Sholto Douglas

Right.

So it's like, I don't know how this... What is concretely happening?

Maybe you can walk me through...

Walk me through a day in the life of... You're working on an experimental project that's going to make the model quote-unquote better.

What is happening from observation to experiment to theory to writing the code?

What is happening?

Sorry, let me interrupt and say, in Carl Schulman's, when he was talking about it on the podcast, he did say that things like improving inference, or even literally helping it make better chips or GPUs, that's part of the intelligence explosion.

Because obviously if the inference code runs faster, it happens better or faster or whatever.

Tell me more about what do you mean by imperfect information?

Are these early experiments?

That's interesting to consider that for every chart you see in a release paper or technical report that shows that smooth curve, there's a graveyard of first few runs and then it's flat or something.

Okay, so...

But presumably it's not just like you run it until it stops and then let's go to the next thing.

There's some process by which to interpret the early data and also to look at your... I don't know.

I could put a Google Doc in front of you and I'm pretty sure you could just keep typing for a while on different ideas you have.

And there's some bottleneck between that and...

just like making the models better immediately.

Right.

Yeah.

Walk me through like, what is the, what is the inference you're making from the first early steps that makes you have better experiments and better ideas?