「好讀，但學不到東西」
陶哲軒說被 AI 過度打磨的證明「好讀，但學不到東西」。他講的是摩擦：作者卡在哪裡的痕跡，也正是別人要跟你學的東西。
#Research
947 Pull Requests, 18 Labels
I wanted to measure whether evidence packs make AI-authored code review faster and safer. Three days of feasibility work killed the experiment, and the reasons were more interesting than the result would have been.
59 Seconds to Avoid 47 Seconds
My master's thesis predicts a CNN's final accuracy without training it. Then I measured what the prediction costs, and on CIFAR-10 it was slower than just training the model.