Gavin Baker

I think the 6 is for next year, but I don't know.

I think the 6 is for next year, but I don't know.

I think AI, and Nick, I sent you a chart. I don't know if you could flash it up, but I think AI is going to make more progress per quarter in 2025 than it did per year in 23 and 24. And the reason is just with 01 and 03, this is Arc AGI, it's designed, you know, we keep changing the goalposts for the Turing test and AGI, I'm sure we're going to change it again.

I think AI, and Nick, I sent you a chart. I don't know if you could flash it up, but I think AI is going to make more progress per quarter in 2025 than it did per year in 23 and 24. And the reason is just with 01 and 03, this is Arc AGI, it's designed, you know, we keep changing the goalposts for the Turing test and AGI, I'm sure we're going to change it again.

we're going to blow through this and i just think what has happened is we were scaling around round one on one axis which was pre-training and then we started scaling around inference time compute and it's very clear that we have now added a third axis of scaling performance and that is reasoning and what this is is these models the internet is composed of answers people giving answers and what the models really benefit from

we're going to blow through this and i just think what has happened is we were scaling around round one on one axis which was pre-training and then we started scaling around inference time compute and it's very clear that we have now added a third axis of scaling performance and that is reasoning and what this is is these models the internet is composed of answers people giving answers and what the models really benefit from

is kind of the internal monologue of somebody getting to that answer. And this is, you know, in AI terms, they call it a reasoning trace. And it's one reason, you know, 18 months ago, everybody was like, oh, wow, the more code you train a model on, the better it does and all sorts of things that have seemingly nothing to do with code.

is kind of the internal monologue of somebody getting to that answer. And this is, you know, in AI terms, they call it a reasoning trace. And it's one reason, you know, 18 months ago, everybody was like, oh, wow, the more code you train a model on, the better it does and all sorts of things that have seemingly nothing to do with code.

But all code, you see kind of the reasoning, the internal monologue, the thought process, the step by step. And so what's happening is you're using models to generate synthetic data that contains these reasoning traces. So you ask a model, you know, solve this problem, it has to be a problem that is functionally verifiable, that has an answer, or you know, the answer, we say show your work.

But all code, you see kind of the reasoning, the internal monologue, the thought process, the step by step. And so what's happening is you're using models to generate synthetic data that contains these reasoning traces. So you ask a model, you know, solve this problem, it has to be a problem that is functionally verifiable, that has an answer, or you know, the answer, we say show your work.

And we have to do that many, many different ways and times. And then you pick the best ones, and you kind of feed those back into the model, you apply some reinforcement learning to it. And so now you're scaling along three axes that are multiplicative with each other. And I just,

And we have to do that many, many different ways and times. And then you pick the best ones, and you kind of feed those back into the model, you apply some reinforcement learning to it. And so now you're scaling along three axes that are multiplicative with each other. And I just,

A guy on the Google team tweeted maybe five days ago, it's going to be a straight shot to ASI, artificial superintelligence. And I think that might be right. Ilya gave a talk at NeurIPS, Ilya Sutskever, one of the original pioneers of this field. And he said something that I thought was...

A guy on the Google team tweeted maybe five days ago, it's going to be a straight shot to ASI, artificial superintelligence. And I think that might be right. Ilya gave a talk at NeurIPS, Ilya Sutskever, one of the original pioneers of this field. And he said something that I thought was...

scary and he said these models that reason are inherently unpredictable so the best reasoning models in the world today are the ones that play games the kind of alpha go alpha zero style models and they are constantly making unpredictable moves that no human grand master ever could have come up with and now these models are going to be making similarly unpredictable leaps in all sorts of domains

scary and he said these models that reason are inherently unpredictable so the best reasoning models in the world today are the ones that play games the kind of alpha go alpha zero style models and they are constantly making unpredictable moves that no human grand master ever could have come up with and now these models are going to be making similarly unpredictable leaps in all sorts of domains

which hopefully will be awesome, but it might not be.

which hopefully will be awesome, but it might not be.

Unquestioned, 1923, season two. I'm normally a science fiction kind of fantasy spy.

Unquestioned, 1923, season two. I'm normally a science fiction kind of fantasy spy.