A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis
The lesions come from MILK10k, a public dataset released through the International Skin Imaging Collaboration in 2025, in which each lesion is photographed twice: a clinical close-up and a dermoscopic image taken through the illuminated magnifying lens dermatologists use. I drew 200 lesions at random, all confirmed by histopathology:
…
You can try the CNN yourself. Remarkably the model is small enough to run entirely in a phone browser using WebGPU. Nothing is uploaded, and its random sample button draws from these same 200 lesions.
This again is part of a deluge of stuff about AI I get the fatigue but this is a wonderful specific example you can play with yourself that demonstrates the headlines that “AI is better at X than humans” should be ALWAYS treated with extreme skepticism, not just for the validity of the tests but also the validity of the basic framing and messaging surrounding results that “confirm” AI is superior.
This article/test also demonstrates that the basic technology behind “AI” really isn’t that new, which means the implications really aren’t that new either…
