Anthropic estimates coding models add 15–20% to productivity

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 37:00

   is qualitatively predicted by studies like this. but two if i just look at the end output obviously you guys are making fast progress but the fact you know the the idea was supposed to be with recursive self-improvement is that you make a better ai the ai helps you build a better next ai etc etc and what i see instead if i look at the you open ai deep mind is that people are just shifting around the podium every few months and maybe you think that stops because you've won or whatever but um but why are we not seeing the person with the best coding model have this lasting advantage if in fact there are these enormous productivity gains from the last coding model so no no no i i mean i mean i mean i think it's all like my my model of the situation is there's there's an advantage that's gradually

2. 37:44

   growing like i would say right now the coding models give maybe i don't know a a like 15 maybe 20% total factor speed up. Like that's my view. Um, uh, and six months ago it was maybe 5%. And so, and so it didn't matter. Like 5% doesn't register. It's now just getting to the point where it's like one of several factors that, that kind of matters. And, and that's gonna, that's gonna keep speeding up. And so I think six months ago, like, you know, there were several, there were several companies that were at roughly the same point because, uh, you know, this, this wasn't, this wasn't a notable factor, but I think it's starting to speed up more and more. I, you know, I, I would, I would also say there are multiple companies that, you know, write models that are used for code and, you know, we're not perfectly good at, you know, preventing some of these other

3. 38:35

   companies from, from, from using, from, from, from kind of using our models internally. Um, so, uh, you know, I think, I think everything we're kind of, kind of everything we're seeing is consistent with this kind of, um, this kind of snowball model where, where, you know, there's no hard. Again, my, my, my, my, my theme in all of this is like, all of this is soft takeoff, like soft, smooth exponentials, although the exponentials are relatively steep.

4. 39:02

   And so, and so we're seeing this snowball gather momentum where it's like 10%, 20%, 25%, you know, 40%. And as you go, yeah, Amdahl's law, you have to get all the like things that are preventing you from, from closing the loop out of the way. But like, this is one of the biggest priorities within Anthropic? Stepping back, I think before in the stack we were talking about, well, when do we get this on-the-job learning?

5. 39:29

   And it seems like the coding, the point you were making of the coding thing is we actually don't need on-the-job learning. That you can have tremendous productivity improvements, you can have potentially trillions of dollars of revenue for AI companies without this basic human ability. Maybe that's not your claim, you should clarify. But without this basic human ability to learn on the job. But I just look at, like, In most domains of economic activity, people say, I hired somebody, they weren't that useful for the first few months.

6. 39:54

   And then over time, they built up the context understanding. It's actually hard to define what we're talking about here. But they got something. And then now they're a power horse and they're so valuable to us. And if AI doesn't develop this ability to learn on the fly, I'm a bit skeptical that we're going to see huge changes to the world. Yeah. So I think two things here, right? There's the state of the technology right now, which is, again, we have these two stages.

7. 40:22

   We have the pre-training and RL stage where you throw a bunch of data and tasks into the models and then they generalize. So it's like learning, but it's like learning from more data and not learning over kind of one human or one model's lifetime. So, again, this is situated between evolution and human learning. But once you learn all those skills, you have them. And just like with pre-training, just how the models know more, you know, if I look at a pre-trained model, you know, it knows more about the history of samurai in Japan than I do.

8. 40:55

   It knows more about baseball than I do. It knows, you know, it knows more about, you know, low-pass filters and electronics than, you know, all of these things. Its knowledge is way broader than mine. So I think even just that may get us to the point where the models are better at – kind of better at everything. And then we also have, again, just with scaling the kind of existing setup, we have the in-context learning, which I would describe as kind of like human on-the-job learning but like a little weaker and a little short-term.

9. 41:30

   Like you look at in-context learning. You give the model a bunch of examples. It does get it. There's real learning that happens in context. and like a million tokens is a lot. That's, you know, that can be days of human learning, right? You know, if you think about the model, you know, kind of reading a million words, you know, it takes me, how long would it take me to read a million? I mean, you know, like days or weeks at least.

10. 41:54

    So you have these two things and I think these two things within the existing paradigm may just be enough to get you the country of geniuses in the data center. I don't know for sure, but I think they're going to get you a large fraction of it. There may be gaps, but I certainly think just as things are, this, I believe, is enough to generate trillions of dollars of revenue. That's one. That's all one.

11. 42:16

    Two is this idea of continual learning, this idea of a single model learning on the job. I think we're working on that too. And I think there's a good chance that in the next year or two, we also solve that. Um, I, I, again, I, I, I, you know, I think you get most of the way there without it. I think the trillions of dollars of, of, you know, the, the, I think the trillions of dollars a year market, maybe all of the national security implications and the safety implications that I wrote about in adolescence of technology can happen without it.

12. 42:51

    But I also think we, and I imagine others, are working on it. And I think there's a good chance that we get there within the next year or two.