Gwern Branwen

And we're just going to focus on the cake for a while.

And now we've actually figured out a good recipe for baking a cake, which wasn't true before.

Before it seemed like you were going to have to kind of brute force it end to end from the rewards.

But now you can do the Lacoon thing of like learning fast on generative models and then just doing a little bit of RL on top to make it do something specific.

Right.

Yeah, I've been thinking about that quite a lot.

What do I want to do?

And what would be useful to do?

I'm doing things now because I want to do them, regardless of whether it will be possible for an AI to do them in like three years.

I do something because I want to, because I like it.

You know, I find it funny or whatever.

Or maybe I think carefully about kind of just doing the human part of it, like laying out a proposal or something.

If you take seriously the idea of getting AGI in just a few years, you don't necessarily have to implement stuff and do it yourself.

You can sketch out clearly what you want and why it would be good and then how to do it.

And then basically just wait for the better AGI to come along and actually do it then.

Unless there's some really compelling reason to do it right now and pay the cost in terms of scarce time.

But otherwise, I'm trying to write more about what isn't recorded.

Things like preferences and desires and evaluations and judgments.

Things that an AI couldn't replace, even in principle.

The way I like to put it is that the AI kind of can't eat ice cream for you, right?