Stanford prof: LLM reviewers catch mistakes scientists keep missing
anshulkundaje · x · 2026-10-28
Stanford computational biology professor Anshul Kundaje pushed back on claims that AI-generated peer reviews are useless: after extensive testing on "idea hardening," he finds LLMs very good for comp bio — imperfect, but reliably surfacing issues many scientists miss or have missed. He concedes LLMs struggle where literature is sparse (echoing an AI-written review that repeated one erroneous point paragraph after paragraph), and notes older Gemini models were poor while "GPT Astra Ultra" has been excellent in his experience.
More from AGI Musings
- Andreessen: AI gives founders managerial superpowers, ending Burnham's managerial society — robleclerc · 2026-10-28
- signull: science is approaching a historic inversion where AI becomes the discoverer — signulll · 2026-10-28
- Data curation for qualitatively different models, and the dream of modular interpretability — teortaxesTex · 2026-10-28
- Researcher questions Dario: where's the evidence Anthropic's 'race to the top' is working? — adam_karvonen · 2026-10-28
- Ivy League-to-consulting elites in their 40s are the most AI-disoriented, turkey-curve edition — 4KTV · 2026-10-28
- Max Tegmark discusses AI manipulation and machine rights with psychology PhD — tegmark · 2026-10-28