✨the glitterati✨

I don't think I've ever seen anyone tell me why they think LLMs are intelligent but my laptop is not

and -- I think -- the only reason for that is LLMs can mimic human speech well and have a bunch of spooky hype propaganda floating around them

i otherwise don't see any difference

Would you claim my laptop is alive? Because it needs to consume to live, it can see and hear me, and it can communicate. It can do things I can't do and act in unpredictable ways

If you gave it to a medieval scribe and told them it was possessed by a demon, they'd probably believe you lol

Chess is also

- a popular game with a long written history (lots of data)
- has no randomness or other ambiguity (100% deterministic)
- has nerd cachet (so it's cool to your peers to make things for it lol)

I think that question isn't well formed, tbh?

Clearly we can create algorithms and such that *play* chess better than people, because people can't search enormous decision trees or have aggregate access to millions of games

I mean, sure, I don't disagree with any of that lol. It's neat as a technology! It's just not smart or alien intelligence, is all

My mildly spicy take is we would not be talking about "intelligence" if the models weren't also creepily good at mimicking human speech patterns. Our brains are crazy good at language and the AI stuff hijacks that

It's like how we're so hardwired to recognize faces that we see them in clouds

ML and friends are super useful tools -- the only sort of thing that really works on high-volume unpredictable data -- but they're not _magic_

I don't think it is, really? Running a statistical analysis to pull out interesting bits of data we can learn from is kind of what big data _does_.

I don't really see any real sea change difference between this and any other ML stuff, other than this is a novel way to do it

Anyways, as for proof, "machines are intelligent" falls under "extraordinary claims" for me.

You don't just need evidence *for*, you need evidence that it can't be explained by *anything else*

It's like UFO videos. The proper default stance is the null hypothesis: it's some ordinary object

I think about that story a lot

People really believed he could do math. It took someone methodically tested things like "does it work with strangers" (yes), "does it work if he can't see the asker?" (no) or "does it work if the asker doesn't know the answer" (no) to come to the real conclusion

It's a trope in AI, but you know the story of Clever Hans? He was a horse that could do math. Super famous.

Well, he couldn't actually do math, he was really good at reading the humans and our involuntary responses when we see the right answer

But it *looked like* he could do math

And in the paper, it's not like the LLM up and decided to be the best chess player

If I understand it correctly, they used the vector properties of LLMs to search for paths that seemed "useful" with some criteria they defined

Because that's not really what that paper is about?

The reason why computers can beat humans at chess is because they can calculate huge decision trees and choose the move with the best probability of winning

A lot of software is like magic -- the professional slight of hand kind lol

It looks amazing and impossible but you look behind the scenes it's a lot of expertise making it happen plus some misdirection and UI sparkle

It's like... I dunno, building a Dyson sphere. Like, you can picture the thing existing and kiiiiind of how parts of it might work

but we'd have to figure out so much stuff and it might not even work

Oh no, not at all. I think we're really complicated machines, too.

And I think it's *technically* feasible to have real machine intelligence, because we exist, but I think people vastly underestimate the complexity.

Eh, agree to disagree lol.

I'm not an *absolute* cynic, but I'm far enough into my career to know that Sturgeon's Law applies to smart people stuff like research, too lol.

I think that's way preferable to the alternative of making science a rare thing to do! But you do have to sideeye things

What trips people up is LLMs are legitimately very good at mimicking human language

I don't see an "intelligence" when reading LLM output, I see a smoke and mirrors software trick lol

A complicated one, but still a trick

Using the statistical model to find interesting patterns is neat but I wouldn't call that "reasoning" by the colloquial standard

It is also neat that we taught sand and lightning to do math, but the computer doesn't "reason" either

Honestly? No. No more than any piece
of software.

If I'm understanding it correctly, the paper says they're using the vector properties of LLMs to find specific trees that are "meaningful" but also not used by humans

ahhhhhhhhhhhhh

something is understandable != something has meaning

markov chains are gonna blow your mind lol

The research question I'd _really_ like to see is in human cognition

My halfassed theory -- quarterassed, really-- is the real true believers have a broken theory of mind, and are used to taking people on face value, so accepting a machine is not that weird

I'm in the industry -- although working in security, not AI -- and it honestly does not seem like a serious question serious people are researching.

The field is like 60% grifter by volume, with the remaining split between actual people doing actual things and Weird Cult Stuff

Their methods of testing was ¯\_(ツ)_/¯ ("we asked four grandmasters what they thought")

So a neat idea but not, like, proof of the machine spirit here

But I think the idea is reasonable - 'find vectors that humans don't use', especially in something like chess

I also don't think that's "reasoning" in the human sense - they're searching a vector space and extrapolating, I think that's "reasoning" in the data science algorithmic approach sense 2/

My reaction to that is a not serious (1) yeah that's research published by Google DeepMind, of course _they_ think that lol and more seriously (2) that paper is fine but not strong

I'm reading it quickly through a kinda cynical frame of mind and reading papers properly takes time and context 1/

You're just Clever Hans-ing yourself lol

You ever see a kitten see themselves in a mirror and instantly poof at the strange cat? AI is like that for some people lol

They're just statistical models stuck to a giant training corpus with some extra code stapled on

They're ok if you need _an_ answer (i used them to give me related words to a concept) but not ok if you need _the_ answer