Hacker News Reader

> The problem is that humans don't play that way, by evaluating a tree a few moves ahead and selecting from that.

As a modest 1400 blitz player, I definitely evaluate future states. That's not all I do, but it's certainly one thing I do. AlphaZero can also run just intuiting the next move without evaluating any additional states also fwiw. Though it is much better when it is allowed to do so.

> Consider the task of picking apples from a tree. You could easily make a machine to do that. But what if you had to make a machine that to an onlooker (who couldn't see the robot's "face", let's say) would appear to be a human picking apples from a tree? That would be much harder.

This is a very deep and nuanced task, made doubly difficult by obscure robotic hardware requirements. Not a great parallel.

> There are many chess AIs on chess.com, specifically designed to play "like" a specific grandmaster or well known chess streamer. I don't think any titled player would not be able to guess they're playing a computer if they played a few games against the AI without being told. It's very well known that computer moves are very different from human moves.

Maybe? A quick google suggested Maia is a close match to what I was suggesting. People are suggesting it does feel like a human in the thread. https://www.reddit.com/r/chess/comments/k4o6z1/introducing_m...

I don’t know that human calculation is so fundamentally different than computer evaluation, it’s just that it’s much, much slower. The human GM’s eval of any given position, without thinking ahead at all, is probably better than most engines. The problem is that strong evaluation isn’t worth much compared to a weaker evaluation that is nevertheless several candidate moves wider and many plies deeper (you might argue AlphaZero disproves this, but Stockfish was regularly beating the neural network engines even before its own neural network, for example).

Humans see fewer candidate moves. They regularly miss quiet moves that are the strongest. They often calculate to a certain depth (even a very shallow one) and stop based on gut instinct. But it’s still just trees of moves with some eval function.

I don’t think it’s fundamentally that hard to mimic, and it would actually be genuinely interesting for didactic purposes. But for fairly obvious reasons it’s not a priority outside of a couple of projects.

I think the proposal is to train neural networks on games played by human players at a given ELO.

Even if it were possible, I don't think there's a way to do it in a way that would make any sense. Would we set up an AI now and then forever use the same one? In that case people will learn any small quirks of the AI and optimize against those vs human play, fundamentally changing the game. If the proposal is to be constantly updating this AI, well even in that case I'd argue that people will over time identify any small difference between machine and human play, but in that case people will also be upset that someone got normed against the "easy" AI, or against the AI that didn't know about the dumdenmorph-joyce-allens countergambit yet, etc.

That kind of engine exists! It's called Maia, as mentioned in a sibling comment.

https://maiachess.com/