Sam Altman says AI model costs will drop 1,000-fold. Compute costs are already down 40 times year-over-year, shifting the bottleneck from training to inference. The speakers think the industry is drastically underestimating how fast AI will improve.
Chapters
Sam Altman predicting, you know, a 1,000x drop in cost.
Do you believe that number, Alex? I do. It's consistent with the 40x year-over-year hyperdeflation that we've discussed on the pod previously. The hyperdeflation of cost has been extraordinary. It's reasoning models that are enabling this massive increase in capabilities. It's not necessarily training time compute. We see the shift to inference time compute or action time compute that, that's really enabling this 1,000x increase in, call it, capability per unit price?
For all of neural network research history going back 40 years, nobody cared about inference time because the training was the bottleneck and the model wasn't smart enough to care about making it inference really fast. Now, chain-of-thought reasoning is the biggest breakthrough ever. Yeah. And you can use inference time compute to build a smarter and smarter and smarter AI. And we as a society have really just started on it in the last really 2 years or less. And that's why we're getting these massive gains. But that's also why everyone's really underestimating the next year because you look, we did 1,000x, like, what are you expecting in the next year? Oh, 2x. Like, what are you talking about?
