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.

26 Replies

0 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.