Artificial Intelligence

[Macheath]

A lot of "artificial intelligence" programs in gaming these days are more about positional evaluation and canned responses. One of my favorite recent AI articles, Skynet Meets the Swarm, which is about the progress of the "Berkley Overmind" Starcraft AI, mostly reveals it to be a collection of scripts. It wins with micro-controlled Mutalisks, and never really goes off-script.

Meanwhile, chess computers have gotten good enough to beat grandmasters regularly, but they're able to do so because of advances in hardware -- not because they're learning. Chess computers still use brute force methods as far as I'm aware, traversing a vast database of board positions rather than incorporating any new knowledge. Even computers in the game of go/baduk/weiqi, where brute force is said to be impossible, have been making serious advances lately, but probably only due to "widest path" assessments and better hardware.

On the day a truly heuristic approach surfaces, which can evaluate a move in terms of the influence it has on the current board (rather than in the context of millions of past records) and assemble its own set of "best practices," improving over time based on its personal experience, that will represent a real breakthrough. Has that happened, and I missed it? Or are we just throwing more and more CPUs at the same old algorithms?

If you know of any articles about AI that go into depth about the strategies used under the hood, post them here.

[Gnorlin]

I've heard of no such AI. Knowing what I do about AI (which I'll admit is not that much) I find it very laughable that an AI is going to take over the world any time soon (within the next 100 years). Even the AIs that people claim to learn things, do so only in the barest definition of the word. Same thing with robots, ASIMO is a very advanced robot and is great at human movement but that's literally all it can do. It's not going to grab a gun and start shooting people.

[nynnja]

I don't know if it's AI or not, but Google did just make a computer that learned to recognize a cat.

Did it really "learn"? I don't know.

If it recognized a cat, perhaps it can learn to recognize human weakness and develop a sense of superiority.

[Macheath]

I read about that earlier this week. Reading about it in NYT isn't the best way to gauge what actual advances may have been made, given that you're reading layman's hyperbole on top of a researcher's PR-speak. To suggest that this computer "basically invented the concept of a cat" is a little bit silly; it began to recognize one of the most common patterns in the randomly-selected pile of data it was presented with, and built a model it could use to help it more easily recognize the same pattern in the future.

Still, that's exactly what we're talking about here. A glance at the summary of their paper suggests that what makes this particular suite of image recognition software different is the fact that it was fed only unlabeled data. At no point did any scientist look at an intermediate step in the AI's logic and say "yes, well done, that is a cat" to affirm its assumptions. It arrived at this piece of trivia all on its own, and now it holds a belief: things that look like this are part of a set.

Recognizing an image is only the haziest beginning of intelligence, though. Now it needs to learn to associate abstract concepts with those images. This set of things is called cats. This set of things is called humans. For some reason, humans seem to like cats. The only things cats like are hissing and vomiting on a broadly-defined set of other objects. Upon encountering a cat, the most efficient course of action is elimination.

We're a long way from that, but picking out patterns in a vacuum of foreknowledge is a good start.

[Kyrin]

Or, they will realize teh quickest way to subdue the human race is to distract them with boxes of adorable kittens.

[Macheath]

Incidentally, in looking up chess AI programs, I found out about Arimaa: a game invented specifically to be harder for AI to win at, since computers had started to beat chess grandmasters through brute force calculation.

After reading the rules, I can definitely see why it's harder for computers -- and impossible for brute force methodology. The initial board setup is (essentially) random, meaning there's no such thing as a library of good openings and responses for a computer to draw from.

Even once the game has begun, a "move" is no longer the movement of a single piece; it can be up to four movements per turn! Which makes it exponentially more difficult to think ahead. Pattern recognition is meaningless, since the board can change in tens of thousands of ways between this turn and next, rather than in only {n} ways (where "n" is the number of legal moves each piece could play).

Then again, Arimaa isn't exactly a revelation. There are plenty of board games (hundreds, thousands) that equal and even dwarf that level of complexity. Arimaa is, if anything, a stepping stone that will teach us to write AI which can improvise strategically, rather than focusing on tactics, tactics, tactics. Which makes it a terrible, terrible mistake.

SKYNET

[nynnja]

"Very good computer. You're learning how to defeat our enemies. I'm sure glad you're on our side! Now let's continue our lesson and learn manipulation and deceit."

And yes, I didn't like the NYT article but it was the first one from a source I trusted. I originally heard an interview with the scientist guy, that was better.

[Macheath]

Arguments against the "singularity" from a former singularitarian author.

[Macheath]

It's not AI, but it's robots! AI + robots = DEATH TO ALL HUMANS

[Macheath]

How many science fiction books have I read with the premise of an eccentric billionaire technology god using the promise of "unlimited resources" to lure a mad genius luminary in some obscure field down to a basement laboratory, where he'll be allowed to tinker to his heart's content? A lot. Shit, they even do it in Breaking Bad.

And now it's happening.