Raw LLM Responses

With respect, Dr. Fry, the "improvisational actor" metaphor is doing a lot of heavy lifting here and I think it's leading people astray. An LLM doesn't "roleplay" — it performs conditional probability estimation over token sequences. There's no actor, no character, no improvisation. Those are anthropomorphic framings that reinforce exactly the kind of confused thinking about these systems that someone in your position should be working to correct. Telling people to "be polite" because it shifts the output is technically adjacent to true — prompt phrasing does influence token probabilities — but wrapping it in theatrical metaphor for a 6-million-view audience is the kind of imprecision that breeds misconception. People walk away thinking there's something in there that appreciates their manners, when what's actually happening is statistics. You hold a Cambridge professorship in the public understanding of mathematics. The public understanding of mathematics is not served by replacing one bad metaphor (AI as encyclopedia) with another (AI as actor). It's served by helping people understand that these are statistical models — powerful, useful, and worth engaging with seriously — but not entities that respond to social cues. They respond to distributional shifts in input sequences. The bar for communicators scales with the platform. At 6 million views, "close enough" isn't good enough. https://www.perplexity.ai/computer/tasks/calling-out-anti-ai-bigotry-JX08wLHbTwC4.lPCHr46DA?view=thread