For a couple of years I've been playing with the idea of AI chess coaches.
Why?
I don't have the money to pay a real coach! Call me cheap but I want a machine to teach me.
Of course there are materials out there one can use to improve like books, interactive digital courses like Chessable etc.
But these are general purpose marketed for profit to fit the general chess learning public. Everyone will agree with me if you had a personal coach, there would be leaps and bounds in chess improvement.
That's where a digital coach comes in. It's personal. It's available 24/7/ It's adaptable, learning from your good and bad habits. The ideal and perfect coach. And also improbable as if it was possible, someone would have already created in this fast moving age of AI tech.
I've posted many #syioksendiri posts on why it's not possible. At least not yet. But will be possible in the near future.
I think the future is now.
My initial thought was that you need a specialised Chess LLM - not those silly general purpose LLM like ChatGPT, Gemini, Claude etc which is great at many things but horrible at chess. Many still play illegal moves and cheat.
A new LLM based on chess would have been too expensive in terms of hardware, time and monetary costs.
The solution would be to use a free local LLM and do "fine tuning" - teaching it all it needs to know about chess so it can become good enough to commentate like a digital version of GothamChess.
Think of it. You have GothamChess (Levy Rosman) to explain games to you in his highly effective and engaging style. You can stop him at any time and ask about the position. He tells you all he knows about the position, what goes in the mind of both black and white, plans etc. He is "just" an IM but as a teacher, I think he is better than most Grandmasters. I just want a digital version of him.
How?
By investing in the right hardware and choosing a good LLM.
Hardware Requirements
To achieve this, hardware investment will be on GPU VRAM.
The "Sweet Spot": A setup with dual RTX 3090, 4090, or the newer 5090 cards (providing 24–48GB of VRAM) is ideal for running 8-bit QLoRA (Quantized Low-Rank Adaptation). This allows to fine-tune a powerful model like Gemma-4-31B, which is considered the "gold standard" for the linguistic nuance required for grandmaster-level commentary.
Entry Level: A single RTX 4090 (24GB) can handle 4-bit LoRA, which is faster but may result in slightly less "fluent" coaching.
The Hybrid Architecture
It is critical to understand that even the best fine-tuned LLMs can hallucinate tactics or illegal moves. Therefore, the most effective feasibility path is a hybrid system:
Chess Engine (Stockfish): Acts as the "source of truth," generating evaluation scores, best lines, and identifying tactical threats like blunders or forks.
Fine-tuned LLM (Gemma 4): Acts as the "translator," taking the engine's raw data and converting it into enthusiastic, "Gotham-style" commentary that explains the why behind moves in plain language.
With this setup, the AI can perform advanced coaching tasks:
Interactive Quizzing: The system can be designed to pause mid-game and ask, "What would you do here?", followed by an explanation of why the user's choice was correct or a mistake.
Narrative Structure: Instead of move-by-move data, the AI can split a game into opening, middlegame, and endgame phases, summarizing the "narrative" of the match just like a human coach.
Psychological Insight: By fine-tuning on human commentary, the AI can point out "panic" moves or "desperate" central lunges, making the learning process engaging rather than purely mathematical.
Training, Training, Training!
By leveraging open-source datasets from Lichess or Chess.com and using parameter-efficient tuning (QLoRA), we can (hopefully) create a professional-grade coach that runs fully offline on our own machine.
We'll feed it PDFs with high quality chess analysis and commentary and of course YouTube chess commentaries from top commentators like GothamChess and Sagar Shah (Chessbase India). I would not even to explain how"magical" this process takes place and many AI expert fail to explain how the machine learn from just looking at examples. Only that it works!!
For example Suno was fed with tons of songs and in the end, it learnt how to compose songs better than most humans. MidJourney was fed with the world's best photographers pictures and artists and it can now mimic photos and art that few can differentiate if it was from a professional photographer or artists - and the examples can go on and one. Suffice to say, given enough example data, the AI can mimic the expert to a high degree!
Of course it's going to take an investment in a PC with a good GPU or maybe some rented time on hosted GPUs at Amazon servers. This is the least of it! The biggest investment is time. The humungous amount of time to train, since we are not billionaires or startup and cannot afford data centres to create and learn from training data with expensive servers.
But the thing is, it is now feasible and possible.
