Transcript:
Elon: The actual number of moves that are not utterly stupid in chess is tiny and chess will be fully solved one day
Chess.com: Skill issue
Elon: Chess is hard for humans, but not for computers. One day, it will be fully solved like checkers. Between now and then, or even after, have fun. People enjoy may games/sports where machines are vastly better.
Chess.com: There are more possible chess games, by about 40 order of magnitude, than atoms in the observable universe. To solve chess, you’d have to map them all. Good luck.
Elon: Wrong, all legal chess positions are ~40 ooms less than the estimated atoms in the Universe. And ASI will figure out ways to solve and compress that are far beyond what we could possibly comprehend. (Here I am arguing with some random intern at a chess website)
Chess.com: I’m a full-time employee
Elon: (something-emoji)
Chess.com: Even my friends don’t text me back that fast
And Elon is reading the correct CS/math blogs by pointing out that only legal moves need to be considered.
If he was a little more on his toes, he might point out that Go has a vastly larger state-space than chess, yet it was one of the first problems that fell to GenAI, which is able to find crazy-complex high-dimensional heuristics/“intuition” navigating such spaces.
Naw it wasn’t genAI, it was the deep learning stuff that came before, when ML was still cool.
God damn I feel that in my bones.
It was one of the first instances of genAI. The team at Google figured out that you take a classifier, flip it round to make a generator, then set it against the classifier to train itself
GAN have been around way before…
Except, Go isn’t a solved game. Which is what we’re taking about here.
Elon doesn’t read anything except social media posts.
Depends what you mean by “solved”
Mapped the entire state space so next move is just a lookup in a huge table? No, Universe not big enough.
Game is no longer generally winnable by humans no matter their level of skill? Yes, that happened.
Moving the goalposts.
What I mean is simply the definition of a solved game. From any position, the outcome of the game can be correctly predicted assuming both players play perfect moves from then on.
You’ll note that Go is not on the list, but is down under ‘partially solved’ as it has weak solves (solved under limited conditions) only at a 5x5 and 7x7 board in 2015.
Humans play on a 19x19 board, which is 145 orders of magnitude above 7x7. It’ll not likely be solved by an LLM until it has storage space, RAM, and processing speed several-dozen orders of magnitude larger than all that which currently exists on Earth.
I wouldn’t say Go “fell to GenAI”
Go grandmaster Shin defeats AI KataGo in historic human victory
As the article shows, the Pros don’t treat AI the same as a human player and often don’t attack its weaknesses and instead try to find more interesting ways to defeat it and learn more about the game.
Although it should be noted that Shin Jin-seo did start with a 2 stone handicap. But this also alters the AI to play more risky moves it wouldn’t normally.
One thing that is super cool about Go is the ability to add a handicap.
A handful of prodigies are still able to sometimes beat the AI.
I count that as solved, the same way a few crumbs on my plate means I have “finished” my meal