Gildilinie Gremlin 🏳️‍⚧️

btw this is what we're all arguing over. this is what vibe coders are defending to the death. this is what financiers are investing in. this is what we're dumping every carbon atom available into the air for. this is what we're doubling electricity prices for. this is what we're quintupling ram prices for. this is what people are going insane for. this is what we're addicting people to. this is from today

usually directed search isn't chosen unless nothing else is available because it takes a metric ass ton of compute. but that's what they have

you can apply directed search to any problem with an unknown answer but where a guess is easy to check + a guesser that sometimes moves in the right direction. It's literally one step above brute force search in terms of sophistication

YouTuber: you can't say that AI isn't smart when it solved a math problem

holy shit it guessed it's way through with a proof checker (LEAN) telling it "no, try again" an unlimited number of times

(this is known as directed search in The Literature and it's decades old)

Imagine if you got that treatment. Imagine taking an math exam you get infinite tries at. Would the fact that you eventually got one of questions right mean you're good at math? I want to actually claw my eyes out

(edit : I've been informed that this may not be entirely accurate, especially with respect to the level of LEAN involvement in the early stages, but it's hard to tell because OpenAI/Anthropic are very secretive)