Topol vs Derya: Can GPT-4 Skill-Loss Findings Generalize to Newer Models?
Cardiologist and AI-in-medicine commentator Eric Topol and immunologist/AI commentator Derya Unutmaz (@DeryaTR) got into a heated X debate over a study examining whether doctors using AI still achieve sustained upskilling. Topol noted the study's key goal was to test whether using current frontier models produces lasting skill gains, but no one in the ensuing argument offered any explanation or evidence on that point. The crux: the study's conclusions rest on GPT-4—can they be extrapolated to today's or future models?
Confirmed
- Both sides' positions are clear: Derya Unutmaz firmly opposed Topol, arguing current models have advanced enormously compared with GPT-4 three years ago, that GPT-4-based research is outdated and cannot be extrapolated to future models like GPT-6, and that journal review delays mean the models in published papers lag far behind the LLMs actually available
- Topol's response: until comparable studies are done with future models (e.g., GPT-6, Opus 5.5), findings based on GPT-4 still hold and remain applicable
- Topol also repeatedly challenged critics to provide evidence of "sustained skill gains from frontier models," saying no one had done so
Unconfirmed
- The specific design, samples, and data details of the study under debate were not disclosed in the posts, so they cannot be verified
- Whether current models have actually solved the skill-decay problem was asserted by neither side with evidence—it was purely a clash of positions
Why it matters
- The debate touches a core issue for AI in medicine: if relying on AI erodes doctors' own skill development, and whether that effect changes across model generations remains unsettled, it will directly shape how medical schools and hospitals regulate AI-assisted care. The mismatch between journal review cycles and model iteration speed also makes "are these findings already obsolete?" a broader methodological controversy across AI research
2026-10-27 ~ 2026-10-27 · 5 related posts
Primary sources
- [source] Experts Clash: Today's Models Dwarf GPT-4, But Journal Delays Lock Out New LLMs — DeryaTR_ · 2026-10-27
- Eric Topol challenges frontier-model study: where's proof of sustained upskilling? — EricTopol · 2026-10-27
- Eric Topol and Derya Unutmaz clash over extrapolating GPT-4 study findings to frontier models — DeryaTR_ · 2026-10-27