Your network blocks the Lichess assets!

lichess.org
Donate

AI Chess Coaches?

A coach does more than explain what is happening on the board. They decide what matters, what can be ignored,

I think this sentence hits the bullseye. Perhaps true intelligence, the general one you are talking about, is less about making connections and more knowing which connections NOT to make. AGI by Saint-Exupéry.

As you well know, this kind of question - how to make computer teach - is one that I am asking and trying to find an answer to on a daily basis. Here are some ideas:

  1. defining what something is not is just as important as defining what it is. Applying it to the AI Coach, it's important to find what part of the attempts you have already made are not "coachy".

  2. you were mentioning finding what SF wants to do. The answer is simple: it doesn't. And using point 1. we should try to analyse what is the difference between a SF line and a human line. A friend of mine recently was noting that the best move in a position reached by a gambit is not "in the spirit of the opening". So then defining that spirit becomes imperative. Perhaps a plan is not only about what to do, but also about NOT to do. Again, point 1.

  3. there are two parts of what I would call "a plan". The first is finding some general ideas that apply to the position and sticking with them. The second is determining when to abandon the plan. I believe that AlphaChess-like training can be achieved for an AI that "plans" by enforcing very strong constraints like demanding a plan before a move, punishing deviations from the plan, updating plans with exit conditions (when to stop following a plan should be a part of that plan), use a different engine to perform the moves within the plan - in other words force the training in some sort of high level DSL that does not contain actual moves and restricts a SF kind of engine.

  4. the beauty aspect was always the one that broke my focus from the first three mentioned above. I always felt that without a proper definition of beauty, there is no point in making plans, finding moves or anything really. However, beautiful moves are not really defined by long term strategy, more by their relation to chess principles. You may want to play the O'Sullivan or making moves to lead to Legal's Mate, but that's more like move memorization, not generating beauty with your moves. I have no real answer to this. All I have is some vague ideas like "find a list of games densely annotated by humans and train AIs on the emotional cues of the comments for beautiful moves" or "train AIs to differentiate between the move made in the game and the one that should be made in puzzle position", but you see that they are not real plans, pardon the pun. I guess one could try to find emotion in chat messages or even facial expressions of people watching games in real time and train on the emotion generated by the moves, but then the realization of the consequences of a chess move are rarely immediate. It's chess, not an action movie.

  5. statistics lead to mediocre choices. I think that's the biggest issue of chess engines (and chess learning) today. You train against the most common moves and they train against most common responses to those moves, you get a ton of theory, all leading to equal positions. What I consider beautiful you might not consider so. What plans I make (I never make plans, but let's go with this thought experiment) might not be the ones you make. There is a style of play that people talk about. A coach (as well as a proper chess AI) should have a style. It should clash and conflict with the styles of other coaches. It should never be about 'what is the best AI coach' but 'what is the coach that best fits MY style'.

  6. we humans mostly use "tricks" to play chess. Memorizing an opening rather than analysing it every time you play is an optimization trick. Chess principles are just shortcuts. We use them because we can't really think that much, that fast or that deep and our memory is crap. AIs should be forced to derive tricks as well because of similar limitations imposed to them in the self learning phase.

I am sure there are others, but I want to have a shorter comment than your blog post ;-)

> A coach does more than explain what is happening on the board. They decide what matters, what can be ignored, I think this sentence hits the bullseye. Perhaps true intelligence, the general one you are talking about, is less about making connections and more knowing which connections NOT to make. AGI by Saint-Exupéry. As you well know, this kind of question - how to make computer teach - is one that I am asking and trying to find an answer to on a daily basis. Here are some ideas: 1. defining what something is not is just as important as defining what it is. Applying it to the AI Coach, it's important to find what part of the attempts you have already made are not "coachy". 2. you were mentioning finding what SF wants to do. The answer is simple: it doesn't. And using point 1. we should try to analyse what is the difference between a SF line and a human line. A friend of mine recently was noting that the best move in a position reached by a gambit is not "in the spirit of the opening". So then defining that spirit becomes imperative. Perhaps a plan is not only about what to do, but also about NOT to do. Again, point 1. 3. there are two parts of what I would call "a plan". The first is finding some general ideas that apply to the position and sticking with them. The second is determining when to abandon the plan. I believe that AlphaChess-like training can be achieved for an AI that "plans" by enforcing very strong constraints like demanding a plan before a move, punishing deviations from the plan, updating plans with exit conditions (when to stop following a plan should be a part of that plan), use a different engine to perform the moves within the plan - in other words force the training in some sort of high level DSL that does not contain actual moves and restricts a SF kind of engine. 4. the beauty aspect was always the one that broke my focus from the first three mentioned above. I always felt that without a proper definition of beauty, there is no point in making plans, finding moves or anything really. However, beautiful moves are not really defined by long term strategy, more by their relation to chess principles. You may want to play the O'Sullivan or making moves to lead to Legal's Mate, but that's more like move memorization, not generating beauty with your moves. I have no real answer to this. All I have is some vague ideas like "find a list of games densely annotated by humans and train AIs on the emotional cues of the comments for beautiful moves" or "train AIs to differentiate between the move made in the game and the one that should be made in puzzle position", but you see that they are not real plans, pardon the pun. I guess one could try to find emotion in chat messages or even facial expressions of people watching games in real time and train on the emotion generated by the moves, but then the realization of the consequences of a chess move are rarely immediate. It's chess, not an action movie. 5. statistics lead to mediocre choices. I think that's the biggest issue of chess engines (and chess learning) today. You train against the most common moves and they train against most common responses to those moves, you get a ton of theory, all leading to equal positions. What I consider beautiful you might not consider so. What plans I make (I never make plans, but let's go with this thought experiment) might not be the ones you make. There is a style of play that people talk about. A coach (as well as a proper chess AI) should have a style. It should clash and conflict with the styles of other coaches. It should never be about 'what is the best AI coach' but 'what is the coach that best fits MY style'. 6. we humans mostly use "tricks" to play chess. Memorizing an opening rather than analysing it every time you play is an optimization trick. Chess principles are just shortcuts. We use them because we can't really think that much, that fast or that deep and our memory is crap. AIs should be forced to derive tricks as well because of similar limitations imposed to them in the self learning phase. I am sure there are others, but I want to have a shorter comment than your blog post ;-)

Thanks for the shoutout to my ChessAgine project. AI coaching, especially with LLMs on their own, is very hard to almost impossible with current LLM tech. The thing is reddit companies and social media have made AI in chess so hard to work with; people automatically slap labels on newer projects that are trying to work from different angles, like my agine project. To this day, people confuse my project with an AI coach, or I hear things like "Oh its wrong haha your project is no way near a chess coach," but they don't check the agine docs and read that it's just an engine to the LLM interface, not meant to be a coach to start with lol. You will encounter this as you build, I do warn you. Also, my interface is a work in progress, as more devs get interested, we will have more sophisticated systems in future, but again, no way near a real human chess coach. I think the premise of "AI Chess coach" in 2026 is just a marketing scheme, and you are right, this will take a lot of work.

Just keep building

Good luck!

Thanks for the shoutout to my ChessAgine project. AI coaching, especially with LLMs on their own, is very hard to almost impossible with current LLM tech. The thing is reddit companies and social media have made AI in chess so hard to work with; people automatically slap labels on newer projects that are trying to work from different angles, like my agine project. To this day, people confuse my project with an AI coach, or I hear things like "Oh its wrong haha your project is no way near a chess coach," but they don't check the agine docs and read that it's just an engine to the LLM interface, not meant to be a coach to start with lol. You will encounter this as you build, I do warn you. Also, my interface is a work in progress, as more devs get interested, we will have more sophisticated systems in future, but again, no way near a real human chess coach. I think the premise of "AI Chess coach" in 2026 is just a marketing scheme, and you are right, this will take a lot of work. Just keep building Good luck!

@TotalNoob69 said in #2:

defining what something is not is just as important as defining what it is. Applying it to the AI Coach, it's important to find what part of the attempts you have already made are not "coachy".
you were mentioning finding what SF wants to do. The answer is simple: it doesn't. And using point 1. we should try to analyse what is the difference between a SF line and a human line. A friend of mine recently was noting that the best move in a position reached by a gambit is not "in the spirit of the opening". So then defining that spirit becomes imperative. Perhaps a plan is not only about what to do, but also about NOT to do. Again, point 1.

Yeah, that is also a very interesting thought, the spirit of the position is also an interesting concept, as how does one capture this. I have been looking at comparing Stockfish with Maia to see about the deviations, but that is tricky, and the hardest part is explaining this.

@TotalNoob69 said in #2: > defining what something is not is just as important as defining what it is. Applying it to the AI Coach, it's important to find what part of the attempts you have already made are not "coachy". you were mentioning finding what SF wants to do. The answer is simple: it doesn't. And using point 1. we should try to analyse what is the difference between a SF line and a human line. A friend of mine recently was noting that the best move in a position reached by a gambit is not "in the spirit of the opening". So then defining that spirit becomes imperative. Perhaps a plan is not only about what to do, but also about NOT to do. Again, point 1. Yeah, that is also a very interesting thought, the spirit of the position is also an interesting concept, as how does one capture this. I have been looking at comparing Stockfish with Maia to see about the deviations, but that is tricky, and the hardest part is explaining this.

@TotalNoob69 said in #2:

there are two parts of what I would call "a plan". The first is finding some general ideas that apply to the position and sticking with them. The second is determining when to abandon the plan. I believe that AlphaChess-like training can be achieved for an AI that "plans" by enforcing very strong constraints like demanding a plan before a move, punishing deviations from the plan, updating plans with exit conditions (when to stop following a plan should be a part of that plan), use a different engine to perform the moves within the plan - in other words force the training in some sort of high level DSL that does not contain actual moves and restricts a SF kind of engine.

Yes, this is something that I have been thinking about, how do you define and codify a plan? Also, how can you heuristically determine what the plan is, like in the blog, I have looking at scanning the line all the way through to try and deduce what plan SF is going for, but that requires a lot of work analysing each position and the defining the line based on the plan. I also looked at running the main line for Maia for several moves to see what plan might exist there, but the challenge is definitions.

@TotalNoob69 said in #2: > there are two parts of what I would call "a plan". The first is finding some general ideas that apply to the position and sticking with them. The second is determining when to abandon the plan. I believe that AlphaChess-like training can be achieved for an AI that "plans" by enforcing very strong constraints like demanding a plan before a move, punishing deviations from the plan, updating plans with exit conditions (when to stop following a plan should be a part of that plan), use a different engine to perform the moves within the plan - in other words force the training in some sort of high level DSL that does not contain actual moves and restricts a SF kind of engine. Yes, this is something that I have been thinking about, how do you define and codify a plan? Also, how can you heuristically determine what the plan is, like in the blog, I have looking at scanning the line all the way through to try and deduce what plan SF is going for, but that requires a lot of work analysing each position and the defining the line based on the plan. I also looked at running the main line for Maia for several moves to see what plan might exist there, but the challenge is definitions.

@HollowLeaf said in #4:

the spirit of the position is also an interesting concept, as how does one capture this. I have been looking at comparing Stockfish with Maia to see about the deviations, but that is tricky, and the hardest part is explaining this.

Well, I think that's backwards. SF doesn't have a plan and neither does Maia. So one has to:

  • start with a plan, complete with exit conditions
  • restrict moves to said plan
  • accept that it might not be best
  • explain the plan, rather than the moves
@HollowLeaf said in #4: > the spirit of the position is also an interesting concept, as how does one capture this. I have been looking at comparing Stockfish with Maia to see about the deviations, but that is tricky, and the hardest part is explaining this. Well, I think that's backwards. SF doesn't have a plan and neither does Maia. So one has to: - start with a plan, complete with exit conditions - restrict moves to said plan - accept that it might not be best - explain the plan, rather than the moves

@TotalNoob69 said in #2:

the beauty aspect was always the one that broke my focus from the first three mentioned above. I always felt that without a proper definition of beauty, there is no point in making plans, finding moves or anything really. However, beautiful moves are not really defined by long term strategy, more by their relation to chess principles. You may want to play the O'Sullivan or making moves to lead to Legal's Mate, but that's more like move memorization, not generating beauty with your moves. I have no real answer to this. All I have is some vague ideas like "find a list of games densely annotated by humans and train AIs on the emotional cues of the comments for beautiful moves" or "train AIs to differentiate between the move made in the game and the one that should be made in puzzle position", but you see that they are not real plans, pardon the pun. I guess one could try to find emotion in chat messages or even facial expressions of people watching games in real time and train on the emotion generated by the moves, but then the realization of the consequences of a chess move are rarely immediate. It's chess, not an action movie.

Funny thing, I actually have a blog that I am writing about the hunt for Brilliancies, and what this means, I guess that is a good a starting point. Right now, brilliancies is a marketing term, but how do you define it? a piece sack, for mate in 2 is not a brilliancy, but I was combining Maia and Stockfish to try and find the probable moves that a human would make vs the engine, then have something, but that is a very narrow attempt to capture something.

But it is still a tough one, as like my blog, Chess is science and art, and art is subjective, I saw a video where Hikaru critiqued Mikhail Tals games, and while they were beautiful and amazing at the time, todays Super GMs, would be able to defend and win...

@TotalNoob69 said in #2: > the beauty aspect was always the one that broke my focus from the first three mentioned above. I always felt that without a proper definition of beauty, there is no point in making plans, finding moves or anything really. However, beautiful moves are not really defined by long term strategy, more by their relation to chess principles. You may want to play the O'Sullivan or making moves to lead to Legal's Mate, but that's more like move memorization, not generating beauty with your moves. I have no real answer to this. All I have is some vague ideas like "find a list of games densely annotated by humans and train AIs on the emotional cues of the comments for beautiful moves" or "train AIs to differentiate between the move made in the game and the one that should be made in puzzle position", but you see that they are not real plans, pardon the pun. I guess one could try to find emotion in chat messages or even facial expressions of people watching games in real time and train on the emotion generated by the moves, but then the realization of the consequences of a chess move are rarely immediate. It's chess, not an action movie. Funny thing, I actually have a blog that I am writing about the hunt for Brilliancies, and what this means, I guess that is a good a starting point. Right now, brilliancies is a marketing term, but how do you define it? a piece sack, for mate in 2 is not a brilliancy, but I was combining Maia and Stockfish to try and find the probable moves that a human would make vs the engine, then have something, but that is a very narrow attempt to capture something. But it is still a tough one, as like my blog, Chess is science and art, and art is subjective, I saw a video where Hikaru critiqued Mikhail Tals games, and while they were beautiful and amazing at the time, todays Super GMs, would be able to defend and win...

@TotalNoob69 said in #2:

statistics lead to mediocre choices. I think that's the biggest issue of chess engines (and chess learning) today. You train against the most common moves and they train against most common responses to those moves, you get a ton of theory, all leading to equal positions. What I consider beautiful you might not consider so. What plans I make (I never make plans, but let's go with this thought experiment) might not be the ones you make. There is a style of play that people talk about. A coach (as well as a proper chess AI) should have a style. It should clash and conflict with the styles of other coaches. It should never be about 'what is the best AI coach' but 'what is the coach that best fits MY style'.

Yeah, I think this is why 960 is picking up a bit, but still not mainstream, but I wonder if it would get there. I do agree that like in Boxing, style makes fights, and finding your own style is probably the hardest part of coaching, and more interesting, we grow over time, so our style changes, and a coach should take that into account. I think I read somewhere that before Magnus became super solid, he was a super attacking player in his youth.

@TotalNoob69 said in #2: > statistics lead to mediocre choices. I think that's the biggest issue of chess engines (and chess learning) today. You train against the most common moves and they train against most common responses to those moves, you get a ton of theory, all leading to equal positions. What I consider beautiful you might not consider so. What plans I make (I never make plans, but let's go with this thought experiment) might not be the ones you make. There is a style of play that people talk about. A coach (as well as a proper chess AI) should have a style. It should clash and conflict with the styles of other coaches. It should never be about 'what is the best AI coach' but 'what is the coach that best fits MY style'. Yeah, I think this is why 960 is picking up a bit, but still not mainstream, but I wonder if it would get there. I do agree that like in Boxing, style makes fights, and finding your own style is probably the hardest part of coaching, and more interesting, we grow over time, so our style changes, and a coach should take that into account. I think I read somewhere that before Magnus became super solid, he was a super attacking player in his youth.

@HollowLeaf said in #5:

how do you define and codify a plan? Also, how can you heuristically determine what the plan is

That's the hard part :D I have no idea. I believe starting with generating plans would be easier. Then, for existing games, just generate more plans and calculate the likelihood for each that it was what was intended. Third step, apply it to women and get rich.

@HollowLeaf said in #5: > how do you define and codify a plan? Also, how can you heuristically determine what the plan is That's the hard part :D I have no idea. I believe starting with generating plans would be easier. Then, for existing games, just generate more plans and calculate the likelihood for each that it was what was intended. Third step, apply it to women and get rich.

@TotalNoob69 said in #6:

the spirit of the position is also an interesting concept, as how does one capture this. I have been looking at comparing Stockfish with Maia to see about the deviations, but that is tricky, and the hardest part is explaining this.

Well, I think that's backwards. SF doesn't have a plan and neither does Maia. So one has to:

  • start with a plan, complete with exit conditions
  • restrict moves to said plan
  • accept that it might not be best
  • explain the plan, rather than the moves

It is an interesting challenge, the question is how do you deterministically work out a plan from any given position, because there could be multiple, all with pros and cons. If I gave you a random position today, how can you figure out the plan programmatically? It is very hard, whereas, if you look through the stockfish line or Maia lines, you can maybe deduce that this is what SF was going for, and hence the plan i.e. what is this position is the line and what is the line trying to achieve, therefore what is the plan. You could possible branch out a little and go through multiple lines and figure out the possible options, but that is a tough challenge.

@TotalNoob69 said in #6: > > > the spirit of the position is also an interesting concept, as how does one capture this. I have been looking at comparing Stockfish with Maia to see about the deviations, but that is tricky, and the hardest part is explaining this. > > Well, I think that's backwards. SF doesn't have a plan and neither does Maia. So one has to: > - start with a plan, complete with exit conditions > - restrict moves to said plan > - accept that it might not be best > - explain the plan, rather than the moves It is an interesting challenge, the question is how do you deterministically work out a plan from any given position, because there could be multiple, all with pros and cons. If I gave you a random position today, how can you figure out the plan programmatically? It is very hard, whereas, if you look through the stockfish line or Maia lines, you can maybe deduce that this is what SF was going for, and hence the plan i.e. what is this position is the line and what is the line trying to achieve, therefore what is the plan. You could possible branch out a little and go through multiple lines and figure out the possible options, but that is a tough challenge.