Post by @zzt@mas.to

when the kids come at you with “not all AI is bad you know”, it’s good to remember that John McCarthy, one of the people who came up with the term AI, was a eugenicist who also believed that mercury thermostats and other simple machines could have the same types of beliefs as people. Lisp, an otherwise very good family of programming languages, was advertised as being uniquely good for AI on the basis of that crank belief, alongside a bunch of noise about expert systems and the ability to process symbols (literally a hash in a nametable) being the key to AGI. the latter being proven false many times led to the bursting of the first AI bubble.

the AI that works might be called machine learning, but that still gives it too much credit. it’s statistical methods applied intentionally to problems with a limited scope. that’s it. there’s no magic here, and certainly no ability for a thermostat or a lever to believe as a human mind does.

8 Replies

17 Boosts

[object Object]

@zzt@mas.to yesterday

symbolic processing is pretty damn good if you’re writing programming language tools and compilers, and that’s why the early AI voices tried to claim high level languages as part of their field, until the shine wore off and compilers became normal technology.

expert systems… are just databases, actually. the shine wore off of that shit so hard that the grift collapsed and that popped the first AI bubble. Lisp alone isn’t even particularly good at expert systems, but Prolog is. and Prolog… isn’t magic at all, it’s a very easy deduction algorithm to understand and a set of rules for how symbols are bound to values.

and you can take the non-grift ideas from an expert system and come up with a better database, and people do that. you can even take Prolog’s syntax and have a better time working with data than you will in SQL.

but symbolic processing and expert systems were both grifts as marketed. they were simple algorithms dressed up as pure magic that would result in AGI soon.

[object Object]

@zzt@mas.to yesterday

don’t let this trick you into thinking there’s a worthwhile non-grift technology hiding inside of LLMs. history doesn’t always rhyme, and there were plenty of projects from the first AI bubble that were just pure grifts. see for example the work that branched off from SHRDLU that was supposed to make robotic spatial reasoning easy, and it just… didn’t. they admitted it was a grift for grant money. it got us better text adventures (combined with Lisp and some algorithms from expert systems at Infocom) and that’s pretty much it. dead ends are possible, and they happen more frequently than successes.

[object Object]

@zzt@mas.to yesterday

no I don’t think LLMs will result in better text adventures either, writing good interactive fiction is like writing a book and LLMs are fucking terrible at writing books

I don’t think it’s worth worrying about what good is at the core of the technology that’s currently fucking us. if we’re very lucky it’ll be a couple of optimizations to Postgres. but LLMs require damage on a mass scale in order to exist, otherwise they’re not Large. that’s why all of the “ethical” LLMs are still trained on, drumroll, huggingface data that was scraped without consent. the people working on this shit don’t consider violating consent and ignoring open source licenses to be unethical.

[object Object]

@zzt@mas.to yesterday

and this is why, by the way, whenever you hear someone excitedly tell you about how LLMs work algorithmically, your response is probably “wow, that’s really fucking boring”

like, we already had tokens from compilers and words being broken into multiple tokens is a pretty common thing in search, but boosters talk about that like it’s revolutionary (or when they’re trying to pretend the slopbot not being able to comprehend the spelling of the word strawberry isn’t an issue)

it’s all real boring math at its core, and most of it isn’t new

~21 more replies

[object Object]

@zzt@mas.to yesterday

“no you see LLMs do this because <extended inaccurate description of tokens>” wow! that sounds like it’s unfit for purpose and unfixable

[object Object]

@zzt@mas.to yesterday

whenever someone comes at you with the ridiculous claim that an LLM or a generative art model works just like a human brain and claims that somehow that’s been proven, keep your eye on the ball. does your brain process tokens in such a way that you consistently can’t spell words or do basic math? does your brain process tokens (words split up and converted into numbers) at all? no? amazing

it’s almost like artists don’t make art by denoising random noise into the shape recommended by a textual prompt either.

these are the same claims as people like McCarthy telling you that mercury thermostats have beliefs, that symbolic processing (again, looking up a hash in a nametable) or expert systems (a garden-variety database or Prolog) are close to being AGI. McCarthy, an alleged expert, claimed that with the right software a Lisp Machine from 1985 would magically become intelligent!

[object Object]

@zzt@mas.to yesterday

and like, we studied complexity so we didn’t fall into this trap, right? we study the complexity of both algorithms and brains so we don’t think that a mercury thermostat or an 80s vintage Lisp Machine can do cognition. cause if you think that a thermostat or a Lisp Machine or an LLM can think because you’re very impressed with their output, you’re a crank.

[object Object]

@zzt@mas.to yesterday

and I say this as someone who likes thermostats and Lisp Machines for reasons McCarthy would likely be utterly unable to comprehend: they both make for cozy environments

I fucking hate LLMs though