When talking about "AI" there's a lot of chat about "hallucinations" but I think it is a bit misleading. Because it's _all_ hallucinations, some just accidentally true. https://tante.cc/2025/03/16/its-all-hallucinations/ - ThreadSky

tante.tldr.nettime.org.ap.brid.gy • 8 days ago

When talking about "AI" there's a lot of chat about "hallucinations" but I think it is a bit misleading. Because it's _all_ hallucinations, some just accidentally true.

https://tante.cc/2025/03/16/its-all-hallucinations/

Comments

mattjhodgkinson.scicomm.xyz.ap.brid.gy•8 days ago

@tante I'm quite LLM skeptical, but real-world grounding and recursion, which you allude to as RAG, will probably give better reliability and reasoning. This is not a new idea, e.g. https://arxiv.org/abs/1506.04366 is from 2015, but the language fluency of LLMs coupled with semantic reasoning / […]

kihbernetics.bsky.social•8 days ago

Humans and LLMs operate in two different (non-intersecting) domains.
Like all living beings, humans evolved in the #physical domain, while the LLMs were constructed to operate in the #language domain devised to communicate ideas between humans. Humans are LLMs' "transducers" for the physical world.😉

j2bryson.mastodon.social.ap.brid.gy•8 days ago

@tante exactly https://joanna-bryson.blogspot.com/2019/06/ai-works-like-dreaming.html

lauerhahn.sfba.social.ap.brid.gy•8 days ago

@tante Great piece, thank you for writing. This section strikes me as especially important to get out in the world:

"They don’t “know”, or “think” or “lie” or do any such things. They iteratively calculate the most probable set of words and characters based on the original data."

Kraemer-HB.mastodon.social.ap.brid.gy•7 days ago

@tante I suspect your analysis of the meaning of "hallucination" could work for the meaning of "truth", too.

tante.tldr.nettime.org.ap.brid.gy•7 days ago

@Kraemer_HB Truth is a very complex idea and one would have to agree on a definition of truth first which isn't as simple as one might think

Kraemer-HB.mastodon.social.ap.brid.gy•7 days ago

@tante That's true *scnr* but I'm not even sure if any of the existing definitions of truth are capable of dealing with the "truth" of AI outputs. Maybe Nietzsche: "truths are illusions about which one has forgotten that this is what they are".

williampietri.sfba.social.ap.brid.gy•8 days ago

@tante Totally agreed. A normal human speaker will monitor what they're saying and make sure it comes out right, which usually includes correspondence to reality. "AI" doesn't have that at all. But the chatbot form is designed to make people think otherwise. There's a giant boom built on a hole […]

Comments

Posting Rules

Reply