If you use modern engineering techniques, I believe you can meaningfully exceed the quality of the best chess engines currently in existence. I believe I am capable of doing this, and have evidence to show this. This task becomes much easier with more compute, and so I am looking for compute.
Chess engines are programs that play chess. Chess engines have been vastly superhuman since 1997. The best chess engine currently, Stockfish, is something like 800 elo better than Magnus Carlsen, the best human chess player, roughly equivalent to the gap between Carlsen and a strong club player. The second best engine, Leela Chess Zero (LC0) is roughly 50 elo behind. They have been #1/#2 for the last 6-7 years. All modern chess engines have two parts: a search algorithm that searches through positions to find the best move, and a neural network to evaluate positions. The main improvement over the last 5 years in LC0 has been in improving the quality of the networks they use over time.
LC0 is run by a small group of volunteers. Their most recent model, BT4, took 6 months to train on 8xA100, running at something like 1% MFU. This makes it difficult to do serious tuning or experimentation. They have not released a new model in the last year and a half. Their most recent big run, BT5, failed. They see this as evidence that RPE, a new technique for injecting positional information, doesn't work. I think they are wrong, and I have experiments showing it does work with some tuning.
I have made infrastructure of workable quality such that I can do rigorous experimentation. Once I've fully tuned a LC0-style architecture, I plan on doing architecture research that I can be confident in. I am working with a friend from OpenAI model training who is giving advice on how to do tuning, what to look for, etc. Most of the arch ideas that make sense here don't make sense for LLMs so he has a fairly free hand. My scaling results can be found here.
As stands I'm spending ~$100 per tuning run. Work would progress much faster if I had funding so I could 1) do more tuning runs 2) train bigger models 3) tune multiple things in conjunction instead of independently. The vast majority of my compute is research/testing compute, so spot instances or other unreliable compute is fine for me. A big block of H100s for a month is not ideal.
I think it's fairly likely I will make the best chess program out there. If this happens it will probably make news in programming/nerd circles. I will do write ups about it and talk about how you made it possible. Could also write a blog post for you or do other promotional work.