Casey Newton

Yeah.

You know, I don't want to waste David's time here.

All right.

We'll save that for later.

The description I loved, this came from Gavin Leach today, was that talkie is creating beautiful prose by a terrible person, which is consistent with some of my tests, which is like, this thing actually does write quite well.

And actually, to my ear, like much more literary than some of the more recent models trained on more recent data.

But yeah, it is not, it is clearly the product of its time, or at least the time of its data.

Yeah.

Now, are you going to point Taki at any big sort of scientific discoveries and see whether it can make them?

I mean, Demis Hassabis at Google DeepMind has this sort of theory that AGI should be able to discover Einstein's theory of relativity if you just give it all of the pre-existing scientific literature at the time.

Are you hoping to use this model or a descendant of this model for anything like that?

David, what are you building next?

Are you going to build a bigger version of Taki and keep trying to get it to perform better?

You can ignore him.

Anyways, I'm going to go.

Well, David, fascinating experiment.

And people can go try talkie for themselves.

It's at talkie-lm.com.

What's a good goodbye to a podcast guest?

Don't worry.