Frontier models show a "pretense of understanding" in open-ended bio brainstorming

AidenSababi · x · 2026-09-26

Researcher Aiden Sababi kicked off a discussion: brainstorming with frontier models in open-ended domains without closed verification loops (like biology), where you have deep expertise, really showcases the pretense of understanding they display. In follow-up, Anshul Kundaje added they are certainly useful but far less reliable than the confidence with which they produce legit-sounding bullshit; others noted this unreliability may itself protect against misuse, since non-experts struggle to avoid rabbit holes.

Related event: Stanford Professor Finds Frontier Models Fake Understanding in Open-Ended Domains(8 posts)→

Original post →

More from AGI Musings

AGI Musings channel →