Exactly this 👇 LLM are exactly what the name Large Language Models means. They model language. They construct sentences that sound plausible, even probable. They are not reproducing facts, but rather language that reads like facts. What makes them dangerous is that people don’t understand that. - ThreadSky

dollyjorgensen.bsky.social • 375 days ago

Exactly this 👇
LLM are exactly what the name Large Language Models means. They model language. They construct sentences that sound plausible, even probable. They are not reproducing facts, but rather language that reads like facts. What makes them dangerous is that people don’t understand that.

Reposted from Sonja Drimmer

Ppl need to stop referring to AI “hallucinations.” LLMs are maybe-most-likely-machines not tools for accurate summarization. It is HIGHLY LIKELY that I would have written an article titled “Manuscript Mediation & Reproduction of Authority” but I didn’t. That’s not hallucination; that’s probability.

Comments

rtr.tombob.com•375 days ago

Best summary I've read so far: "Here's a fragment of text. Tell me how this fragment might go on. According to your model of the statistics of human language, what words are likely to come next?"
From https://cacm.acm.org/magazines/2024/2/279533-talking-about-large-language-models/fulltext

thryse.com•375 days ago

LMMs don’t hallucinate, they *confabulate*

wheneverends.bsky.social•375 days ago

even the inventors thought they were reverse engineering the brain

took a class with Hinton and he legit thought machine learning would make psychology obsolete

petermiles.bsky.social•375 days ago

There's a similarity with Trump there.

Perhaps AI and (some) politicians are on a convergent evolutionary path...

andrask.bsky.social•374 days ago

This is also how the novel as a genre is supposed to have started: fictions that read like truth.

ellieprobably.bsky.social•375 days ago

I think that's the key. No one knows what the fuck is, allowing an industry to built on what it isn't. In reality, LLMs have a narrow slice of legitimate use cases, and none of what people are doing with them are those.

(And for those legit uses, we need a LOT more research to say they're safe.)

nug.monster•374 days ago

i think LLMs could be really helpful in customer service not to replace customer service PEOPLE, but to make it easier for them to serve customers because eg, it's been trained on the policy docs for a govt agency. But you have to know it doesn't hallucinate before you can let anyone use that

lessbadproblems.bsky.social•375 days ago

Ted Chiang's "Blurry JPEG" analogy helped me see how the errors will always be there, and the fact that it's now "infecting" the internet means it's starting to become a blurry copy of a blurry copy.

qichin.bsky.social•375 days ago

Garbage in garbage out but for the entire internet

Comments

Posting Rules

Reply