Hacker News new | past | comments | ask | show | jobs | submit
I don't think the "centaur" approach will go away any time soon. People like to point out chess as an example where humans are pretty much liabilities for chess engines - but chess is a much simpler game, with only a few possible actions you can take and a fairly limited board ;)

In programming you've got a lot of decisions to take, many aren't verifiable and even more have tradeoffs so you can't hillclimb your way out of architecture design and requirements.

Most of what people call AI slop is literally making AI do design decisions. AI is much faster at executing than you are but making wrong decisions 20x as fast doesn't get you to fulfilling your requirements any faster.

{"deleted":true,"id":50554127,"parent":50548933,"time":1795046614,"type":"comment"}
> Most of what people call AI slop is literally making AI do design decisions.

I dunno about "most"; the memetic monoculture is also a big part of it, even when it comes up with something that seems novel the first time you encounter it, it's a cliché almost immediately.

While clear patterns in software is not itself a bad thing, this behaviour applies equally to some of the AI's failure modes, and those become instantly-recognisable annoyances.

From a search, because i've been out of the chess scene for a bit now: apparently centaur chess was really big until 2020 or so, but it's fallen behind pure chess machines now. Bummer.

From another search, it looks like go leapfrogged from ai-sucks to ai-is-unbeatable and didn't really get a centaur period.

Please correct me if wrong. I'm disappointed to learn this.

Can anybody clarify the relationship between that adversarial deep learning where it can play games against itself to improve and llms? I figure the clear win conditions in chess and go make it easier to train models, and I'm not very clear on how llms know if they improved or not