Hacker News Reader

Comment by ks2048 | Hacker News Reader

ks2048·

I read something about Go - that very unusual (maybe even considered bad) playing could beat the super-AIs. They are so tuned to opponents in a "typical" style, that they don't know how to beat a player outside this distribution.

Maybe an NBA team will come up with something like that.

View on HN

I'd be interested in seeing a link to this. Decades ago, playing defensive "computer chess" used to be a relatively optimal strategy against Chess AIs.

However, I believe Kasparov famously tried to employ this tactic against Deep Blue but by that time it wasn't particularly viable.

I'm not sure if this is the study I was remembering, but here is one link I found: https://far.ai/post/2023-07-superhuman-go-ais/

I can see it in the case of champion-trained AI, but in the case of Alpha Zero trained AI, it should have encountered virtually every kind of senseless play.

I can't speak for go, but I've seen beginners get an edge in other strategy games by playing a very unusual strategy. The trouble is, it doesn't tend to last - it turns out either that their strategy really is weak, or that it's viable but they don't have the ability to follow up on the early edge they gained. I can imagine a similar thing might happen with sports strategy.

That can't be true, especially not in a territorial game like Go. Could you find any link to that?

The sister comment has a link. Maybe I shouldn’t say it can beat AlphaGo, but rather you can set up adversarial states that it is bad at.

i forgot what year but the year the spurs won the championship against lebron would be unusual today. tons of passing not necessarily for the 3 but to just dislodge the defense enough for a guaranteed bucket

Reply on news.ycombinator.com