Scoring an experiment before training it, Monte Carlo tree search style

Mm.

⁓ And that to me, adding to the tree of knowledge is what is really super intelligent in some ways. To your point on the, yeah, the Monte Carlo tree search is really a great way of branching out and like exploring the space of experiments. But the key thing that they had is also determining how good certain rollouts are. So if you play a game of go and or even chess and you roll it out. Basically, the player makes a certain set of moves and the other player makes its other set of moves. How good is that initial play? Namely, how good was your first move after having rolled out a ton of moves? So you sort of see into the future the effect that you have by making one action today, one action now.

Mmm. Mm.

to me actually is very analogous to what we're doing here because we need to understand what experiment to run next in order that the next set of experiments are going to be really good and having that sort of sort of the proxy for understanding how good that experiment might be is You know, there are lots of different ways to do this. So like Monte Carlo tree search uses like upper confidence bound and these various techniques. But I actually think that we can use something like a regression language model, take, you know, like the program itself, go ahead and like predict, assign it like one score. That score is it's predicted validation loss. So how, how well the program scores, how well the model performs would perform out of distribution without having to train it. ⁓ that's the, that's the key part because training it takes hours, right?

Mm. Mm.

But what if we just have this really good model that can tell us like how good the Singh performs without training it and that is our score that we use in the sort of Monte Carlo tree search and That I think is how we get to move 37 and AI research