On Thu, Dec 26, 2024 at 11:02 PM PGC <[email protected]> wrote:
> *> I’d note first that your analogy with chess and engine moves ignores an > asymmetry: a grandmaster can often sense the “non-human” quality of an > engine’s play,* > * Superhuman would be a better word for that than nonhuman. * > *> whereas engines are not equipped to detect distinctly human patterns > unless explicitly trained for that. That’s why a GM plus an engine can > typically spot purely artificial play in a way the engine itself cannot > reciprocate.* > *Nope. When AlphaGo beat Lee Sedol, the best human GO player in the world, its winning move was move 37. It was such an unusual move that many GO experts at first thought it was a huge blunder, even the people who wrote AlphaGo were worried so they check their readouts because whenever AlphaGo makes a move it automatically estimated the likelihood a human would expect it; and they found that AlphaGo thought there was only one chance in 10,000 of a human making such a move. Today it is generally agreed among GO experts that move 37 was one of if not the most brilliant and creative move in the entire history of the game. AlphaGo knew it was making what you would call a nonhuman move and that everybody else would call a superhuman move. * > Even then, the very subtleties grandmasters notice—i*ntuitive > plausibility, certain psychological hallmarks—are not easily reduced to a > static dataset of “human moves.”* > *AlphaZero is better at Chess and GO than AlphaGo (and better at any two player zero sum game) and it contains NO dataset of human moves, static or otherwise. NOR DOES IT NEED ONE.* *> **Chess self-play is btw trivial to scale because it operates in a > closed domain with clear rules.* > *There are clear rules about what moves in Chess and GO are legal, but there are no clear rules about what moves are good.* > *> You can’t replicate that level of synthetic data generation in, say, > urban traffic,* > *The question is moot. Tesla has about 4 million cars on the road and it has been collecting data from them since 2015, so by now it has billions of hours of real traffic data. * *> nuclear power plants, surgery, or modern warfare.* > *Both humans and AIs find nuclear reactor simulators and war games to be very useful, but I grant you that at least right now humans have more experience with surgery than AIs. Today surgery and nursing care are the only areas of medicine in which humans still have an edge over machines. We already know for a fact that O1 Preview is far better than human doctors at diagnosis, I can only imagine how good O3 will be. * *> The bait was luring you to clarify your stance that “a problem is solved > or it isn’t” while simultaneously implying that AI failure is more > tolerable than human failure. That is self-contradictory.* > *Yes, t**hat certainly would be self-contradictory IF I had said that an AI error is less serious than a human error, BUT I did not. I did not imply it either although you may have inferred that I did when I said that AIs are constantly getting smarter and thus are constantly producing fewer errors, but human beings are not getting smarter. * *John K Clark See what's on my new list at Extropolis <https://groups.google.com/g/extropolis>* ngs -- You received this message because you are subscribed to the Google Groups "Everything List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/everything-list/CAJPayv0VZR90jMb8Xf0PGRMdo-atGqeQaubbJi5nV3b21KikRw%40mail.gmail.com.
