TechTakes

FAR AI tries to imply that the Go bots are still not merely superhuman, but far superhuman: “this result demonstrates that even far superhuman AI systems can fail catastrophically in surprising ways.” Uh huh. [FAR AI]

fuck it’s so disappointing that everything in this space has to be communicated through the cracked lens of critihype — that even the utterly normal failings of a misengineered system must be misrepresented, in true techfash style, as further proof that the system is powerful and is only one more breakthrough away from perfection

The usual AI pumpers have suggested the bots are superhuman!

They are technically correct. The AI is superhuman when it comes to playing Go, by most measures. I don't know about "far" superhuman - the usual type of machine wins against professional human players every single time if the humans are trying to play well, but it's not as if it's off in another dimension playing moves we could never possibly comprehend. Many strong Go players probably still disagree with me there, but it is at least not as unlikely as some assumed at first that we can understand what it's doing. Its moves can usually be analysed and understood with enough effort, and where they can't the difference measured in points won or lost in the game is often small. Their main advantage is being inhumanly precise, never making the kind of small errors in judgement that humans always do. Over the course of a lengthy game of Go that gradually adds up to an impressively large margin of victory.

Katago is not an artificial general intelligence. It is a Go-playing intelligence. And this class of flaw that's been found in it is due to the particular algorithm it uses (essentially the same one as AlphaGo.) It lacks basic human common sense, having found no need or ability to develop that in its training. Where humans playing the game can easily count how much space a group has and act accordingly, the program has only its strict Monte Carlo-based way of viewing the game and has no access to such basic general-purpose tools of reasoning. It can only consider one move at a time, and this lets it down in carefully constructed situations that do not normally occur in human play, since humans wouldn't fall for something so stupid.

Its failing is much more narrow than those of the LLM chatbots that everyone loves so much, but not so different in character. The machines are super-humanly good at the things they're good at. That's not too surprising; so is a forklift. When their algorithms fail them, in situations that to naive humans appear very similar to what they're good at, they're not. When it works, it's super-human in many ways. When it goes wrong, it's often wrong in ways that seem obviously stupid.

I suspect that this problem the machines have with Go playing would be an excellent example for the researchers to work with, since it's relatively easy to understand approximately why the machines are going wrong and what sort of thing would be required to fix it; and yet it's very difficult to actually solve the problem general way through the purely independent training that was the great achievement of AlphaGo Zero rather than giving up and hard-coding a fix for this one thing specifically. With the much more numerous and difficult failure modes they have to work with, the LLM people lately seem busy hacking together crude and imperfect fixes for one thing at a time. Maybe if some of them have time to take a break from that, they could learn something from the game of Go.

Wish they hadn't figured out that the 3x3 invasion works well enough, though. Shit is annoying and happens like every game now

I don’t see any of them play an if err != nil { return err } so they can’t be all that smart now, can they

Is AI rally not capable of "miscounting" the score or at least furiously flipping the game board or table??

Sadly somebody trained them on the wrong dataset

@dgerard