Comment by adamrezich | Hacker News Reader

Comment by adamrezich | Hacker News Reader

there's no way that the Bing language model actually "knows" (has "learned") the rules of chess (from training data) such that it can play a game, without ever breaking a rule—from what I understand, this would be a monumental achievement in the language model space.

without looking into this deeper, I have to assume they simply hooked in a chess AI somewhere in the stack. which is what I foresee a lot of going forward: software being presented as "an AI", which is just a text or voice input frontend to software that parses said input and routes it to internal subsystems, then wraps the subsystem with black-box pre-prompted language model text output, and emits it as output, creating the appearance of "AI", when in reality it's no such thing at all.

personally I find this boom of increasing amounts of fakery to improve the appearance of increased "I" in "AI" to be extremely exhausting, even moreso than the cryptocurrency boom ever was. it's all just tacking more and more stuff onto a black-box language model—as well as pre-altering the prompts which the user provides, in increasingly arcane linguistic ways that really are just the equivalent of throwing shit at the wall to see what sticks—all to deliver on the "promise" that language models seem to show when one is initially exposed to them.

the only saving grace with the "'AI' boom" as opposed to the "'crypto' boom" is, it's kinda fun to watch the big tech companies scramble to "outdo" each other, not even necessarily in terms of the quality of the software they're building, but in terms of faking it up enough to be presentable to the user such that they buy into the "AI" concept, which is, for now, seen as The Future Of Computing.

I have some good (or bad?) news for you then. Bing chat has no access to a Stockfish or other chess API. There is strong evidence that LLMs can learn and play games, including Chess. See this paper Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task by Kenneth Li et al.: https://arxiv.org/abs/2210.13382 or a more accessible write-up here: https://thegradient.pub/othello/

> Bing chat has no access to a Stockfish or other chess API.

is there any evidence to back this claim?

It's very unlikely. The moves it makes are not always good. Any public facing chess engine API or the API they chose would have been at least 'good'.

if that's the case then I agree it is unlikely, but then I would posit it is highly unlikely that it flawlessly follows the rules of chess. why is this interesting then? of course an LLM can create the appearance of "knowing how to play chess" that passes basic scrutiny, given that chess notation is a thing and all, but that's totally separate from having "learned" an ironclad "understanding" of the rules of a relatively basic turn-based board game.

if Microsoft wanted to tout that its "AI" could actually play chess, then the obvious thing to do would be to hook in a chess AI system as I suggested, as opposed to somehow engineering (or prompt "engineering") a better-able-to-follow-the-rules-of-chess language model. again, this is how nearly everything is going to be going forward in this space: creating a product that has the appearances of being able to do so many different things, but in reality is many different specific subsystems wrapped in a nice language model package.

I'm getting really tired of seeing all these "look at what I made the 'AI' do!!" posts that really aren't interesting at all when you get past the headline.

It's interesting because it does demonstrate an understanding of chess, just like othello gpt demonstrated an understanding of othello.

It follows the rules of chess and makes legal rules the vast majority of times. Making illegal moves occasionally doesn't mean you don't understand chess. Not if you learned chess in a self supervisory fashion. It's also a vast improvement over the chess performance of cGPT which makes legal moves on basically random chance and flies off the handle quickly.

It's also just plain interesting that a language model can pick up how to play these games from training on text in the manner they do. If you don't find the idea that LLMs are universal computation engines interesting then i don't know what to tell you.

One paper, making one claim, supported only by empirical results, in the current climate (where thousands of papers with similar claims are posted on preprint servers every month) is not "strong evidence". It's more like people saying stuff on social media.

There are tons of chess game transcripts online -- next token prediction with attention and rules/strategy discussions should be enough to get moderate skill in the Chess domain. It would be exceptionally unlikely that Bing would wire up a chess engine of all things into their production stack.

> There are tons of chess game transcripts online -- next token prediction with attention and rules/strategy discussions should be enough to get moderate skill in the Chess domain.

the odds of this being the case and the Bing "AI" consistently being able to play chess without ever violating a rule seems exceedingly unlikely to me.

> It would be exceptionally unlikely that Bing would wire up a chess engine of all things into their production stack.

why? pragmatically it makes the most sense. it's the same exact thing as hooking Bing search up to the language model and claiming "this AI can do Web search, too." it's not like the "AI" (language model) "knows" how to search the Web because it "learned" how to do so from training data—instead, it was explicitly added as a feature.

Microsoft is making a user-facing "AI" first and foremost, not an LLM. all that matters is how the user perceives the interaction with the service. the underlying LLM is useful only insofar as how it can make the user/computer interaction seem more magical and/or "natural".

Reply on news.ycombinator.com