Grant Harvey

Or if you give it a complex, subjective question, it's going to give you some different answers.

But how the hell do you decide which is better?

That's where evaluation is so critical.

I think Andre Carpenter- Someone has to read-

Well, Henri Carpenter had a tweet about six months ago where he said, like, the entire AI is just an evaluation problem now.

Because it's like the models are getting so complex that understanding what better means, they're not linearly better.

You might get an improvement in, say, let's imagine you get an improvement in creativity, but that also makes it more verbose and more likely to hallucinate.

You might be more creative, which is great.

I mean, AI is not magic.

It's one of the biggest problems.

People think AI is not magic.

AI is just better predictive text.

Like it's very good predictive text.

But fundamentally, if you go into your phone and you go into WhatsApp and you type a word and you keep pressing next word, that's predictive text too.

It's just not as good because it's based on a much simpler model.

Like all AI is trying to do is find the next character or the next pixel or whatever.

It has no concept of truth.

It has no concept of anything.

And so these are just various different mechanisms to make it better at finding the next token.

And that could be because you've given it a load of label data, which is making it be rewarded for certain behaviors, or you've given it loads of information, which is making it be punished for certain behaviors.