Thoughts on generative A.I
3 points by kghose
3 points by kghose
In 2000 if you told me that a computer would generate a novel coherent sentence, let alone several paragraphs of grammatically correct English I would have called you a liar.
The dissociated-press algorithm was published in 1972. Don Mitchell's markov was the basis for Mark V. Shaney and similar works in the early 80s. Racter was available for home computers by 1985. The bidirectional integrated interactive-fiction environment Inform 6 was published in 1996.
I do think that there's an interesting nuance here which might be steering your thoughts. We didn't have a reliable way to parse English until about 2013, described here. Chatbots are usually not used for generating arbitrary text, but for generating text conditioned on some context ("prompt"), which in full generality implies that the bot ought to be able to parse the prompt to answer questions like "how many 'r' are in this sentence?" or "what is the main verb of this sentence?" In this sense, although our generative capabilities have been slowly increasing since the 1960s, it is only recently that we have been able to complete the loop which makes the ELIZA effect so potent by fooling the user into thinking that the bot understands English by both reading and writing English.
It would seem that generative A.I. is proof by demonstration that thinking is the regurgitation [...] of thoughts we’ve experienced other express to us with some amount of novelty (randomness) sprinkled in.
What about reasoning or deliberation? The current frontier genAI models are poor at—or even incapable of—catching simple logical errors.
What is novel about randomness when it's not critically evaluated against what you know?
I'm expecting this to be slammed with "vibecoding" so no one sees it, and I suppose that is the sad state of lobste.rs today, but I would have felt comfortable sharing these philosophical and practical thoughts on generative AI with this crowd a few years ago, so for old time's sake, I'm posting it here.
Would be worth it for me to have just one person reply with how they see this historic time and how it's affected them.
There are possibly four articles here, (The one directly linked here, and the three that are linked from that landing article). Read as little as you want :)
Would be worth it for me to have just one person reply with how they see this historic time and how it's affected them.
Utterly batshit insane and rather badly.
My most recent exposure to confabulated LLM output (last week) was a document on installing and configuring Docker. I don't know anything about Docker, so it's hard for me to know what's correct and what was outright made up but plausible. I do not want to be responsible for the results in such a case and pushed back hard enough (basically "fuck no!") to force a course change with someone who is pro-LLM.
The UK postal scandal was bad enough, but I can see that as being a trivial mistake to what might be coming in the future as expertise is seen as worthless and more and more people stop thinking.
I'm expecting this to be slammed with "vibecoding" so no one sees it, and I suppose that is the sad state of lobste.rs today
I don't blame people for not wanting to see stories about generative AI, they can get pretty tiring. The vibecoding tag doesn't apply to this story, but I'd assume that even if it gets misapplied, (edit: lol) people who filter out vibecoding are going to be relatively likely not to want to read an article titled "Thoughts on generative A.I" anyways.
The tag situation is indeed pretty sad, though. We ought to have a more general generative-ai tag or something.
I hate the tag name. But setting aside the naming kerfuffle and focusing on how the tag is used, the tag itself is a good thing. Telling readers up front, "This is the topic and you have the choice of whether to interact," is a fundamentally empowering act.
(And I'm trying my hardest to word this politely, but when an author says, "I know this will get buried, but…" I have a knee-jerk reaction—it makes me want to avoid what they wrote.)
I didn’t even look at the tag. I just read the title: here’s yet another person’s “thoughts” on LLMs. Maybe I’m being cruel but I don’t generally care about anyone’s “thoughts” on this stuff any more.
That's pretty much my point! You should be empowered to hide stories you don't want to see. It's better to skip it than to be cruel.
We ought to have a more general generative-ai tag or something.
Please no, this would imply that generative-ai was on-topic, which apart from the narrow technical aspects on how it is produced, it is not.