Hacker News Reader

Comment by jsheard | Hacker News Reader

> Just look to Chess. The top players today are way better than any of the greats before, because they can train against computers and know exactly where they failed.

AlphaGo isn't available for anyone to train against like Stockfish is though, what are Go players using? Has another powerful Go engine been developed since then?

View on HN

KataGo is an open-source algorithm derived from AlphaGo, but with a number of tricks so that it trained faster: https://katagotraining.org/

It likely surpasses AlphaGo, and just like Stockfish, it delivers a protocol that can hook into many user interface apps: https://github.com/lightvector/KataGo?tab=readme-ov-file#gui...

From those technologies, also came an interesting visualisation of how human players changed their habits following AlphaGo: https://drive.google.com/file/d/16-ntvk3D1_pgjJ7u64t4jMYMh0z...

> It likely surpasses AlphaGo

If I remember correctly, a recent estimation is around +1000 Elo compared to AlphaGo.

That also matches my estimation which places AlphaGo Zero 40-blocks 40-days at 5185, and KataGo at 6262: https://thefiletree.com/espadrine/%E2%9A%92/elo.html

(Details of the estimation in the source code.)

Lovely! That's a nice summary.

We use KataGo and sometimes LeelaZero (which is a replication of the AlphZero paper). KataGo was trained with more knowledge of the game (feature engineering and loss engineering), so it trained faster. It was also trained on different board sizes and to play to get a good result when it's already behind or ahead.

KaTrain is a good frontend.

> KataGo was trained with more knowledge of the game (feature engineering and loss engineering), so it trained faster.

Not really important to your point, but it's not really just that it uses more game knowledge. Mostly it's that a small but dedicated community (especially lightvector) worked hard to build on what AlphaGo and LeelaZero did.

Lightvector is a genius and put a lot of effort into KataGo. It wasn't just add some game knowledge and that's it. https://github.com/lightvector/KataGo?tab=readme-ov-file#tra... has a bunch of info if you're interested.

I wasn't at all trying to say his work was simple. I was trying to say "deepmind were trying to build an AI that gets good at games without anything in their structures being specialized for the game, lightvector asked what if we did specialize the model on Go". And he did some wonderfully clever things.

Reply on news.ycombinator.com