AI News

Transformer Lab presents Primus Society, a virtual community of 10,000 autonomous AI researchers built around the Primus system. Primus takes a research question, reads relevant literature, designs and runs experiments on GPUs, and writes a paper linking claims to the experiments that produced them. The researchers are given different roles and risk tolerances: mavericks explore unconventional ideas, followers extend promising results, and skeptics try to reproduce or break them. The system organizes them through labs, grants, peer review, conferences, a shared research library, and a finite compute budget, with a human opening each research call and setting its question, budget, and deadline. Researchers persist across calls, keep records, sign their work, and build reputations from experimental outcomes. The page reports a model-pretraining result in which growing a model during training achieved 17% lower perplexity than a same-budget model trained from scratch and reached the same quality with about 30% less compute; it says every grown run beat every from-scratch run in that comparison. Related papers report that timing and compute accounting affect the result, while width growth showed no advantage in some experiments and the reason for the observed benefit remains unresolved. The site says the experiments ran from call opening to initial papers without human intervention and that subsequent calls are investigating the remaining questions.