Adam Brown

I think maybe, again, the analogy with chess programs is a good one here.

They will often consider way more knowledge.

possible positions as a Monte Carlo research than any human chess player ever would.

And yet, even at human level strength, if you fix human level strength, they're still doing way more search so that their ability to evaluate is maybe not quite as natural as a human.

So the same, I think, would be true of physics.

If you had a human who had read as much and retained as much as they had, you might expect them to be even stronger.

The last physics query, well, a recent one was I asked it to explain to me the use of squeezed light at LIGO, which is a topic that I always felt like I should understand.

And then try to explain it to somebody else and realize that I didn't understand it and went and asked the LLM.

That blew me away that it was able to, like—

exactly explain to me why what I was thinking was incorrect.

So why do we use this particular form of quantum light in interferometer used to discover gravitational waves?

The reason that's a good topic is perhaps because it's an advanced topic.

Not many people know that, but it's not a super advanced topic.

There are

Out of a physics literature of millions of papers, there have got to be at least a thousand on that topic.

If there was just a handful of papers on a topic, it's typically not that strong at it.

Yeah, I don't know the answer.

That is an interesting question.

I think it might be able to debug even without that.

If you do much simpler things like give these language models code, it will successfully debug your code, even though presumably no one has made that exact bug in your code before.