There is a post that’s circulating around the Lemmyhood where someone asks an LLM to solve the “there’s a goat, a wolf and a cabbage that need to cross a river” problem, and it returns grammatically correct logically impossible nonsense. I think this is instructive as to how LLMs work and how useless they really are.

Presented with a logic problem, it doesn’t attempt to solve any problems or apply logic. That it does is search through the sumtotal of all human communication, finds dozens if not hundreds of cases where this or a similar problem has been asked, and then averages the answers. Because answers might be phrased in different orders or different sentence structures, or some people published wrong or joke answers sometimes but it has no means to detect that, they get averaged in with equal weight and so the answer it puts out begins with “Take the wolf and the goat, leave the boat behind. Take the boat back.” It has a fascinating ability to output seemingly relevant and grammatically impeccable worthless noise. Just like everything I say.

The only compelling use case I’ve seen for these things is writing frameworks for fictional stories. There was an episode of the WAN Show back when LMG still existed where Linus gave ChatGPT a prompt to create a modern take on the premise of the movie Liar Liar. And it came up with an actually compelling outline, I’d go see the movie made out of that outline. Because it’s fictional, it doesn’t have to conform to reality.

I doubt it could write an entire acceptable movie script though, it would have gaping plot holes, would have no theme or cohesive narrative structure, but every individual line of dialog would make grammatical sense and some conversations might even seem coherent.

As a research or information gathering tool, they’re worse than useless because it has no way of detecting if information is up to date or obsolete, serious or farcical, correct or incorrect, it just averages them all together, basically on the same theory as the Poll The Audience lifeline on Who Wants To Be A Millionaire: Most of the crowd is almost always right. Except with this approach what happens is it will cite a completely fake made up paper and attribute it to a genuinely real scientist who works in the relevant field and allegedly published in a real reputable scientific journal. It looks right, it passes the sniff test. It’s also completely useless.

And that’s when they’re not throwing weird emotional tantrums.

The AI only fools us into thinking it’s intelligent because it picks the most likely text response based on what it’s read before. But often, the output is confidently wrong as it’s really just a parlor trick.

that’s basically what many humans do

i think AI still has really cool applications, it’s just that the vibe is getting destroyed by shitty companies putting it in everything, harvesting stealing data for it, the awful spam, and the built in restraints which make it act like you’re a child

The way that people use and trust these chat bots reminds me of stories about executives in the '80s climbing the corporate ladder using a Magic 8 Ball

As long as they stop trying to treat it like a hammer and literally everything else like a nail, I’ll take it. It’s nfts all over again.

“Any hole’s a goal” is a relationship goal, not a social one!