I remember when several months (a year ago?) when the news got out that gpt-3.5-turbo-papillion-grumpalumpgus could play chess around ~1600 elo. I was skeptical the apparent skill wasn’t just a hacked-on patch to stop folks from clowning on their models on xitter. Like if an LLM had just read the instructions of chess and started playing like a competent player, that would be genuinely impressive. But if what happened is they generated 10^12 synthetic games of chess played by stonk fish and used that to train the model- that ain’t an emergent ability, that’s just brute forcing chess. The fact that larger, open-source models that perform better on other benchmarks, still flail at chess is just a glaring red flag that something funky was going on w/ gpt-3.5-turbo-instruct to drive home the “eMeRgEnCe” narrative. I’d bet decent odds if you played with modified rules, (knights move a one space longer L shape, you cannot move a pawn 2 moves after it last moved, etc), gpt-3.5 would fuckin suck.

Edit: the author asks “why skill go down tho” on later models. Like isn’t it obvious? At that moment of time, chess skills weren’t a priority so the trillions of synthetic games weren’t included in the training? Like this isn’t that big of a mystery…? It’s not like other NN haven’t been trained to play chess…

Particularly hilarious at how thoroughly they’re missing the point. The fact that it suggests illegal moves at all means that no matter how good it’s openings are the scaling laws and emergent behaviors haven’t magicked up an internal model of the game of Chess or even the state of the chess board it’s working with. I feel like playing games is a particularly powerful example of this because the game rules provide a very clear structure to model and it’s very obvious when that model doesn’t exist.

Here are the results of these three models against Stockfish—a standard chess AI—on level 1, with a maximum of 0.01 seconds to make each move

I’m not a Chess person or familiar with Stockfish so take this with a grain of salt, but I found a few interesting things perusing the code / docs which I think makes useful context.

Skill Level

I assume “level” refers to Stockfish’s Skill Level option.

If I mathed right, Stockfish roughly estimates Skill Level 1 to be around 1445 ELO (source). However it says “This Elo rating has been calibrated at a time control of 60s+0.6s” so it may be significantly lower here.

Skill Level affects the search depth (appears to use depth of 1 at Skill Level 1). It also enables MultiPV 4 to compute the four best principle variations and randomly pick from them (more randomly at lower skill levels).

Move Time & Hardware

This is all independent of move time. This author used a move time of 10 milliseconds (for stockfish, no mention on how much time the LLMs got). … or at least they did if they accounted for the “Move Overhead” option defaulting to 10 milliseconds. If they left that at it’s default then 10ms - 10ms = 0ms so 🤷‍♀️.

There is also no information about the hardware or number of threads they ran this one, which I feel is important information.

Evaluation Function

After the game was over, I calculated the score after each turn in “centipawns” where a pawn is worth 100 points, and ±1500 indicates a win or loss.

Stockfish’s FAQ mentions that they have gone beyond centipawns for evaluating positions, because it’s strong enough that material advantage is much less relevant than it used to be. I assume it doesn’t really matter at level 1 with ~0 seconds to produce moves though.

Still since the author has Stockfish handy anyway, it’d be interesting to use it in it’s not handicapped form to evaluate who won.

LLMs sometimes struggle to give legal moves. In these experiments, I try 10 times and if there’s still no legal move, I just pick one at random.

uhh

@gerikson @BlueMonday1984 the only analysis of computer chess anybody needs https://youtu.be/DpXy041BIlA?si=a1vU3zmOWs8UqlSQ