Maybe for the first time, this is no longer true, right? So the famous fact here is that an H100 costs more to rent now than it did 3 years ago, even though we have much superior technology and we have much more compute in the world. Um, because as models get smarter, the opportunity cost of compute gets higher.

The amount of information per flop has been decreasing as you had to unroll 2 days worth of thinking in order to see if you even did something correctly.

We've seen the valuations of a bunch of software companies crash because people are expecting AI to commoditize software.

by late 2027, or certainly by 2028, there's going to be such wide diffusion that even open source models will be able to match the performance that the frontier had 12 months prior.

while this year might cost $30 billion, next year it'll cost $3 billion, the year after that $300 million, and by 2030 it'll be less expensive to monitor every single nook and cranny in this country than it is to remodel the White House.

However, that bottleneck goes away with AI. There are 100 million CCTV cameras in America, and you can get pretty good open-source multimodal models for 10 cents per million input tokens. So if you process a frame every 10 seconds, and if each frame is, say, 1,000 tokens, then for $30 billion, you can process every single camera in America.

Models keep getting more impressive at the rate that the short timelines people predict, but more useful at the rate that the long timelines people predict.

But people are trying to launder the prestige that free-training scaling has, which is almost as predictable as a physical law of the universe, to justify bullish predictions about reinforcement learning from verifiable reward, for which we have no well publicly known trend.