10-year-old CNN beats frontier LLMs on skin cancer
The gist
A dermatologist reports a fine-tuned 2015 CNN beat two frontier LLMs on 200 skin lesions.
Narrow, task-specific models can still outperform general-purpose AI on medical image checks.
How it unfolded
- 2015Google first publishes Inception v3, the convolutional neural network later fine-tuned for this test.
- 2025The MILK10k public dataset is released through the International Skin Imaging Collaboration, with each lesion photographed both close-up and dermoscopically.
- Oct 25, 2026Dr Magnus Lynch publishes a post reporting that the fine-tuned Inception v3 beat GPT-6 Astra and Claude Fable 5.1 on 200 biopsy-proven skin lesions.