[asparagui]

I would argue that the Alpha Zero algorithm does well at learning chess via self play to get to a master level, but the high level chess engines are super-optimized for their specific domain problem and so will remain out of reach.

[thom]

Leela and AlphaZero are already comfortable beyond master level and into the 3200s in ELO terms, at least. As for the best engines today (who I've no doubt we'll see win out at TCEC - Wasp has already been holding its own in division 4) I think they're super optimized on one axis - search, but not optimized very much (in the grand scheme of things) on the other - evaluation.

AlphaZero and LCZero show have orders of magnitude better evaluation is available and can go toe to toe with existing engines despite radically fewer nodes searched. I wouldn't be surprised however, after this current generation of NNs get good, if the traiditional Alpha-Beta engines can't encode _some_ of the more subtle positional knowledge they've picked up, in a way that works quickly with their current architectures.

[thom]

And I expect the flip side is also true - radically better hardware and algorithms for both training and executing neural networks. Even after that, I think there’s a huge gap in the market for a system that not only sees tactics deeply and had an almost flawless positional sense, but can explain and train humans in what it sees.