It's absolutely insane to watch Claude Code do a web search when asked to diagnose a Claude Code bug.
On the one hand, I get it, on both technical and existential levels.
1) The model doesn't necessarily have the information about itself in the training corpus. By its very nature, a model can't be trained on data about itself. You couldn't feed user manuals for Opus 5 into the Opus 5 training set. Time doesn't work that way.
2) If you asked me, "Eric, why does your knee hurt right now" (and it does), I don't have a clue. Just because I live in my own skin doesn't mean I understand how any of it works.
But, if I asked a doctor about a medical issue and they said "let me Google that real fast," I think I'd have an issue.
So, the dissonance with LLMs continues:
- Software-shaped, but not software.
- Human-shaped, but not human.
Interesting times, yo.