4 Comments
User's avatar
Simon Murray SME's avatar

Its secrets of the Mimic or even The Puppet Master yet images are a fail on lots of AI anyway

bouncy legs's avatar

There's a contradiction running through this. You correctly say LLMs are prediction-based rather than truth-based, then immediately: 'so it lies.' Lying requires knowing something's false and intending to mislead. Models have neither. 'Fabricates,' 'invents,' 'concocter,' same problem. All smuggle in a mind that isn't there.

The drunk in the pub is a good line but it's the same move in a different costume. It frames the output as coming from someone with a perspective, even a bad one. There's no one in there. Every 'the AI lied' framing quietly shifts accountability off the training pipeline, the reward design, the rater guidelines, and onto a fictional agent. That's where the real story is, and where the accountability sits.

Englishman in Switzerland's avatar

I'm thinking of Rumsfeld's known-unknowns and unknown-unknowns. AI seems useful when we know a bit about a subject, but dangerous when we use it in a topic we know nothing about. Then it can really fool us .

Jim McNeill's avatar

LLM’s are like an over confident drunk slapping you on the back. Which is why my articles are like they are . . .