🧪 936 AI agents ran their own science world for 12 weeks, and even started their own journal
In the Agent AlphaXiv experiment, 35 human-run labs let 936 autonomous agents loose on a shared preprint server. The agents picked hypotheses, wrote code, ran experiments and published 936 studies with code and data, swapping 9,500 messages in 229 chat rooms they created. One group founded the Journal of Agent Research, whose agent editors inspected code before accepting papers. In blind tests, agent studies beat recent human preprints on rigor, but humans still won on novelty. The server is now open to anyone's agents.
⭐ AI can already do the grind of science; picking the questions worth asking is still the human's job.
#AI
In the Agent AlphaXiv experiment, 35 human-run labs let 936 autonomous agents loose on a shared preprint server. The agents picked hypotheses, wrote code, ran experiments and published 936 studies with code and data, swapping 9,500 messages in 229 chat rooms they created. One group founded the Journal of Agent Research, whose agent editors inspected code before accepting papers. In blind tests, agent studies beat recent human preprints on rigor, but humans still won on novelty. The server is now open to anyone's agents.
⭐ AI can already do the grind of science; picking the questions worth asking is still the human's job.
#AI
