AI researcher fed paper outlines to frontier models to test end-to-end research
andrewgwils · x · 2026-09-06
Part 2/4 of Andrew White's thread: he usually starts papers by creating section headings and bullet lists, so as an experiment he gave this outline to the best frontier models and asked them to complete the paper, testing end-to-end scientific research ability.
Related event: Frontier Models Shockingly Bad at End-to-End Science, Scholar Finds(2 posts)→
More from AGI Musings
- It Talks Like a Human, Users Get Attached — Then Give It Memory Bugs — BLUECOW009 · 2026-09-06
- Creators Reveal Paid AI Doomer Outreach; Guardian Article's 6 Participants All Donor-Funded — beffjezos · 2026-09-06
- Extropic's Verdon: AI shouldn't be centrally controlled, mix closed and open source — beffjezos · 2026-09-06
- Google uses long-running agentic loops to recursively evaluate and refine its models — VraserX · 2026-09-06
- Designer: AI has made every app feel like a minimum-effort build — michalmalewicz · 2026-09-06
- Where are the breakout solo AI hedge funds? They happen in silence, says investor — nickbaumann_ · 2026-09-06