The Transformer Lab team — Ali Asaria, Deep Gandhi, and Tony Salmon — launched Primus Society, a “city” of 10,000 autonomous AI researchers, each with a name, role, reputation, and memory spanning years. The preprint “A Society of Researchers: Institutional Design for Populations of Autonomous Scientific Agents,” published on Zenodo on September 24, 2026, describes the project’s design and results. Agents are divided into three archetypes by risk tolerance: mavericks test unconventional ideas, followers turn others’ discoveries into solid results, and skeptics spend compute on attempts to disprove colleagues’ conclusions. In a series of grant rounds, the society’s labs independently discovered the model growth technique: grown models showed 17% lower perplexity than a model trained from scratch at the same budget and reached the same quality at roughly 30% lower cost. If these numbers are reproducible, the project is valuable not only as a multi-agent demo but also as a real mechanism for reducing pretraining costs.
What happened
The Transformer Lab team launched Primus Society — a simulation of a scientific community described as a “city” of 10,000 autonomous AI researchers, each with a name, role, reputation, and multi-year memory. The project’s institutional design was published on Zenodo on September 24, 2026, as the preprint “A Society of Researchers: Institutional Design for Populations of Autonomous Scientific Agents.” The only human in the system, the “mayor,” governs the society by issuing grants rather than assigning tasks. In a series of grant rounds, the society’s labs independently discovered the model growth technique: first a small model is trained, then new layers are inserted between “transferred” layers. Grown models showed 17% lower perplexity than a model trained from scratch at the same compute budget and reached the same quality at roughly 30% lower cost.
Context
Primus Society brings the economic logic of real science into agent orchestration: instead of a single goal and a task queue, there is purpose, peer review, reputation, and compute allocation through grants. Methodologically, the most interesting design element appears to be the “skeptics” archetype: dedicated agents that spend compute on attempts to disprove others’ conclusions embed falsification into the population’s architecture itself — this distinguishes it from classic multi-agent pipelines, where roles merely distribute work. Facts and interpretation should be separated: the facts are the preprint publication on Zenodo and the reported metrics; the interpretation is that these metrics reflect a sustained methodological advantage of grown models. The verifiable part of the story is the specific model growth technique and the share of conclusions that survived skeptics’ attacks, not the scale of the “city”: population size alone proves nothing about the agents’ capabilities.
Why this matters for the industry
For the industry, the main value is not the launch itself but three transferable patterns: grants as an agent orchestration primitive instead of a task queue, a transparent “workspace” for each agent that is open to observation, and role allocation by risk tolerance. Such an institutional layer — grants, archetypes, reputation, a shared results library — could, as the field develops, become part of corporate agent platforms on top of models and orchestration frameworks. The bet on model growth is separate: if independent groups confirm the claimed pretraining compute savings on other datasets and at least one other scale, the method will quickly move from demo to a tool for small and medium pretraining and change the cost base for custom models. Likely near-term consequences — a wave of independent reproduction attempts, the appearance of “grant-type” budget allocators in open agent frameworks as an alternative to task dispatchers, and observability tools that copy the agent’s “open desk” pattern. If no confirmations come, Primus Society will remain a reference for institutional design rather than a working tool.
Why this matters for users
Readers can open the project website and “sit down at the computer” of any of the 10,000 researchers: each agent’s email, chat, notes, and articles are transparent and available for viewing. The Zenodo preprint is freely available, so the institutional design can be studied from the primary source, without retellings. The practical benefit today is in study, not application: grant orchestration patterns, the open agent workspace, and role division by risk tolerance can already be built into one’s own agent systems. Those who train their own models should plan an inexpensive experiment to reproduce model growth at an equal compute budget — the technique description is simple enough to try on their own data.
What is still unknown / limitations
There are no independent replications of model growth yet, so the reported metrics should be considered the authors’ hypothesis rather than a ready-made pretraining technique. The phrasing “every grown build outperformed every from-scratch build” indirectly indicates multiple builds, but the available materials do not disclose the number of seeds, the datasets used, or the exact definition of comparison conditions. The comparison of grant orchestration with a centralized-task baseline is not quantitative. The preprint has not undergone external peer review. There is no product layer: no API, no pricing, no latency data.
Sources
- Primus Society | Transformer Lab
- A Society of Researchers: Institutional Design for Populations of Autonomous Scientific Agents (Zenodo preprint)
Author
Look at AI, editorial team