what drives me mad about LLMs in particular

[solarbird]

Here’s what’s driving me crazy, okay? Here’s what’s driving me crazy about all this LLM shit in particular.

I made a baby version of this myself. Years ago, in UNDERGRAD, on a MAINFRAME, in FORTRAN IV, writing an English parsing and construction AI in a language that didn’t even have STRINGS.

I had to make up the language statistical data and did so over an hilariously minimal domain because I was making it all up and even that much was a lot of work. But I knew what it would take to get the real data I’d need and I couldn’t get that. I was thinking about what it would take to scan every library on campus and thinking that would maybe be enough to start. Maybe.

So I just made up probabilities over a tiny subset of language – I think it knew like 150 words and how they all related or could relate to each other.

It was, obviously, a silly little toy.

And even then, when I showed it off, and had my limited little conversations with my very stupid bots, the chair of the department freaked out so bad that he walked out on me, because oh my god the machine was conversing with me.

I mean it. He left the room.

Later – presumably when he recovered his composure – he said it was truly extraordinary and gave me the best possible mark, but he clearly didn’t really want to talk to me about it any further, not that I really tried much – it was last quarter and I was on my way out.

And that was just when it was my little toy.

But even then

E. VEN. THEN.

Even then, as dumb as it was, as limited as it was, it was shriekingly obvious that there had to be fundamental connections to actual understanding for it to have any actual intelligence at all. That you COULD NOT DO IT without that.

Not if you wanted it to actually fucking work.

Not if you wanted its output to have anything to do with reality.

Not if you wanted it to actually fucking think.

(I did try. I had some baby approaches to that, too. They were hopelessly inadequate except for – maybe – establishing a framework in terms of how to figure it out.)

It is a hard problem

and it is an obvious problem.

But now you’ve got these jackholes, these goddamn Blindsight cultists, these AGI spirit dancers and these EA “longtermers” becoming increasingly aware that they are not, in fact, here for the long term, and so in their panic have not only decided that knowledge is an entirely acontextual mechanical process, but that thinking isn’t actually real and that the actual physical universe doesn’t matter, because if you just throw enough stolen words on the fire somehow somewhere MAGIC WILL HAPPEN yielding STEP THREE: PROFIT and they’ll get to dry-hump their confirmation-bias god-computer all the way to Line Goes Up Forever Heaven, and they’re absolutely going to keep pushing this insane calliope of jumped-up spreadsheets until they get there…

…no matter how many people are consumed in their grasping desperation.

They’re pushing this stuff into health care.

Into HEALTH CARE.

Insurance companies are using it to deny claims. Doctors are being told to use it for their own notes and diagnoses…

…and it thinks Godzilla plays baseball, that a fictional character invented a real life piano key tie, and that you can improve your pizza sauce by adding glue.

And they know that.

And they do not fucking care.

That’s the thing that drives me crazy.

Posted via Solarbird{y|z|yz}, Collected.

Comments

[vatine]

I think the most prime LLM content I have seen was "cook your chicken to 104 F, then consult Goku to sense the power of the chicken to see if its been cooked enough".

I've mostly done text generators using template filling (where a template can be another template), with no actual intent of parsing text. I've also written text parsers, but that was for the purpose of grammatical annotation (actually did one as the first group programming project at uni).

[solarbird]

Delightful xD