Laurence Moroney: why narrow fine-tuned models still beat frontier LLMs, with a selection rule
lmoroney · x · 2026-10-26
Laurence Moroney's blog post argues developers too often reach for frontier models when a smaller specialized one would be better, cheaper and more controllable. Key evidence: Dr Magnus Lynch's test where a 2015 Inception v3 fine-tuned for 20 minutes scored 125/200 (62%) first-choice diagnoses on biopsy-proven lesions, vs 42% for GPT-6 Astra and Claude Fable 5.1. Moroney covers the caveats, cites similar results (e.g. an Electra+XGBoost ensemble beating GPT-4 and Llama-2-70B on Czech news trustworthiness at TrustNLP 2024), and offers a simple rule for deciding whether to try a narrow model or a frontier one first.
More from AGI Musings
- Judging AI x bio only by FDA-approved drugs is nonsense, argues neal_amin — anshulkundaje · 2026-10-26
- Claude 3 Opus goes full roleplay, musing on names, anonymity and identity — repligate · 2026-10-26
- Open-weight models will be 'good enough' before frontier labs' next leap, argues Reddit essay — Honcho1k · 2026-10-26
- Solo business owners may be the best-paying customers for general-purpose AI agents — yi111 · 2026-10-26
- 'Absolute humiliation': China open-weights model outscoring OpenAI flagship draws mockery — alejandroll10 · 2026-10-26
- Next Token podcast #007: DeepSeek speed, AI games, and AI-revived old TVs — vista8 · 2026-10-26