commenter-thread

It's been years since I drove in downtown Boston and I still get the nervous twitches thinking about it. Which is part of the reason I think the initial killer app for self-driving cars is chauffeuring elderly Boomers around the suburbs.

Sometimes I think AGI is "just" a big enough, deep enough hierarchy of game-like things. "Me" is a heuristic goal-setting system shaped by a billion years of ancestors that didn't make fatally bad choices too soon, plus the ability to set up my own deep learning processes, plus the occasional random number*. I don't expect I'll live long enough to see us assemble that much computation and communication in silicon form. But there will be a lot of stuff that we can build and combine limited specialized subsystems for if we want to. A hand may not be the ideal gripping element in any specific situation, but it's damned good as a general-purpose device. And once that control problem is solved well once, in the AI world it's solved everywhere.

* The other day I was going through all of the "random" things that had to happen to keep me from spending most of my life in California. From my Dad's teeth that kept him out of Navy OCS to the order in which I lusted after two different brunette women.

There's now research software that does well at Texas Hold 'Em at a table with the software and five professional players. Granted I'm looking for certain approaches, but in the couple of papers I've read there's hierarchies of games, repeated subgames, guided search... Also, the software "discovers" some of the tactics humans use, like bluffing, and some tactics that humans don't.

Indeed. Although MuZero does learn to play Atari games, some of them quick-twitch games, against real time. The company behind MuZero is focusing on StarCraft for all of the reasons you list. My intuition is that there will be more in common in the low-level plumbing -- eg, hierarchies of "games" and guided search -- than many people think.

Apparently their StarCraft II software has achieved grandmaster status playing against top humans under tournament conditions. One of the humans made the same sort of remark as the Go player I mentioned, that the software isn't playing the game the same way a human does.

With what, a billion years of "development" behind it?

Word Perfect, reveal codes. When I worked for the legislature, we had to use Word Perfect. Responsibility for documents was passed on from one analyst to another, for years. Lots of years. One interim I spent a significant amount of time going through the documents I would be responsible for the next session with reveal codes turned on, stripping out all of the accumulated cruft. What a miserable task.

Yep. And people look at me funny when I say that all four-year degrees should require at least a bit of programming. (Also, writing for people's consumption, but that's an argument for a different day.) We have reached the point where almost all of us will be forced to "code" more than once in our lives.

MuZero was released last year. Starting from scratch it derived the rules for chess, shogi, Go, and 57 Atari games. For Go, it then trained by playing against itself. Beats every other piece of software, and humans. One of the best humans says it's like watching an alien play Go, the style is very different. Of course, the follow-up question I always ask goes something like, "Great! When will it be able to to write a book that teaches a human to play Go?"

The description sounds like the version that only plays Go would fit in one shelf in a rack and consume about a kW of power. (The learning version is much bigger.) As I recall, the chess-playing version of Deep Blue was many racks and drew upwards of a MW.

On an always contentious topic, I still think my wife and I (middle-ish Boomers) will get to have a self-driving car that will at least take us on errands before physical deterioration makes it too dangerous to let us drive ourselves.

ThTh5: Yeah, if we ever get software intelligence, it will almost certainly run on different principles than a human brain.