That anecdote makes me wish we could make LLMs less compliant … I’m imagining some code wildly inappropriate to shoehorn a <canvas> into and GPT happily complyi

That anecdote makes me wish we could make LLMs less compliant … I’m imagining some code wildly inappropriate to shoehorn a <canvas> into and GPT happily complying, instead of responding “are you nuts? How would a <canvas> be useful to this method that validates a certificate?” Obviously it’s harmless here, but I can picture “evil genie” results when someone asks a question to an LLM entrusted with important privileges. Like if you said “Open ‘uranium containment’” meaning a file with that name, and the LLM is like “don’t want to argue with humans!” opens up the uranium containment doors An extreme and stupid example, obviously, but you get the idea. If I am to trust an LLM it ought to be able to say “That sounds like a stupid idea. Surely we have a misunderstanding.”