3 ms·

Few weeks ago I did an experiment after a discussion here about LLMs and chess. Basically inventing a board game and play against ChatGPT and see what happened

Few weeks ago I did an experiment after a discussion here about LLMs and chess. Basically inventing a board game and play against ChatGPT and see what happened. It was not able to do a single move, even having provided all the possible start moves in the prompt as part of the rules. Not that I had a lot of hope about it, but it was definitely way worst than I expected. If someone wants to take a look at it: https://joseprupi.github.io/misc/2023/06/08/chat_gpt_board_game.html https://joseprupi.github.io/misc/2023/06/08/chat_gpt_board_g...

golergka 3y ago
You haven't specified what model did you use, and the green ChatGPT icon in the shared conversation usually signifies GPT-3.5 model. Here's my attempt at similar conversation — it seems GPT-4 is able to visualise the board and at least do a valid first move. https://chat.openai.com/share/98427e21-678c-4290-aa8f-da8e939a2ad9 https://chat.openai.com/share/98427e21-678c-4290-aa8f-da8e93...
- melenaboija 3y ago
  Interesting. The model was whatever was up that that time, so probably was 3.5 if you say so.
  - golergka 3y ago
    Your conversation is from June, GPT-4 was available for almost half a year at that point.
    - melenaboija 3y ago
      Ok
  - xcv123 3y ago
    If you were using the free ChatGPT then you were playing with an obsolete LLM. GPT-4 has an estimated 10x parameters of GPT-3.5 (1.8 trillion vs 175 billion), and other improvements.
FrustratedMonky 3y ago
I'm older. I've bought 'new' board games for kids. Then, I have been un-able to play because the instructions were pretty bad. Humans also need to 'learn'. Need a few play-throughs. No human is going out and 'in a vacuum' with no experience, buying Risk and from scratch, read instructions and play perfect game winning strategy.
- melenaboija 3y ago
  The thing is that I wanted to prove that ChatGPT was not able to learn from the rules and that is indeed a Language Model that puts one token after the other, if it know how to play chess it is because it has seen games in the past, as I say in my post: > If it is not memorizing, how do you think is doing it? (me) > by trying to learning the general rules that to explain the dataset and minimize its loss. That’s what machine learning is about, it’s not called machine memorizing.
  - xcv123 3y ago
    > Language Model that puts one token after the other, The interesting thing about the one-token-at-a-time process in OpenAI transformer LLMs is how the Attention Mechanism is executing ~1600 processes in parallel over the entire context window for each new token generated. So it is dynamically re-evaluating the entire context (including the rules of the game) in relation to the next token at each step.
GaggiX 3y ago
I have played some moves with GPT-4 and they seem right to me, what does this mean? That the model switched from not understanding to understanding, from unintelligent to intelligent? I don't think so, GPT-4 is just a more intelligent model than GPT-3.5 and it does understand more. Also in this game if I don't move the queen I force a draw, right?
- melenaboija 3y ago
  > what does this mean? I don't know, take your own conclusions, I tried what I tried with the results I got. And the reason I created a Monte Carlo Engine to play the game was specifically because of this, I expected ChatGPT to be able to make moves but actually not being good with the game. You can try yourself, the code is available. > Also in this game if I don't move the queen I force a draw, right? I don't know as there is no time but I assume it is mandatory to move when, what happens in a chess game with no time if someone does not want to move? Same applies here.
  - GaggiX 3y ago
    >I don't know, take your own conclusions The API cost of the game is getting noticeable, but I think you were just being naive about LLM limitations, there is simply no way that it can answer all questions simply by memorization. A simpler way is to just invent a programming language and ask the model to solve problems with it, at least I don't have to write down the position of a game ahah Also I have trained models to do additions in the past, I removed many possible combinations of digits to show in the dataset, but after training the model was able to solve all of them, meaning it learned the algorithm and not just memorized the answers, I did it because a friend of mine thought like you that LLMs just memorize the answers from the dataset and cannot learn, but that is not how they work. About the game, I realize that I cannot move the queen like in chess, so in this game I will eventually fall into a zugzwang, trying not to move the queen.
    - melenaboija 3y ago
      I was not being naive, as I said above I was expecting the model to be able to move but not to be good. That is why I created an engine to play against, to see how good/bad it is. My exact entire point was that it would be really surprising that chatgt only knowing chess rules could play at 1400 ELO level. And I did not give a crap about how to play a single game and was not meant to see how smart HN users are but how chatGPT is, but whatever conclusion a human would take without playing a single game (not trying to move the queeen) is what chatgpt should do, right? BTW, I simulated few games of the engine against the engine to see what happened and it does not seem as obvious as you say.

golergka 3y ago

You haven't specified what model did you use, and the green ChatGPT icon in the shared conversation usually signifies GPT-3.5 model. Here's my attempt at similar conversation — it seems GPT-4 is able to visualise the board and at least do a valid first move. https://chat.openai.com/share/98427e21-678c-4290-aa8f-da8e939a2ad9 https://chat.openai.com/share/98427e21-678c-4290-aa8f-da8e93...

melenaboija 3y ago
Interesting. The model was whatever was up that that time, so probably was 3.5 if you say so.
- golergka 3y ago
  Your conversation is from June, GPT-4 was available for almost half a year at that point.
  - melenaboija 3y ago
    Ok
- xcv123 3y ago
  If you were using the free ChatGPT then you were playing with an obsolete LLM. GPT-4 has an estimated 10x parameters of GPT-3.5 (1.8 trillion vs 175 billion), and other improvements.

FrustratedMonky 3y ago

I'm older. I've bought 'new' board games for kids. Then, I have been un-able to play because the instructions were pretty bad. Humans also need to 'learn'. Need a few play-throughs. No human is going out and 'in a vacuum' with no experience, buying Risk and from scratch, read instructions and play perfect game winning strategy.

melenaboija 3y ago
The thing is that I wanted to prove that ChatGPT was not able to learn from the rules and that is indeed a Language Model that puts one token after the other, if it know how to play chess it is because it has seen games in the past, as I say in my post: > If it is not memorizing, how do you think is doing it? (me) > by trying to learning the general rules that to explain the dataset and minimize its loss. That’s what machine learning is about, it’s not called machine memorizing.
- xcv123 3y ago
  > Language Model that puts one token after the other, The interesting thing about the one-token-at-a-time process in OpenAI transformer LLMs is how the Attention Mechanism is executing ~1600 processes in parallel over the entire context window for each new token generated. So it is dynamically re-evaluating the entire context (including the rules of the game) in relation to the next token at each step.

GaggiX 3y ago

I have played some moves with GPT-4 and they seem right to me, what does this mean? That the model switched from not understanding to understanding, from unintelligent to intelligent? I don't think so, GPT-4 is just a more intelligent model than GPT-3.5 and it does understand more. Also in this game if I don't move the queen I force a draw, right?

melenaboija 3y ago
> what does this mean? I don't know, take your own conclusions, I tried what I tried with the results I got. And the reason I created a Monte Carlo Engine to play the game was specifically because of this, I expected ChatGPT to be able to make moves but actually not being good with the game. You can try yourself, the code is available. > Also in this game if I don't move the queen I force a draw, right? I don't know as there is no time but I assume it is mandatory to move when, what happens in a chess game with no time if someone does not want to move? Same applies here.
- GaggiX 3y ago
  >I don't know, take your own conclusions The API cost of the game is getting noticeable, but I think you were just being naive about LLM limitations, there is simply no way that it can answer all questions simply by memorization. A simpler way is to just invent a programming language and ask the model to solve problems with it, at least I don't have to write down the position of a game ahah Also I have trained models to do additions in the past, I removed many possible combinations of digits to show in the dataset, but after training the model was able to solve all of them, meaning it learned the algorithm and not just memorized the answers, I did it because a friend of mine thought like you that LLMs just memorize the answers from the dataset and cannot learn, but that is not how they work. About the game, I realize that I cannot move the queen like in chess, so in this game I will eventually fall into a zugzwang, trying not to move the queen.
  - melenaboija 3y ago
    I was not being naive, as I said above I was expecting the model to be able to move but not to be good. That is why I created an engine to play against, to see how good/bad it is. My exact entire point was that it would be really surprising that chatgt only knowing chess rules could play at 1400 ELO level. And I did not give a crap about how to play a single game and was not meant to see how smart HN users are but how chatGPT is, but whatever conclusion a human would take without playing a single game (not trying to move the queeen) is what chatgpt should do, right? BTW, I simulated few games of the engine against the engine to see what happened and it does not seem as obvious as you say.