Hacker News Reader

Comment by ganeshkrishnan | Hacker News Reader

It's really amazing and impressive what LC0 has achieved with almost no money and only community pooled training matches. They were struggling to pay $600 per month for the cloud server so have a baremetal server.

Pretty sure it can beat AlphaZero at this point of time.

View on HN

It's insane. Basically a hobby project and it obliterates other engines which are the result of 30+ years of computer chess research.

I feel the opposite way: it's insane that Stockfish, running on CPUs and crude handmade heuristics, is able to keep up at all.

Sure, Stockfish is a very impressive piece of technology, and I'm not trying to take away from that.

The way I see it, and what stands out as really cool to me is that we seem to be living through a phase transition to the third major era of chess, based on what type of player was strongest:

500 AD - 1997 AD: Humans.

1997 AD - 2019 AD: "Classical" minmax computer chess engines.

2019 AD - ???: Reinforcement learning engines.

It's extremely exciting to witness.

Just like the invention of the car didn't make horse races obsolete, I wonder if we'll see high-level computer chess competition split into two categories, one where using pre-trained models is not allowed.

Kasparov complained that Deep Blue had access to a myriad of his historical games, so it could prepare purposely just against one opponent. At the same time, Kasparov hadn't any access to train against Deep Blue's way of playing.

Stockfish had a similar disadvantage against Alphazero. With LC0 being more openly available, it can start improving its heuristics against this type of opponents.

It is called zero because it only play with itself. Follow alpha go zero and leeks zero.

It has not studied human play. Unlike alpha go.

I agree with you in general, but to be perfectly fair, even the Zero-style AIs have a few hyperparameters such as block size and CPUCT value etc that can be tuned to produce engines that play differently. And if Stockfish is the measuring stick by which you gauge progress, you'll tune those values to produce good anti-stockfish nets.

That said, I think the effect here is probably negligible.

Except that LC0 wins decisively against all other engines too. You just hear about the wins vs. Stockfish because Stockfish is the best conventional engine so that's more newsworthy.

Sure, but that doesn't really change my point, only the dates involved. You can debate how fair the Kasparov - Deep Blue match was, but even being maximally generous, by the mid 2000s it was completely impossible for a human to beat a computer in a serious, fair match.

> it's insane that Stockfish, running on CPUs and crude handmade heuristics, is able to keep up at all.

Seems like another nail in the coffin for symbolic AI.

Only if you think skill in chess translates to actual intelligence. As a chess master, I tend to benefit from this myth, but it’s bogus.

The idea that computer chess is a stepping stone to common sense AGI is also fairly absurd to me.

Don’t get me wrong. Leela is a great project. Just see it for it is, an impressive demo showing how machine learning algs paired with massive compute has led to increased performance in a perfect-information game like chess. It’s not a “nail in the coffin” of anything other than possibly Stockfish’s reign as strongest chess player. No more, no less.

Pretty sure Alpha zero would destroy Leela.

Maybe so, but Deepmind isn’t training AlphaZero chess any further! Leela will hopefully just keep on getting better and better :)

Eh, If they want, they can train Alpha zero in one day Leela can be trained in 10 years. In any case indeed Leela's progress is impressive.

Training AlphaZero for more than 4 hours on DeepMind's system didn't improve its strength. Makes you wonder if that's as close to ideal chess as we're ever going to get...

No, I don't think so. Remember that AlphaZero played stockfish 8 which is pretty outdated. Some of it's moves against SF was pretty amazing and you could be right but everything is speculative right now.

My only vote for LC0 is that it can win against the latest iteration of SF

Stockfish 8 evaluation was in 2017. In 2018, DeepMind re-did everything with Stockfish 9 and nothing changed.

https://deepmind.com/blog/alphazero-shedding-new-light-grand...

I remember an annonce saying that the go branch of leela-zero had finaly reachen the level of facebook's network. It is arguably not Alpha zero but it shows that, if they have not reached the level of Alpha-zero yet, it is only a matter of time.

I believe AlphaZero's CNN uses 40 residual layers and LC0 uses 20 layers. So in theory, AlphaZero could have twice as much chess knowledge. So that's a reason to believe that AlphaZero might be stronger than LC0. But there's no real way to know at this point.

Reply on news.ycombinator.com