Dwarkesh Patel

Oh, interesting.

And the thing it was protecting against was probably another disease?

Maybe.

Prepping for this episode required a full lit review.

I needed to understand why other methods had failed to find evidence of natural selection over the last 10,000 years.

What exactly did Reich and Akbari do differently?

Honestly, this was quite subtle because the most important points were distributed across a bunch of different papers.

And it was frustrating to talk to LLMs about it because they kept getting confused.

One of them would fail to understand an important crux.

And so I switched over to a different model and that one would get tripped up on the very next point.

I ended up using Cursor to kick off a handful of models at the same time and compare the results after.

I could have one model critique the response of another.

This was super useful because while I'm not a geneticist, I do have enough taste to be able to say, hey, this answer makes sense.

These ones don't.

I also had Cursor turn this work into a flashcard so I could retain what I learned.

Cursor started as a programming tool, but I found it really great for this kind of research.

There's no other interface where I can get answers from a bunch of independent LLMs all while reading the relevant paper on the same screen.

Go to cursor.com slash thwarkesh to try it out.

One of the big takeaways for me from the paper was just that something weird happened in the Bronze Age.

And that...