models struggle a lot.

I mean, so a version of this is, I mean, the physicists have a dream of getting the AIs to discover new laws of physics.

Right.

The dream is you just feed it all this data, and here is a new patent that we didn't see before.

But the current state of the art even struggles to discover old laws of physics from the data.

Or if it does, there's a big concern of contamination, that it did it only because somewhere in its training data it already somehow knew Boyle's law or whatever law you're trying to reconstruct.

Part of it is that we don't have the right type of training data for this.

For the laws of physics, we don't have a million different universes with a million different laws of nature.

A lot of what we're missing in math is actually the negative space.

So we have published things of things that people have been able to prove and conjectures that ended up being verified or maybe counterexamples produced.

But we don't have data on things that were proposed and they're kind of a good thing to try.

But then people quickly realized that it was the wrong conjecture and then they said, oh, but we should actually change

our claim to modify it in this way to actually make it more plausible.

There's a trial and error process, which is a real integral part of human mathematical discovery, which we don't record because it's embarrassing.

We make mistakes and we only like to publish our wins.

And the AI has no access to this data to train on.