> But ultimately, Seltz-Axmacher believes, the tools he’s developing will be good for truckers. He cites a new book by Garry Kasparov, Deep Thinking, in which the chess great observes that middling chess players who play with the help of a standard computer are reliably better than either grandmasters or supercomputers by themselves. “I think humans and technology working together are always going to be better than either one alone,” Seltz-Axmacher says. “But maybe that’s just because I like humans.”
This is amusing because this stopped being true a long time ago. Even by 2007, it was hard for anyone to improve, and after 2013 or so, the very best centaurs were reduced to basically just opening book preparation (itself an extremely difficult skill involving compiling millions of games and carefully tuning against the weakness of possible opponent engines), to the point where official matches have mostly stopped (making it hard to identify the exact point at which centaur ceased to be a thing at all).
I've never really understood Kasparov's pro-biological opinion along those lines. I suspect it's a combination of an observation that was true at that time (that humans are better are intuition), and his own comment “But maybe that’s just because I like humans”.
If you think about it, AIs have a significant advantage in that their evolution is completely artificially guided. They can be focused on tasks in isolation, and the results can be observed and improved upon. Biological evolution does not have this luxury, overall the entire organism has to "work", and survive long enough to reproduce. Sure, current technology is in its infancy, but extrapolating into the far future, things are interesting.
I think it's a mix of outdated information, wishful thinking, and the contrarian appeal. It was true for a few years that a grandmaster or a random chess player with good 'mechanical sympathy' could crush a solo chess engine (which is why discussions of this tend to stop in 2007 despite being a decade ago), it would be very convenient if AIs required humans for the best performance (https://github.com/JackToaster/Reassuring-Parable-Generator/... 'A computer will never play the best chess in the world'), and it provides a meta-contrarian (http://lesswrong.com/lw/2pv/intellectual_hipsters_and_metaco...) take on AI risk/technological unemployment which is clearly Tyler Cowen's main motivation in popularizing it (object: 'AIs will never be able to do X better than humans as they have no soul'; contrarian: 'they totally will, just look'; meta-contrarian: 'ah, but the human element is still critical for a last bit of performance due to the creativity and ingenuity of the human spirit!').
It's also interesting in that it demonstrates that there's at least two levels of 'superhuman' performance in chess and other domains: there's superhuman in being able to beat any living human one on one (Deep Blue, Sedol AlphaGo), and superhuman in being unable to be improved upon by a human collaborator (post-2013 chess engines; AlphaGo Master?). I'm not sure this gap was really expected, but though it only lasts a few years (or months), it's still interesting to note.
This is amusing because this stopped being true a long time ago. Even by 2007, it was hard for anyone to improve, and after 2013 or so, the very best centaurs were reduced to basically just opening book preparation (itself an extremely difficult skill involving compiling millions of games and carefully tuning against the weakness of possible opponent engines), to the point where official matches have mostly stopped (making it hard to identify the exact point at which centaur ceased to be a thing at all).