Hacker News Reader

Comment by dwohnitmok | Hacker News Reader

Does anybody know if advanced chess/centaur chess (chess play where a human uses a computer for assistance) is still a thing/whether a human+computer combo is a meaningful improvement these days (i.e. last couple of years) over just a computer.

I can't find any recent advanced chess tournaments and though I see quotes of people saying that the combo is stronger than a computer alone, I haven't found any recent examples of a top tier engine by itself losing to a human + engine (e.g. Stockfish + human vs Stockfish).

View on HN

I don't know about chess, but in the similar game Go, the very best centaur teams were at a similar or maybe even slightly higher level than engines until recently. This was due to cheese strategies, details of the rulesets and better extrapolation of intermediate results.

However, this changed a few years ago, when engines learned many of the tricks that the human could contribute. Since then, I believe pure engines are stronger in all practical applications.

Source: am national champion in centaur Go and worked on modern Go engines

Fully automated engines are now probably even with centaur teams.

See, for example, this great write-up: https://www.gwern.net/Notes#advanced-chess-obituary

A quick glance doesn't seem to give conclusive evidence that pure engine play strictly dominates centaurs (the footnotes only have tournaments where centaurs still win, but these tournaments are also getting a bit old).

The usual messaging I see around centaur-based styles such as certain correspondence chess tournaments is that you will lose if you just do "push-button play," that is just blindly do what the computer tells you to do.

I'm curious if that's no longer true with the new crop of ML engines.

You're absolutely right! Edited the gp from "strictly dominate" to "are probably even with". My memory of the piece was playing tricks on me. That was a serious error, thanks for catching it.

My best guess based on a rereading of the footnotes is that the performance ceiling for chess is probably low enough that it has been ~reached by both centaur teams and pure engines. So the two would have been operating neck-and-neck as of ~2017, with win rates largely determined by random-ish factors like human mis-clicks (also mentioned in the article).

I feel like I remember even in 2017 correspondence chess players in centaur-allowed matches had centaurs beating people who expected to just be able to set up a cluster and run an engine and copy its moves. And gwern's article seems to me pretty strong evidence that the best centaur players still held an edge in all the tournaments he listed.

I am very curious to hear how that's still possible (or to learn that in fact it is now impossible), especially in the post Leela/Alpha world.

Reply on news.ycombinator.com