Please feel free to test. Ask anything about chess rules. English only. You can do follow up questions, argue, or give opinion. If you think your session with the AI will be good for its learning, send me message or comment here. It does not do machine learning automatically, but I can compile good session for consideration to the future answers. You also can login, so that I can identify who talked and base on your TD certification if any, that will matters to the weight of the sessions.
When you say “all basic AI training,” do you mean you did some training (fine-tuning) of an existing model? Or just that you’re using a standard model? If it’s a standard model, are you using a RAG pattern with the rulebooks as your sources to pull from? I tried that with gpt4 and the USCF rulebook a while ago and got pretty mediocre results at the time, but I still think it’s worth iterating on it. The best case would be to always include a reference to the appropriate section of the rulebook so the user can quickly go to the right place to validate the answer. It would be even better if the rules were in a web form flexible enough to link to specific sections.
I tried yours but I’m getting HTTP 500s to requests.
AI tools have improved several fold over the last 12 months, but they still have a long ways to go, and tools trained to use specific data without revealing non-public information are still difficult to do.
Even asking a question like “Who was the youngest player to achieve a US Chess regular rating of 2200?” would probably have to be disallowed on a public interface because it has to have access to individual birthdates. We have looked into having such a tool for staff usage, though.
I’ve heard of people training Clyde on a data structure without access to the underlying data, apparently it can write complex queries faster than humans can and the humans can then execute them. Whether it can optimize them as well as an experienced programmer remains to be seen, but that should be learnable as well.
True, but we’ve been running that query for several years now internally, and the last few potential candidates for the youngest master missed by just a handful of days.
It is not a training. The AI term is just simple, it is MCP, so we hook the LLM to a player lookup. We read the intention of the messages and if asking for player lookup, then it triggers the lookup and retrieve the info.
