You know for a blog that’s on its face about computational complexity, you’d think Scott would show a little more skepticism to the tech bro saying “all we need is 14 quintillion x compute to solve the Riemann hypothesis”

they phrase it so your brain will think otherwise, but I just want to point out the maybe obvious so nobody else has to do a double-take at the quoted paragraphs:

their example of what an ultra-intelligent AGI could do is of a human speedrunner using a glitch to beat Minecraft in 20 seconds. this asshole is taking something humans are excellent at and saying “but what if AI could do this, and also what if Minecraft was real life?” this is literal baby shit. like, even tool-assisted speedruns are the product of a shitload of human research into the problem space, and the tool’s just executing impossibly precise game inputs programmed, again, by a human. this is another space where AI sucks compared with regular human effort.

and speaking of which, does anyone remember the early OpenAI and Google marketing where they had an LLM play pac-man or some shit at supposedly superhuman levels? can anyone dig up an outcome for any of those in the form of a record on any credible speedrunning site or during an event? cause speedrunning has a ton of categories including stuff like dog-assisted runs, where you train your dog to play the game, and it’s all considered valid as different forms of skill applied to the game. the one thing you can’t do is cheat, and they’re very good at verifying runs (ie, you must be provably using the method you claim, and you can’t splice video together or use emulator cheats to achieve a better run). so where’s the verified LLM speedrun records?

The error bars here are, of course, extremely large. Still,

Me, as a jock bully: Well of course these nerds would find finishing in 20 seconds an achievement