In defense of stupid questions
It's never been easier to get answers to things. You type in a question, some LLM runs a search somewhere and synthesizes a response based on the results, and you often get a pretty reasonable answer.
But LLMs are faulty and stochastic. The model the answering machine used may not be strong enough, or maybe it dealt with sparse search results, or maybe your question was too obscure -- and it could end up confidently synthesizing a completely incorrect answer.
The danger then of this pattern is when you -- the human operator -- get into the habit of complacency, of implicitly trusting the machine's responses without validation. The same mechanism that made it easy to get answers also made it easy to be stupid.
Years ago, work in software development involved the tedious hunt for answers: someone looking for technical, architectural, or tool-specific answers often ended up trawling Github issues, StackOverflow, Quora, or a tool's doc pages. During team briefings, we would often say: "ask away, there are no stupid questions."
I recently made the mistake of telling a coworker the age-of-AI version: "ask away, but with all your agents available there are stupid questions. only reserve the real questions for the humans."
See, I found that the trouble with frowning upon stupid questions is that people lose the ability to calibrate what's stupid and what's not depending on whom they're asking. We also don't learn what's worth a human's time versus what an agent can chew on. So we ask the agent everything, including the long-tail stuff like "where's the code for X?" that used to teach juniors how to hunt. And if the underlying data is sparse, the machine still answers like it's sure. That's the complacency trap.
To me, this means that the old tenet holds: we should keep on asking each other questions. Keep asking our fellow humans. There are no stupid questions.
Actually, maybe there still are. But it's worth asking them anyway.