2026-10-26 / Signal #7
A 2015 CNN beat GPT-6 Astra and Claude at melanoma
“The ten-year-old model kept the melanoma. The frontier models sent it home.”
The Story
Dermatologist Magnus Lynch ran 200 biopsy-proven lesions from MILK10k through GPT-6 Astra, Claude Fable 5.1, and a 10-year-old Inception v3 CNN fine-tuned for ~20 minutes. The CNN’s first-choice diagnosis was right 62% of the time vs ~42% for both frontier models. On one thin melanoma of the ear, both LLMs called it seborrhoeic keratosis and said discharge; the CNN caught it. The chatbots were more cautious overall, but still discharged some melanomas.
Why It Matters
The future-shock isn’t “AI doctors.” It’s that the shiny generalist discharged cancer and the dusty 2015 specialist didn’t. Builder utility plus a sick visual.
Evidence
Dr Magnus Lynch, Guy’s Hospital, lynchlab.net, Oct 25; large HN discussion
Sources
Daily scan: 2026-10-26