Comment by Animats

"The hallucination problem is the difference between a clever toy and a system a hospital or a bank or a court can actually rely on. It is the whole ballgame for enterprise AI."

This is the big problem with "agentic" AI. If you let the AI system do anything important, it's going to screw up reasonably often, and screw up in an expensive way occasionally. The usual solution to this is to make the errors an externality - dump them on the consumer-grade end user or an employee. As Google Search puts it, at the end of each result, "AI can make mistakes, so double-check responses".

External checking, which Cringley is pushing, has potential for search type systems. It's not likely to help when there's no one source text that can be used as an authority for checking. It's not likely to help with systems that actually do something.

How's end to end neural net driving working out?