Kevin BeaumontJun 12

KPMG issued a report citing all the transformational ways GenAI has transformed industry, it’s been widely cited.

One minor problem: it turns they used AI to write the report, and it made up all of the evidence.

KPMG have now withdrawn the report in full.

https://www.ft.com/content/b3828e92-4961-4b39-84f0-c42f33be3c3f

Show thread⚙️ Roy PhillipsJun 12
@GossiTheDog AI doesn't hallucinate. That is just a marketing ploy to try and suggest it is in some way conscious (it isn't and never will be). It is the people unthinkingly using it that are deluded.
Show threadHugh SimpsonJun 14
@sothach @GossiTheDog the sun doesn't really rise, it's the world that turns. Dough doesn't really rest, it's not animate. Houses, generally speaking, don't have bones - good or otherwise. Language must be a constant source of frustration for you
Show thread⚙️ Roy Phillips
@silasmariner @GossiTheDog Yeah, unlike George Orwell, I'm not so bothered about dead metaphors, like those you highlight.
But the "hallucination" of LLMs is still very much an example of a live metaphor, and as Orwell remarks: "if thought corrupts language, language can also corrupt thought."
Jun 15 at 12:20pmWeb
Show threadHugh Simpson6d ago
@sothach @GossiTheDog dunno if many people would agree that the sun rising is a dead metaphor. Personally I don't agree that language can even be 'corrupted' -- to me that sort of phrasing implies an extant 'purity' to language. 'Hallucinate' is a good metaphor, anyway IMO - it has less implied intentionality than 'makes shit up'. What would your preferred term be?
Show thread⚙️ Roy Phillips6d ago
@silasmariner @GossiTheDog "Erroneous" wouldn't work, as many prompts don't have a right or wrong response. How about: "LLMs are prone to emitting untrustworthy output"?
Show threadHugh Simpson6d ago
@sothach @GossiTheDog a statement of fact, but it doesn't play the same linguistic role. 'That's an [llm hallucination]' vs 'that's an [example of llms being prone to emit untrustworthy output]'. Hardly rolls off the tongue... Even 'bullshit' implies intentionality, and given the context in which language evolves it's hardly surprising. Indeed, even truth/falsity implies intentionality since that's the original basis for utterances having semantic meaning. I don't see you getting away from it...
Show threadHugh Simpson6d ago
@sothach @GossiTheDog I suppose 'nonsense' might do, but that doesn't quite carry the speciousness of it...
Show threadHugh Simpson6d ago

@sothach @GossiTheDog maybe just 'mirage'? The problem is finding one that applies to semantic weight but doesn't imply agency. 'Mirage' doesn't do that, and appropriately places all interpretive weight on the reader.

A propos, if two llms converse in a forest and nobody's around to see them, did anything mean anything?

Show thread⚙️ Roy Phillips5d ago
@silasmariner @GossiTheDog Mirage is good a one. It also has a sense that the observer is projecting meaning where there is none, like a shimmering pool of water in the desert to a thirsty wanderer.