A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

	A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis(lynchlab.net)
	586 points by sungam 19 hours ago \| 250 comments
	tl;dr: A dermatologist tested 200 biopsy-proven skin lesions against GPT-6 Astra, Claude Fable 5.1, and a fine-tuned Inception v3 CNN from 2015, finding the CNN correctly diagnosed 62% versus 42% for both LLMs. The CNN also outperformed on melanoma detection (13/30 vs. 8 and 11) and was dramatically better at identifying benign lesions safe to discharge. The LLMs showed strong bias toward favorite diagnoses (BCC for GPT, actinic keratosis for Claude), and when queried as patients, recommended seeing a doctor in nearly every case regardless of actual risk.
	HN Discussion: ↑Unsurprising that specialized models outperform general-purpose LLMs on narrow tasks ↓Comparison is unfair because LLMs lack additional patient context dermatologists use ↓The CNN is benchmaxxed on in-distribution data, making the comparison misleading ~LLMs have untapped diagnostic potential if given richer inputs like notes and patient dialogue ↑Finding reinforces using LLMs as tools to build specialized classifiers rather than as classifiers themselves