[fferen]

Something I wonder: are computers better at chess because they can think faster? What if we give a team of top grandmasters 1 day per move, versus like 1 hour for Stockfish. Will they at least consistently draw? To my knowledge, top human vs computer matches these days only use piece odds, not time odds.

[317070]

No, on long time controls, humans cannot see enough of the tree in some positions. They might be able to push the game sometimes into a position where they can hold if the engine is not set up to prefer complicated positions.

But

This is one of my favorites. On extreme short time controls, humans can handle some really good engines some of the time! Here you have Andrew Tang beating Leela in hyperbullet: https://youtu.be/Wf-wFXRpwgo

Another approach that sometimes works is to be really booked up in dubious lines with an engine that does not have too much time, something Jonathan Schrantz manages from time to time https://youtu.be/FC2P6VUYu78

[gjm11]

It's a little debatable whether Andrew Tang counts as "human" when playing hyperbullet chess.

[NewEntryHN]

"Classical" chess engine (those based on alpha-beta search) are indeed better because they calculate faster, although so insanely faster that there is probably not enough time in a _lifetime_ for a human to win a single game against it.

The more recent chess engines (LeelaChess, and the latest versions of Stockfish) use neural nets to allow for more shallow calculations without decreasing game quality. In this case the engine is better because it's just "more intelligent" (and nobody understands how).

[amalcon]

An easier way to do this would be to limit the hardware resources available to Stockfish. You can get the 24x time ratio just by giving Stockfish 1/24th of the compute cycles.

That said, 24x is not even close to enough advantage there -- that probably removes less than one ply of depth. Stockfish can read far deeper than any human; you'd need a dramatically larger reduction to compensate for this.

You could probably put Stockfish on a desktop computer from the mid 90's, and expect the team of GMs to win.

[gjm11]

24x is more than 1 ply. Typical branching factor isn't so far from 24, and with alpha-beta the effective branching factor is more like sqrt(nominal branching factor), and I think modern engines are pretty good at pruning and extending the tree to make the scaling even better than that.