The limits of LLMs

A practical failure model for LLMs — what they can do alone, what needs evidence or tools, and what must stay under human control. Covers the one root limitation (likely text is not automatically true), why capability is jagged, where exact operations break, why knowledge is incomplete and stale, why long context doesn't guarantee recall, why instruction-following is probabilistic, the limits of long-horizon autonomy, the kinds of bias that appear, why benchmarks don't settle capability, which decisions not to delegate, the defense-in-depth that builders use, and a go/no-go checklist for putting an LLM in a workflow.