Gwern about centaurs: there is no chance that any useful man+machine combination will work together for more than 10 years, as humans soon will be only a liability

post by avturchin · 2018-12-15T21:32:55.180Z · LW · GW · 4 comments

This is a link post for https://www.reddit.com/r/reinforcementlearning/comments/a3rbm3/alphazero_shedding_new_light_on_the_grand_games/

This is a quote from the discussion in Reddit (responding to another commenter):

"I remember and I kind of subscribe to this idea of "man+machines as a future of work". I think Tyler Cowen implanted this idea into my head"

Yes, Cowen was big on that too in stuff like Average is Over.

Very irresponsible of them to try to foster complacency like that. It should have been beyond obvious that there was no reason chess engines wouldn't keep improving and that at some point very quickly, far from representing a new stable paradigm and a reason to not worry about technological unemployment, the 'centaur' would be a net liability. As far as I can tell, in chess, the centaur era lasted barely a decade, and would've been shorter still had anyone been seriously researching computer chess rather than disbanding research after Deep Blue or the centaur tournaments kept running instead of stopping a while ago. In Go, it lasted a year at best (if we assume that world champs like Lee Sedol could spot 'delusions' like made it lose a game to Lee Sedol and contribute at all, but then by the Ke Jie tournament with Master, between Master's performance and the various match settings, it looked like humans were way far behind and liabilities when paired with Master, and even if we doubt that, Zero then came out and superseded Master entirely). Not very comforting precedents... The idea was nice but it doesn't work.

4 comments

Comments sorted by top scores.