Fine-tuned 2015 CNN beats GPT-6 and Claude on skin lesion diagnosis, 62% vs 42%
lmoroney · x · 2026-10-26
Dermatologist Magnus Lynch fine-tuned a 2015 Inception v3 CNN for 20 minutes and tested it against GPT-6 Astra and Claude Fable 5.1 on 200 biopsy-proven skin lesions from the public MILK10k dataset. The old CNN got the first-choice diagnosis right 125/200 times (62%), vs 84 (42%) for GPT-6 Astra and 85 (42%) for Claude Fable 5.1. Caveats: small sample, all hard biopsy-level cases, and the CNN isn't state of the art — yet the pattern holds: fine-tuned narrow specialists can beat prompted frontier giants on fixed labeling tasks. Laurence Moroney's writeup covers when to still build narrow models and a simple rule for choosing which to try first.
More from Models
- Block unveils Dessa, a full-duplex speech model trained only on human conversation — daluoseo · 2026-10-26
- DeepSeek V4.1F tops Chinese models on FrontierMath Tier 4 at 46% — teortaxesTex · 2026-10-26
- Casado slams Claude's safety classifier for flagging 'fetching my own documents' as reasoning extraction — AccBalanced · 2026-10-26
- Yacine MTB: DeepSeek is 'undoubtedly' a frontier model — yacineMTB · 2026-10-26
- Single-neuron steering: injecting a "joyful neuron" visibly alters a model's self-description — Sauers_ · 2026-10-26
- MillionInt is spending lots of time thinking about human eval curves — MillionInt · 2026-10-26