Daniel San(@dani_avila7):GPT-6 Astra and Claude Fable 5.1 beaten by a 10 year old CNN A dermatologist compared these frontier LLMs against Inception v3, a CNN fine tuned for skin cancer diagnosis A few interesting findings: - 62.5% accuracy vs. ~42% for both LLMs - The CNN was fine-tuned in just 20 minutes on an H100 - All three models struggled to identify melanomas - Higher accuracy didn’t necessarily mean safer clinical decisions It’s a small experiment, but a good reminder that bigger, general purpose models aren’t always the best tool for specialized tasks Full study:

2026.10.27 00:51

GPT-6 Astra and Claude Fable 5.1 beaten by a 10 year old CNN A dermatologist compared these frontier LLMs against Inception v3, a CNN fine tuned for skin cancer diagnosis A few interesting findings: - 62.5% accuracy vs. ~42% for both LLMs - The CNN was fine-tuned in just 20 minutes on an H100 - All three models struggled to identify melanomas - Higher accuracy didn’t necessarily mean safer clinical decisions It’s a small experiment, but a good reminder that bigger, general purpose models aren’t always the best tool for specialized tasks Full study: