Stanford prof: DNA model GPN-STAR defies bitter lesson narrative, ignored for a year
anshulkundaje · x · 2026-09-10
Stanford professor Anshul Kundaje argues GPN-STAR, a DNA language model (DNALM), "goes completely against the entire narrative of bitter lesson, no inductive biases, scaling laws" — yet performs spectacularly for its class.
He notes the preprint has been out for a year while VCs, AI/GPU CEOs and influencers stay silent about it, preferring to hype behemoth models measured by parameters and GPU hours. His point: the field systematically overlooks small, inductive-bias-heavy approaches that actually work.
More from AGI Musings
- Developer says AI safety community's grey goo fears undermine its credibility — Darpinian · 2026-09-10
- Ex-OpenAI/Anthropic pretraining researcher resigns, says labs are gambling on superintelligence — Miles_Brundage · 2026-09-10
- As "AI will kill us all" headlines spread, observers warn doom-mongering could fuel real-world panic — TheMoonMidas · 2026-09-10
- Why AI can tackle Millennium Prize Problems but still gets everyday questions wrong — davidpattersonx · 2026-09-10
- AI x-risk debate: the real danger is people with models, not the models — arian_ghashghai · 2026-09-10
- Scientists grapple with AI agent fleets reshaping research and peer review — vykthur · 2026-09-10