Nathan Lambert: belief

Transcript context

…I’d like to start with the technical definition of a scaling law- …which kind of informs all of this. The scaling law is the power law relationship between… You can think of the x-axis—what you are scaling—as a combination of compute and data, which are kind of similar, and then the y-axis is like the held-out prediction accuracy over our next tokens. We talked about models being autoregressive. It’s like if you keep a set of text that the model has not seen, how accurate will it get when you train? And the idea of scaling laws came when people figured out that that was a very predictable relationship. I think that technical term is continuing, and then the question is, what do users get out of it? And then there are more types of scaling, where OpenAI’s o1 was famous for introducing inference-time scaling. And I think less famously for also showing that you can scale reinforcement learning training and get kind of this log x-axis and then a linear increase in performance on the y-axis. So there are kind of these three axes now where the traditional scaling laws are talked about for pre-training—which is how big your model is and how big your dataset is—and then scaling reinforcement learning, which is like how long can you do this trial and error learning that we’ll talk about. We’ll define more of this, and then this inference-time compute, which is just letting the model generate more tokens on a specific problem. So I’m kind of bullish; they’re all really still working, but the low-hanging fruit has mostly been taken, especially in the last year on Reinforcement Learning with Verifiable Rewards, which is this RLVR, and then inference-time scaling. That’s why these models feel so different to use, where previously you would get that first token immediately. And now they’ll go off for seconds, minutes, or even hours generating these hidden thoughts before giving you the first word of your answer. And that’s all about this inference-time scaling, which is such a wonderful kind of step function in terms of how the models change abilities. They enabled this tool use stuff and enabled this much better software engineering that we were talking about. And this is, when we say enabled, almost entirely downstream of the fact that this Reinforcement Learning with Verifiable Rewards training just let the models pick up these skills very easily. So if you look at the reasoning process when the models are generating a lot of tokens, what it’ll often be doing is: it tries a tool, it looks at what it gets back, it tries another API, it sees what it gets back and if it solves the problem. The models, when you’re training them, very quickly learn to do this. is: it tries a tool, it looks at what it gets back, it tries another API, it sees what it gets back and if it solves the problem. The models, when you’re training them, very quickly learn to do this. And then at the end of the day, that gives this kind of general foundation where the model can use CLI commands very nicely in your repo, handle Git for you, move things around, organize things, or search to find more information—which, if we were sitting in these chairs a year ago, is something that we didn’t really think of the models doing. So this is just something that has happened this year and has totally transformed how we think of using AI, which I think is very magical. It’s such an interesting evolution and unlocks so much value. But it’s not clear what the next avenue will be in terms of unlocking stuff like this. I think that there’s—we’ll get to continual learning later, but there’s a lot of buzz around certain areas of AI, but no one knows when the next step function will really come. So you’ve actually said quite a lot of things there, and said profound things quickly. It would be nice to unpack them a little bit. You say you’re bullish basically on every version of scaling. So can we just start at the beginning? Pre-training: are we implying that the low-hanging fruit on pre-training scaling has been picked? Has pre-training hit a plateau, or are you still bullish on even pre-training?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.