[mlurp]

Is this true? From what I've read about Alpha Zero/etc, in both Go and chess, it's going interesting move sequences that people hadn't considered viable before. That certainly seems like an interesting thing to learn.

Also, sure, you can dismiss it all as statistics. But how sure are you that what's happening with humans isn't in some form? I'd also say that MCTS is something people kind of do too in games: look a few moves deep and try to judge the value of that position, which is definitely more interesting than simple RL/bookkeeping/stats.

[dfan]

No, it isn't. The top humans are much better at Go than they were four years ago, largely due to learning from the new engines. If it were all just about sampling the phase space zillions of times, this would not be the case.

[breatheoften]

This is interesting to me ... how exactly is this assessed ...?

Do people keep around versions of “alpha go year 2017” and play against it in order to measure human improvement over time?

If the basis for observing improvement has become “I can beat old versions of the ai more reliably than I used to be able to” or if I have learned to beat players who have not studied alpha zero I suppose that’s a form of usefully learning “about go” by analyzing the games played by alpha zero ...

I wonder if we might ever arrive at a point where human vs fixed-year x ai performance at go pretty much stops increasing over time ...?

[dfan]

I admit that I do not have a quantitative measure to support my claim (as you note, constructing one is difficult). But qualitatively:

1) People have learned a lot from new engines: joseki (corner patterns), general strategy (e.g., moves on the side are now considered less valuable, making large moyos (largely empty space loosely surrounded by your stones) is less attractive because AIs have demonstrated that they're more invadable than previously thought) and are able to actually explain the new principles in human terms.

2) All go professionals now play in the new style, to some degree; ones who tried to continue in the old pre-AI style performed badly.

So I am comfortable claiming that human play has improved by learning from the new engines.

[selestify]

Is there a name for the new style?

[dfan]

I'm not aware of a single "official" name for it that everyone uses, like the Hypermodern style of chess in the early 20th century. In English, people say things "AI-inspired" or "AlphaGo style" (although a lot of the ideas come not from AlphaGo directly but from the public engines that followed in its wake).

[Madmallard]

The sequences are viable because alpha go assessed state much deeper than humans do. Doesn't mean humans will be able to utilize them correctly.

[tfha]

Nah at least in chess all the grandmasters lean heavily on the chess engines, including studying top games between the best AI

[protomikron]

> Nah at least in chess all the grandmasters lean heavily on the chess engines, including studying top games between the best AI

They may use it for training and analysis, but they don't play their style - they play an inferior style (not to dismiss their achievements).

[dfan]

I am not sure what you mean by "style", but top chess grandmasters certainly play in a different way today compared to, say, 20 years ago. Current play is much more concrete (less based on general strategic principles) and players are willing to take greater risks for rewards such as material gains, since AIs have shown that there are often many more defensive resources than was previously thought as long as the defender remains tenacious.