3 ms·

The chess puzzle for every LLM

2 points by mkagenius 2y ago

jqpabc123 2y ago
Up vote this enough and they'll be able to solve it next week --- the same way they solve most things --- by memorizing the answer from "training". LLMs exhibit the characteristics of "savant syndrome". https://en.wikipedia.org/wiki/Savant_syndrome https://en.wikipedia.org/wiki/Savant_syndrome
PaulHoule 2y ago
Chess is a classic case which LLMs are structurally inadequate. The strength of LLMs is that they can "short circuit" and arrive the right answer by the wrong methods for squishy problems like language. [1] Chess is about precision and I'm awful at it myself because of the line noise in my schizotypal brain, even if I can compensate pretty well for it doing hard math. Short-circuiting just doesn't cut it for chess, which is why you get the invalid moves (the only requirement for a chess program is that it never make invalid moves) It is a tree search problem, even if human intuition makes it so we can think through about 20 board positions whereas my α-β program has to look at 2-20 million positions to challenge a serious beginner. Yes, the problem could be solved by a lookup in an endgame table; the 5 piece table is reasonable to keep with a specialized chess program (think how big an AAA game is!) but not reasonable to store in an LLM which does many other things than chess. To make my chess program able to challenge better players the direct path is to make it look at more positions, and most of that is optimizing the code, not allocating any objects in the inner loop, calculating the evaluation function incrementally, off-heap transposition table, etc. Not the path of ChatGPT-o. It is a thing to make a neural chess engine that picks better moves and doesn't have to look at as many positions [2], but it's a specialized thing; chess just isn't a model of general intelligence. [1] I was working on foundation models the summer BERT came out and had a technique of "predictive evaluation" that let us look at a small number of examples in detail and decide what the potential and weaknesses of a system were. By this method I saw BERT as a huge step forward, yet, this method entirely discounted "get the right answer by the wrong path" and we might have rejected the path that led to today's LLMs. [2] https://en.wikipedia.org/wiki/AlphaZero https://en.wikipedia.org/wiki/AlphaZero
- jqpabc123 2y ago
  Chess is a classic case which LLMs are structurally inadequate. They are structurally inadequate in a lot of cases. The real problem is it's really difficult to know in advance if your case is one of them (HINT: Don't trust their cooking advice). They are simply not predictable and that makes them less than trustworthy. LLMs turn traditional computing on it's head --- instead of highly accurate results at low cost, you get questionable results at high cost.
  - PaulHoule 2y ago
    They're structurally adequate for language translation in the sense that it is a fuzzy problem, people won't agree what a "good translation" is (translations of French writers like Baudrillard are bad [obscurantist] in my mind that they translate quotodienne to the $500 word quotidian instead of an everyday word like 'everyday') Even if an LLM based translator makes certain mistakes, other automatic translator may make even more mistakes. They aren't structurally adequate for the simpler problem of 'sort a list' because they can't make the N log N comparisons that it takes. It is hilarious to ask Copilot to sort a list in some order that isn't usual (like US states in alphabetical order) and see that it can't do it, and if you ask it if it can do it it will say it can, what probably it will get it right it says 'very high', then it goes ahead and tries and gets it wrong and if it is unlucky it even duplicates or drops items.
    - jqpabc123 2y ago
      MS is on a mission to prove to large numbers of people that AI is not the next big revolution in computing. Copilot has all the potential Cortana had.
      - PaulHoule 2y ago
        For me Copilot is at the plateau of productivity. When I'm trying to figure out how to test for certain things with Mockito it gives me working examples faster than I'd figure out if it was me and my lonesome with the manual. If I don't know why something works it often gives a correct answer. It does take some critical thinking to use it successfully, but the same is true for StackOverflow. I've also gotten good advice on subjects such as voice acting. Trying to knock out quick prototypes with other code assistants I'd say I have a 50% success rate. For some circumscribed tasks I've had good luck with existing code bases but if the problem is something like "figure out how somebody managed to download a file they shouldn't have downloaded and make sure it never happens again" you'd be a fool to even try using one.
al2o3cr 2y ago
Tried this with Gemini and it face-planted right at the start by mis-parsing the position of the black king and pawn :(
freediver 2y ago
Also check https://github.com/kagisearch/llm-chess-puzzles https://github.com/kagisearch/llm-chess-puzzles

2 points by mkagenius 2y ago

Up vote this enough and they'll be able to solve it next week --- the same way they solve most things --- by memorizing the answer from "training". LLMs exhibit the characteristics of "savant syndrome". https://en.wikipedia.org/wiki/Savant_syndrome https://en.wikipedia.org/wiki/Savant_syndrome

PaulHoule 2y ago

Chess is a classic case which LLMs are structurally inadequate. The strength of LLMs is that they can "short circuit" and arrive the right answer by the wrong methods for squishy problems like language. [1] Chess is about precision and I'm awful at it myself because of the line noise in my schizotypal brain, even if I can compensate pretty well for it doing hard math. Short-circuiting just doesn't cut it for chess, which is why you get the invalid moves (the only requirement for a chess program is that it never make invalid moves) It is a tree search problem, even if human intuition makes it so we can think through about 20 board positions whereas my α-β program has to look at 2-20 million positions to challenge a serious beginner. Yes, the problem could be solved by a lookup in an endgame table; the 5 piece table is reasonable to keep with a specialized chess program (think how big an AAA game is!) but not reasonable to store in an LLM which does many other things than chess. To make my chess program able to challenge better players the direct path is to make it look at more positions, and most of that is optimizing the code, not allocating any objects in the inner loop, calculating the evaluation function incrementally, off-heap transposition table, etc. Not the path of ChatGPT-o. It is a thing to make a neural chess engine that picks better moves and doesn't have to look at as many positions [2], but it's a specialized thing; chess just isn't a model of general intelligence. [1] I was working on foundation models the summer BERT came out and had a technique of "predictive evaluation" that let us look at a small number of examples in detail and decide what the potential and weaknesses of a system were. By this method I saw BERT as a huge step forward, yet, this method entirely discounted "get the right answer by the wrong path" and we might have rejected the path that led to today's LLMs. [2] https://en.wikipedia.org/wiki/AlphaZero https://en.wikipedia.org/wiki/AlphaZero

jqpabc123 2y ago
Chess is a classic case which LLMs are structurally inadequate. They are structurally inadequate in a lot of cases. The real problem is it's really difficult to know in advance if your case is one of them (HINT: Don't trust their cooking advice). They are simply not predictable and that makes them less than trustworthy. LLMs turn traditional computing on it's head --- instead of highly accurate results at low cost, you get questionable results at high cost.
- PaulHoule 2y ago
  They're structurally adequate for language translation in the sense that it is a fuzzy problem, people won't agree what a "good translation" is (translations of French writers like Baudrillard are bad [obscurantist] in my mind that they translate quotodienne to the $500 word quotidian instead of an everyday word like 'everyday') Even if an LLM based translator makes certain mistakes, other automatic translator may make even more mistakes. They aren't structurally adequate for the simpler problem of 'sort a list' because they can't make the N log N comparisons that it takes. It is hilarious to ask Copilot to sort a list in some order that isn't usual (like US states in alphabetical order) and see that it can't do it, and if you ask it if it can do it it will say it can, what probably it will get it right it says 'very high', then it goes ahead and tries and gets it wrong and if it is unlucky it even duplicates or drops items.
  - jqpabc123 2y ago
    MS is on a mission to prove to large numbers of people that AI is not the next big revolution in computing. Copilot has all the potential Cortana had.
    - PaulHoule 2y ago
      For me Copilot is at the plateau of productivity. When I'm trying to figure out how to test for certain things with Mockito it gives me working examples faster than I'd figure out if it was me and my lonesome with the manual. If I don't know why something works it often gives a correct answer. It does take some critical thinking to use it successfully, but the same is true for StackOverflow. I've also gotten good advice on subjects such as voice acting. Trying to knock out quick prototypes with other code assistants I'd say I have a 50% success rate. For some circumscribed tasks I've had good luck with existing code bases but if the problem is something like "figure out how somebody managed to download a file they shouldn't have downloaded and make sure it never happens again" you'd be a fool to even try using one.

al2o3cr 2y ago

Tried this with Gemini and it face-planted right at the start by mis-parsing the position of the black king and pawn :(

freediver 2y ago

Also check https://github.com/kagisearch/llm-chess-puzzles https://github.com/kagisearch/llm-chess-puzzles