game commentary / guidance using Ollama models ?

[venkyyuvy]

prompt:

Player: d2d4
Computer: 98f6
Player: d1d3
Computer: d7d5
Player: d3a6
Computer: b8a6
Player: e2e4
Computer: f6e4

how is my performance so far and what should I do next>

========

response:

Performance Analysis

Current Position

Let's break down the moves you've made so far:

1. d2-d4: A solid opening move, controlling the center.
2. d1-d3: This move supports your pawn on d4 but may not be the most aggressive option. Typically, developing pieces or playing e2-e4 here would be more common.
3. d3-a6: This is an unusual move that seems to be aiming for an attack on the black pieces but does not develop your position effectively.
4. e2-e4: A strong central move, aiming to open lines and challenge Black’s control of the center.
5. f6-e4: Black has captured your pawn on e4, gaining material.

Evaluation

• Positional Strength: You have a central pawn structure, but your pieces are not well developed. The move d3-a6 was particularly passive and doesn't contribute to piece development.
• Material Balance: Currently, you are down a pawn after Black's last move.

Suggested Next Steps

1. Develop Your Pieces

Focus on getting your pieces into play. Consider moves like:

• Nf3: Developing your knight and attacking the pawn on e4.
• c4: This can challenge the center and potentially lead to recapturing the pawn on e4.

2. Control the Center

Maintaining control of the center is crucial:

• After Nf3, if Black responds with something like ...dxe4, you can recapture with your knight.

3. Prepare for Castling

Look for opportunities to castle early to ensure your king's safety and connect your rooks.

Example Move

A good candidate for your next move could be:

• Nf3: This develops a piece while attacking the e4 pawn and prepares for further development.

By focusing on these principles—development, center control, and king safety—you can improve your position moving forward!

[whiletruelearn]

Thanks for exploring this @venkyyuvy . There is definitely a lot we might be able to accomplish with LLMs. I haven't looked into it yet. I feel there is going to be better response if the input was on chess notation itself.

One thing I have thought of is this scenario

• Can termichess become some form a vertical personal agent that helps someone become good at chess. Chess is one of those games where it's easy to objectively tell if someone is getting good at or become worse off. This is definitely something in my mind that will increase the usefulness of the app.

• My initial idea was to use Analysis on the game #17 stockfish directly and give some inputs on how good the recent move is and maybe suggest a better move.

[whiletruelearn]

Few updates on this, Apparently LLM's have a lower ELO than specialised engines like stockfish. Unless we train a specific chess based LLM like here (https://arxiv.org/pdf/2402.04494) . Zeroshot is not going to yield more than existing engines atleast for now.

However I feel this could be a cool chess analyser component and could be a separate project itself in a way.

Here is what I am thinking.

1. From an image of a chess board get the current state of game . This might need a vision model. We can start with chess.com style digital boards and move to real world boards latter. The current state of the game is FEN .
2. Pass over the FEN to stockfish and get the best next possible move.
3. Show the move visually in board.
4. This could be like a gradio/ streamlit app to start with.