Dave Blundin

I just want to make the point that if you look at the benchmarks that Alex was referring to, for AI to self-improve, very hard math problems like humanity's last exam are the trigger of self-improvement.

It is not a great writer, and it's definitely not funny.

But if you said to me, well, it couldn't possibly build a better AI algorithm until it's funny, that's intuitively might make sense.

That's not true at all.

I spent six years as an AI researcher.

I tell you, all you need to do is tweak the algorithms, try different parameters, try different transfer functions, and do it iteratively at very high rates.

And the algorithm just magically, it's more like evolution than writing a book.

You think so?

I don't actually write books with it, so I actually don't even know.

But I do use all the top tier models to design new neural network tests and experiments.

So exactly the self-improvement problem.

And every dot release is dramatically, shockingly better.

Let me give you another thought experiment, just before we... Because people underreact, and it's more dangerous to underreact than to overreact.

But if I go to somebody and I say, look, here's Claude, 4.7, and here's Stephen Kotler.

Let me give them both this very difficult writing task and write something really entertaining in the next five minutes.

And you just kill it.

You crush it.

You go, great, I'm better than Cloud 4.7.

Now you say, Cloud 4.7, I want you to write it a billion times.

And then I'm going to put a prose process on top of that to select the very best article.