It's interesting that while standard chess engines have essentially perfect "sight" of the board in both its present and future state, the AI models play like a human playing blindfolded for the first time. A lot of their moves make some kind of sense, but the mistakes are really glaring.
The fact that it can find ...Rxe3!! is amazing. Though it also played ...Qxh2?? and ...Rxe2?? so maybe there's some pattern recognition around capturing things that is causing both brilliancies and blunders.
One needs to be careful with chess "AI" engine word wielding. LLM based on PGN (only) and say A0 are entirely different world model premises. One is string sequecen bottom sampled world, The other is a multidimension chess piece space a cartesian product of 32 chess thingie planes (64 squares for each of the 32).. So it has space as part of the ambient formalism. While they did cheat and hardcoded the finite mobility rules (and other stuff i don't care to talk about, for fear of missing a critter), they could have let the things learn the rules too. My point being that spatial logic is directly accessible to training.
Contrary to the language or string (SAN move syllables, chain of them) world or even the FEN world, it is a complete problem, and math. can show RL will lead to ever-improving self-play batches trajectories (even if saturating....). Convergence, given engine architecture, finite constraints, and choices, to its best play.
I think the LLM is akin to opening repertoire, high-level game, or top chess engine, move by move long chains, imitation learning theory. The theory of learning from studying the sequences depth-first, I would hunch, is not about the full static spatial single position layout feature shaping the dynamic future possibilities also in future static spatial single position layouts of, say, experience learned "acceptability" as a plan objective. Things like that. That is a caricature, but the LLM is doing it.
I guess it might be an idea while doing this LLM examination as some kind of partial model of the human chess problem to compare it with what I consider better science-based AI, perhaps because it had a math formulation before it went to alchemical technology. (kidding about myself and the technology). LLM, SF(nnue), LC0 (a0). But I think LLM is hotter now. More urgent for shedding light beyond chess on how it affects or what it says about our irrational adoption of that technology all the way to the new dark age promise from the tech lords.
In fact, either in this article or another (I am jumping around not fully reading any of them now, but in the long run I will... over days), Jennifer Shahade (practicing writing names, a weakness of mine) is having similar point of view, maybe not in my convoluted wording.
It feels good to hear some wisdom about that technology, given the loud and asphyxiatic BS out there with it. (and still no scientific research at its foundation, why I "joke" with the word alchemy).
I think that it's similar to my rationale. Assessing their chess can help reveal weaknesses that may apply in other areas. Those weaknesses can be more obvious in a perfect information game. This is important since so many people are relying them more and more and even sometimes for tasks they're ill suited for.
I missed the video. But I find plenty of food here. I am glad to find some echoes to my past impressions, using chess is a good idea. The drosophila. for many things. of scientific curiosity.
I will keep reading here.. at my uncontrollable pace(more than one day). Thanks to you and Nate Solon for this surge of curiosity. Well, I just found out.
It's interesting that while standard chess engines have essentially perfect "sight" of the board in both its present and future state, the AI models play like a human playing blindfolded for the first time. A lot of their moves make some kind of sense, but the mistakes are really glaring.
The fact that it can find ...Rxe3!! is amazing. Though it also played ...Qxh2?? and ...Rxe2?? so maybe there's some pattern recognition around capturing things that is causing both brilliancies and blunders.
Broken clocks and all that, I suspect.
One needs to be careful with chess "AI" engine word wielding. LLM based on PGN (only) and say A0 are entirely different world model premises. One is string sequecen bottom sampled world, The other is a multidimension chess piece space a cartesian product of 32 chess thingie planes (64 squares for each of the 32).. So it has space as part of the ambient formalism. While they did cheat and hardcoded the finite mobility rules (and other stuff i don't care to talk about, for fear of missing a critter), they could have let the things learn the rules too. My point being that spatial logic is directly accessible to training.
Contrary to the language or string (SAN move syllables, chain of them) world or even the FEN world, it is a complete problem, and math. can show RL will lead to ever-improving self-play batches trajectories (even if saturating....). Convergence, given engine architecture, finite constraints, and choices, to its best play.
I think the LLM is akin to opening repertoire, high-level game, or top chess engine, move by move long chains, imitation learning theory. The theory of learning from studying the sequences depth-first, I would hunch, is not about the full static spatial single position layout feature shaping the dynamic future possibilities also in future static spatial single position layouts of, say, experience learned "acceptability" as a plan objective. Things like that. That is a caricature, but the LLM is doing it.
I guess it might be an idea while doing this LLM examination as some kind of partial model of the human chess problem to compare it with what I consider better science-based AI, perhaps because it had a math formulation before it went to alchemical technology. (kidding about myself and the technology). LLM, SF(nnue), LC0 (a0). But I think LLM is hotter now. More urgent for shedding light beyond chess on how it affects or what it says about our irrational adoption of that technology all the way to the new dark age promise from the tech lords.
In fact, either in this article or another (I am jumping around not fully reading any of them now, but in the long run I will... over days), Jennifer Shahade (practicing writing names, a weakness of mine) is having similar point of view, maybe not in my convoluted wording.
It feels good to hear some wisdom about that technology, given the loud and asphyxiatic BS out there with it. (and still no scientific research at its foundation, why I "joke" with the word alchemy).
"Yes, Qxh2 --- I get to play mate in one with an extra tempo!"
now that's a way to look at it!
I can't believe there was actual money and resources spent on a tournament held between entities that can't even be counted on to play legal moves.
I think that it's similar to my rationale. Assessing their chess can help reveal weaknesses that may apply in other areas. Those weaknesses can be more obvious in a perfect information game. This is important since so many people are relying them more and more and even sometimes for tasks they're ill suited for.
I missed the video. But I find plenty of food here. I am glad to find some echoes to my past impressions, using chess is a good idea. The drosophila. for many things. of scientific curiosity.
I will keep reading here.. at my uncontrollable pace(more than one day). Thanks to you and Nate Solon for this surge of curiosity. Well, I just found out.