This is John Batchelor.

A very helpful conversation with Spencer Clavin of the Claremont Book of Review of Books

about using GROC, about how GROC learns from humans, interacting,

and then becomes better at optimizing the next word or the next sentence that GROC gives back.

And what happens if you give Grog instead of human talk, or human information,

if you give it the information from Grog?

The end is not happy.

Here's Spencer Claven describing an exchange with Grock.

He types it out, he doesn't talk it,

and what he learns from it. Much more of this tonight and in coming weeks. The philosophy

of AI to be discovered. I don't tend to quiz GROC at that level of metaphysics. Maybe I should. And I'm probably one of the only people I know who doesn't talk verbally to GROC. I type my questions and get answers. And it feels, I suppose, more impersonal that way. And maybe that reflects my own belief about what's going on. But I do

think it's interesting what you say about learning or adapting. One of the things you'll notice

about Grock and other chatbots is they usually conclude their answer with another question.

Yes.

So yeah, you say, I don't know, what's a recipe for a good cocktail in the autumn?

And it comes up with this elaborate description and then says, would you like me to narrow it down?

Can you tell me how many guests

you're having over? And it does this in part because your answers then become part of its

training data. So having been trained on all this material, these machines need more human

output in order to get better and sharper at predicting how humans

might behave or expect them to behave.

And one of the really interesting things is that they have to get that input from us,

...