If AI is unreliable, what is it good for?

[CarwilB]

(Smartphone ruler app anecdote)

This smartphone app isn't an example of artificial intelligence, just of a particular kind of scam: turning popular trust in computerized intelligence into monthly subscription revenue.

But this app's failure shows exactly what's at risk when we insert an unreliable into the work we're doing.

Tools should do the same thing every time. (That's what makes them tools.)

Large-language models appear to be the ultimate digital multi-tool (just type what you want them to do), but their internal workings make them very unreliable. Let me show you some simple examples of trying to deploy them on routine tasks.