Thanks, and yeah the whole attempt to ban open models probably wouldn’t have hurt to mention. Although, I don’t think it’s that impactful in the long term. Microsoft already tried doing a similar crusade to ban open source, and that failed. There’s also a lot of use of open models by large US companies, so they would have a big incentive to fight such attempts. I think what’s more likely is that they’d manage to pass laws banning access to models hosted in China, but if the whole mainframe era is transitory, that doesn’t really matter either in the grand scheme of things.

And can recommend The Dot-Com Chancers Who Burst The Bubble gives a good overview, it has actual footage from 1999 and follows day traders in New York offices who were making thousands of dollars a week by clicking a mouse capturing the irrational exuberance of the era before the crash actually happened. It’s about 20 minutes long, so a pretty quick watch. For a deeper dive, this podcast breaks down the crash and explains the dynamic with distorted incentives, stock-based compensation, and the transition from value to eyeballs as a metric.

It’s not quite that simple. Models you can run locally without paying tens of thousands for the hardware are not really comparable in quality to frontier models. Even with DeepSeek, you’d need a pretty beefy machine to run the full 671 billion parameter version. But that’s rapidly changing with new techniques requiring drastically less memory and processing power. I expect that within a couple of years, we’ll be able to run something comparable to Claude on a regular consumer laptop. And I think at a certain point the model gets good enough that you don’t really need a better one. Even if the bleeding edge is going to keep getting pushed, there’s going to be a point where it’s not going to be relevant for most people. But we’re not quite there yet today.