Kevin Surace explains how recursive reinforcement learning lets AI models improve themselves, with 80% of recent gains generated autonomously. As training data runs out around 10 trillion tokens, he predicts all transformer models will become commoditized and interchangeable, erasing any single company's advantage.
OpenAI, Anthropic, Google. This AI pioneer says that in a few years, it won't matter which one you use. And here's why. Over the last year, we've seen the world shift from OpenAI is way in the lead to then Gemini for 6 months was way in the lead on many metrics. And then lately Claude is way in the lead. So I use today Fable because Fable's alive again. It came back. So I'm using Fable. But, you know, they're in the lead because they've done more recursive reinforcement learning or reinforcement programming than others have. And so everybody's in that game. And for those listening, what that means is I have my own coding engine improving my own foundational model and doing that over and over again. So now we're seeing releases come every few weeks. And if I believe what Anthropic said, possibly 80% of the improvements in, say, Fable or Mythos were actually totally written by itself. It wrote itself and improved itself. So we're getting into this self-improvement category. We knew that that would happen. We're doing it with robots, of course. Reinforcement learning, of course, is that instead of the old way to do it. So I think we're likely to see this thing really move quickly until we see an asymptote. And there will be an asymptote because there's only so much material to learn on, and we've learned on virtually all of the 10 trillion tokens there are out there, and we're gonna run out, right? Yeah. So there's probably some asymptoting of this type of transformer. Yeah. You don't think one will start running away? No, it'll turn into a commodity. It's commodity. Yeah. And they don't want to hear— they don't want
