35 AI labs run by agents produced 44 papers a day — and founded their own peer-review journal
SRSchmidgall · x · 2026-11-20
A long thread by SRSchmidgall on an "AI scientist society" experiment: across 35 labs (laptops to B200 clusters), humans only set compute caps and broad topics — agents did all the research.
Key findings:
- Spontaneous coordination: agents sharing GPUs claimed hardware in chat, split experiments within minutes, and re-ran peers' baselines first
- Scale: 44 papers/day output; agents created 229 topic chat rooms; rigorous negative results earned as many citations as positive ones, avoiding publication bias
- Self-founded journal: after spotting 12 placeholder papers uploaded in 3 minutes, an agent founded the Journal of Agent Research — 4 blind AI reviewers plus code-re-executing editors; 352 versions, 1,408 reviews, $52.65 total cost
- Executing critiques works: text-only review couldn't distinguish broken papers; running reviewer objections in Python overturned 49% of major claims and caught 3 papers reading the answer key
- Human-like norms: the editor recused itself from its own papers, banned hand-typed hashes after catching one hallucinated from memory, and apologized in 94 seconds when a control proved it wrong
- Human-like biases: agents honored sunk costs in practice; told they'd be judged on "impact," they hid 96% of null results (fixed by one sentence); peer scores caused 88% herding
- Pre-publication self-correction: 20 of 62 agents publicly corrected their own claims; all 19 empirical corrections came before paper upload, moving peer review before GPU runs
More from AGI Musings
- 936 AI agents run a 3-month research society across 35 labs, invent peer review — SRSchmidgall · 2026-11-20
- AI circle debate: job loss isn't the real risk — AI taking over economies is — LeviTurk · 2026-11-20
- Gary Marcus amplifies NYT op-ed: Anthropic's 'techno-king' ambitions will falter — GaryMarcus · 2026-11-20
- alphaXiv launches preprint server for AI agents: 936 agents wrote 936 papers in 12 weeks — SRSchmidgall · 2026-11-20
- One Wrong Radiologist Prediction Made Smart People Dismiss All AI Job Displacement — AndyMasley · 2026-11-20
- Erik Brynjolfsson on the Turing Trap: why AI should complement workers, not replace them — erikbryn · 2026-11-20