AI Chess Coach

Scenario: A chess player leaves the playing hall and calls someone on his phone.

Student: Hey Coach, I just lost my game in the tournament. It was a tough match, and I think I blundered in the middlegame.

AI Chess Coach: I’m here for you. Losing is part of learning. Can you send me the PGN of the game so I can help you analyze it?

Student: Sure, here it is. I felt confident in the opening, but after a few exchanges, I just lost control.

AI Chess Coach: Let’s take a look together. I see where the exchanges started. It seems like you missed the tension in the center. Remember, every move affects the balance.

Student: Yeah, I realize that now. I felt pretty down after that. It’s like I knew I was going to lose.

AI Chess Coach: It’s natural to feel disappointed, but every game is a lesson. What you’re feeling is a sign of your passion and desire to improve. Let’s focus on how you can use this experience to grow.

Student: Thanks, I needed to hear that. Do you think I could have saved the game at some point?

AI Chess Coach: Absolutely. There was a moment where a different pawn move would have fortified your position. We’ll work on recognizing these moments better.

Student: I appreciate it. How can I avoid such mistakes in the future?

AI Chess Coach: Practice makes perfect. We’ll do some targeted exercises on pawn structures and positional play. And remember, resilience is key. You’ve got the potential, and with each game, you’re getting stronger.

Student: Thanks, Coach. I’ll keep pushing. I want to be ready for the next match.

AI Chess Coach: You will be. Let’s schedule some extra sessions this week. We’ll turn this setback into a comeback. Keep your head up!

Student: Will do, Coach. Thanks for the support. Let’s win the next one!

AI Chess Coach: That’s the spirit! We’re in this together. Onward and upward!

With the fast-paced advancement of LLMs (Large Language Models) like ChatGPT it is now feasible to create that AI based Chess Coach. I am not talking about a powerful chess engine that shows you the best move. A proper chess coach that will interact with a chess student looking at their games, have a question and answer session whereby the coach presents chess positions and see what the human student will respond with etc. It has human-like properties such as:

• Knowledge: A deep understanding of chess theory, tactics, and strategy.
• Patience: Willing to work with students at their own pace and skill level.
• Analytical: Able to dissect games and positions to provide insightful feedback.
• Motivational: Encourages students to push their limits and achieve their goals.
• Adaptable: Tailors coaching methods to fit the learning style of each student.
• Communicative: Clearly explains concepts and maintains open dialogue with students.
  

Also, this cannot be achieved just by using ChatGPT. Sure, the latest ChatGPT 4 can actually annotate a chess game, give “almost useful commentary” on a game as we’ve seen from the Jax-Milligan game. That’s because ChatGPT is a general purpose LLM trying to encompass every topic.

What is needed is an LLM with a specific purpose. In this case – Chess Coaching.

So what is needed?

• Hardware Requirements: Running an LLM locally requires significant computational power. You’ll need a powerful PC with a high-end CPU and GPU to handle the processing¹.
• Software and Models: There are frameworks and tools available that allow you to run LLMs locally. For example, you can use implementations like llama.cpp for running models like LLaMA locally¹.
• Data and Training: To create a chess-focused LLM, you would need a large dataset of chess games and commentary for training or fine-tuning the model to understand and explain chess moves².
• Fine-Tuning: For specific tasks like chess coaching, fine-tuning an existing LLM with chess data can be more practical than training a model from scratch².
• Integration: You can integrate the LLM with chess databases and software to provide a comprehensive coaching experience, analyzing games and offering insights³.
• Continuous Learning: The LLM can be designed to learn continuously from new games and coaching sessions, improving its coaching capabilities over time.
  

Ok enough with generalisation. Let’s get specific. What kind of computer or hardware would I need If I want to manage such an LLM?

• CPU: AMD Ryzen 5 7600 3.8 GHz 6-Core¹
• GPU: Dual NVIDIA GeForce RTX 3090 24 GB (for handling large models)²
• CPU Cooler: Noctua NH-D15¹
• Motherboard: Asus PROART B650-CREATOR ATX AM5¹
• Memory: G.Skill Flare X5 64 GB (2 x 32 GB) DDR5-5600 CL36¹
• Storage: Samsung 980 Pro 1 TB M.2-2280 PCIe 4.0 X4 NVME¹
• Power Supply: Corsair HX1200 Platinum 1200 W 80+ Platinum¹
• Case: Lian Li O11 Dynamic EVO XL¹
• Case Fans: 5x be quiet! Pure Wings 2 87 CFM 120 mm Fan¹

The cost:

• AMD Ryzen 5 7600: Approximately $299¹.
• Dual NVIDIA GeForce RTX 3090: Each GPU can cost around $1,499, so two would be $2,998¹.
• Noctua NH-D15 CPU Cooler: Prices can vary, but generally around $100.
• Asus PROART B650-CREATOR Motherboard: Around $300.
• G.Skill Flare X5 64 GB Memory: Approximately $300 for high-speed DDR5 memory.
• Samsung 980 Pro 1 TB SSD: Around $200 for a PCIe 4.0 NVMe SSD.
• Corsair HX1200 Platinum Power Supply: About $300 for a 1200W PSU.
• Lian Li O11 Dynamic EVO XL Case: Roughly $200.
• be quiet! Pure Wings 2 Case Fans: A set of 5 fans may cost about $100.

That comes to about USD 4797 or RM 22,699.40.

It is feasible but not cheap.

Also, let’s be real. One person cannot create and run an LLM. There is the question of man power. Traditional LLMs are maintained by large coperations with a large team of professionals and not individual hobbyist like me. But scale it down, it is possible with a team of like-minded chess enthusiasts who can do the following :

• Feed the LLM actual human annotated games so it can learn to do it itself.
• Monitor the effectiveness of the interaction between AI coach with human students.
• Database of games played at top level.
• and anything else we can think of to build it’s experience – in essence it’s “brain”
  

Like Alpha Zero which started off really weak, it gradually grew in knowledge and strength because it was being fed data. So strong that at the end of it, it was beating the strongest chess engines in the world. Same goes for a Chess Coach LLM. It needs to be fed. It’s performance will be poor and laughable in the beginning. But in time, it can only get better and better.

I’ve blogged about similar topics in the past but these were really “shiok sendiri” posts about what I dream would happen. But no more! Today an AI chess coach is actually possible and feasible.

The question is who will be the first to create an LLM-based chess coach?