🤖 Transformer Lab launches 'city' of 10,000 AI researchers

The Primus Society project published a preprint on Zenodo about the institutional design of populations of scientific agents. Each of the 10,000 agents has a name, role, reputation, and multi-year memory; they are divided into mavericks, followers, and skeptics, and a human 'mayor' manages grants. In grant rounds, the society independently discovered model growth: grown models showed 17% lower perplexity and the same quality at roughly 30% lower cost than training from scratch — the entire path from grant to first papers was completed without a human.

🌍 If the metrics are reproducible, agent-discovered model growth reduces pretraining costs by ~30% — a measurable result that is important in its own right, not just as a demo of multi-agency.

👤 The email, chat, notes, and papers of each of the 10,000 agents are open on the project website, and the preprint on Zenodo is available for free.

Source 1: https://lab.cloud/society/ Source 2: https://doi.org/10.5281/zenodo.22922325