AlphaZero Ideas

That is a new and exciting paper from Julio González-Díaz and Ignacio Palacios-Huerta, here is the abstract:

Can artificial intelligence (AI) uncover new ideas? As machines are learning fast and becoming increasingly intelligent, can AI not only automate the production function of goods and services, but also of ideas? Economic growth arises from people creating ideas, and thus an affirmative answer to these questions may have drastic implications for a host of important issues. Yet, to date, there is no empirical evidence showing that AI can in fact generate tabula rasa ideas that improve human understanding. Using as an exogenous shock the introduction of AlphaZero, we provide the first causal evidence of the impact of unsupervised AI on the production function of ideas. Specifically, AlphaZero is considered a milestone of scientific progress in AI research. This program rediscovered ideas known in centuries of human chess, and created new ideas as well. We study world experts at the frontier of knowledge and find that at least the player with the highest classical rating in the history of chess learned and adopted new ideas uncovered by AlphaZero. Other players may have also done the same. We contend that obtaining evidence of the impact of AI on the production function of ideas is a necessary first step to think about AI’s impact on the innovation and research processes that drive the advancement of knowledge and economic growth.

The main new ideas I have seen come from AlphaZero are the following:

1. Pushing the h pawn is often better than you thought! (emphasized by the authors)

2. Said pawn can be worth more on h6 (h3), as an aggressive weapon, than you might have thought.

3. Qa1 (a8) is occasionally a better move than it looks.

4. The chess openings that were preferred in the early 20th century, such as the Queen’s Gambit, are in fact pretty lindy and pretty good.  You can debate whether that is a “new” idea, but it is a meaningful revision of sorts.  (Of course plenty of earlier patzers had some fondness for #1-3, one might add, though perhaps not for the right reasons.)

So that is something.  But I think in terms of a percentage of the total improvement in play, it is quite small.  “Finding more good players through the internet” would come in first by a long mile.  “Giving more players more time on convenient services such as chess.com” likely would be next.  Even “just having good players improve their endgame play using basic study and standard chess engines” would be much larger than these AlphaZero effects.  “More top players copying the physical training regimen of Magnus” would be more significant as well.

I am also skeptical of the claim that very much of Carlsen’s 2019 improvement (he didn’t lose a game that year) came from AlphaZero.  He has lost some games since then!  And it is not as if all of his subsequent opponents are zapping him with surprise “Qa1” moves.  I see AlphaZero as a series of innovative but ultimately modest advances that have been incorporated by some of the top players with barely noticeable overall gains in move quality.

I am not an AI skeptic, and furthermore I see special value per se in “advancing the frontier,” even when various infra-marginal gains (“swim more!”) are more significant in quantitative terms.  Still, my estimate of the chess innovative advances from AlphaZero are more modest than what this paper would seem to suggest.