Andrej Karpathy

It might take unintuitive forms where you are telling the GPT, like, hey, you have a declarative memory bank to which you can store and retrieve data from, and whenever you encounter some information that you find useful, just save it to your memory bank.

And here's an example of something you have retrieved, and here's how you say it, and here's how you load from it.

You just say, load...

whatever, you teach it in text, in English, and then it might learn to use a memory bank from that.

It's not just text, right?

You're giving it gadgets and gizmos.

So you're teaching some kind of a special language by which it can save arbitrary information and retrieve it at a later time.

And you're telling it about these special tokens and how to arrange them to use these interfaces.

It's like, hey, you can use a calculator.

Here's how you use it.

Just do 53 plus 41 equals.

And when equals is there, a calculator will actually read out the answer and you don't have to calculate it yourself.

And you just like tell it in English.

This might actually work.

Yeah, I think so.

So Gatto is very much a kitchen sink approach to reinforcement learning in lots of different environments with a single fixed transformer model, right?

I think it's a very sort of early result in that realm.

But I think, yeah, it's along the lines of what I think things will eventually look like.

Yeah, I'm not a super huge fan of, I think, all these interfaces that look very different.

I would want everything to be normalized into the same API.