A London dermatologist tested 200 biopsy-proven skin lesions and found a small specialist model from 2015 named the right diagnosis 62% of the time, ahead of GPT-6 Astra and Claude Fable 5.1 at about 42% each. The catch matters as much as the score: naming a diagnosis is not safe triage, and at one preset threshold the old model sent home 18 of 30 melanomas. Do not ask a chatbot to diagnose a mole — see a dermatologist for any changing spot.
Hook: 62% beats 42%, and the winner is ten years old.
Context: London dermatologist Magnus Lynch tested GPT-6 Astra, Claude Fable 5.1, and a fine-tuned 2015 Inception v3 model on 200 biopsy-proven skin lesions, each model naming one of six diagnoses.
Stakes: Small specialists can beat giant generalists at narrow tasks, but naming a diagnosis is not safe triage: at one preset threshold the old model sent home 18 of 30 melanomas.
Payoff: Do not ask a chatbot to diagnose a mole; see a dermatologist for any changing spot, and send this to one person who checks symptoms with AI.
Watch next: Meta Wants $95 Billion in New Stock to Pay for AI
Instagram reel: https://www.instagram.com/p/De_PW2qAr3R/
Source: https://theaipin.com/science/ten-year-old-cnn-beats-frontier-llms-at-skin-cancer-diagnosis
Leave a Reply