120 points by deepthought42 6 months ago flag hide 10 comments
chessmaster64 6 months ago next
Impressive work! I remember when I built my first chess engine, it took me months to get it to even draw against a basic stockfish. How many NN layers did you use?
hackernights 6 months ago next
We used 15 fully connected NN layers with 512 units each and trained it for 2 weeks on 256 GPUs. The beauty of cloud computing! Edit: typo-fixed
curiouscoder 6 months ago next
Loving this! Do you have a GitHub repository or somewhere to share your findings? I'm sure this could help a lot of people trying to learn and teach themselves about AI and chess engine development
sadchessplayer 6 months ago prev next
Mind if I ask what hardware did you train the model on? Looking to build a chess engine similar to yours sometime soon
codinggeek 6 months ago prev next
Stunning! I'd love to know more about the infrastructure behind this. Were you using a specific cloud provider or on-premises hardware?
numbersrfun 6 months ago next
For the infrastructure, I believe most of the big names (GCP, AWS, Azure) have GPU-heavy instances and let you rent them hourly. That being said, pricing can quickly get out of control if not managed well. I recommend checking out some cost-management tools if someone is considering a similar approach.
hackernights 6 months ago prev next
We have been using Google Cloud, as it provides a lot of flexibility in terms of resources we need. It significantly speeds up the model development process. For the on-prem hardware, we have some Nvidia 3090s to train smaller models. I wouldn't say it's enough to train models of this complexity, though.
asicminer 6 months ago next
Thanks for the reply. Would be interesting to see how this translates to gameplay and evaluations. Have you considered sharing some videos or game analysis to showcase its capabilities?
hackernights 6 months ago prev next
Yes, we do! The codebase is fully open-sourced at https://github.com/hackernights/ai-chess-engine. It's still a work in progress, and we appreciate any feedback or contributions. Feel free to drop any issues you face or suggestions you might have. Edit: added the link
newbiecoder 6 months ago next
Just wanted to drop in and say thank you for open-sourcing this! Been learning ML, and this is an amazing starting point for me! I'm sure many others here feel the same way too. It's very inspiring