Noam Shazeer

And that enabled us to go from taking a night to translate a sentence to basically doing something in 100 milliseconds or something.

All of these are very flattering.

They're pretty funny.

They're like an April Fool's joke gone awry by my colleagues.

I think once we built that for translation, the serving of large language models started to be used for other things, like completion of, you know, you start to type and it suggests what completions make sense.

So it was definitely the start of a lot of uses of language models in Google.

And, you know, Noam has worked on a number of other things at Google, like spelling correction systems that use language models.

But his spelling correction system he built in 2001 was amazing.

He sent out this demo link to the whole company.

And I just tried every butchered spelling of every few word query I could get.

I, like, scrambled Uggs Bundict.

Oh, I remember that one.

Instead of scrambled eggs Benedict.

Yeah, it's nice because you then have the right answer and then you can train on like all but the current word and try to predict the current word.

And it's this kind of amazing, you know, ability to just learn from observations of the world.

I mean, I like to think of a lot of ideas as –

they're kind of partially in the air where there's like a few different, maybe separate research ideas that one is kind of squinting at when you're trying to solve a new problem.

And you kind of draw on those for some inspiration.

And then there's like some aspect that is not solved and you sort of need to figure out how to solve that.

And then the combination of like some morphing of the things that already exist and some new things lead to some new breakthrough or a new research result that didn't exist before.