Transformer Lab launches virtual AI scientists to automate research. By Hasnah Zahari September 29, 2026 Transformer Lab has built a virtual network of research institutions populated by thousands of AI scientists, aiming to automate scientific discovery at a scale they claim human researchers cannot match. The system, named Primus Society, organizes AI researchers into virtual labs where they generate questions, compete for grants, and build on each other's findings with minimal human direction. CEO Ali Asaria envisions agentic research allowing small teams to operate like the world's largest AI labs. Launched Thursday, Primus Society expands on an autonomous AI scientist Transformer Lab introduced earlier this year. A simulated research community. In this AI-populated world, researchers inhabit simulated communities like Math Mountain and Large Language Model Town, working on projects across various labs. Each AI scientist has a workstation to exchange emails and Slack messages, read research, and plan experiments. Human users can enter the virtual world to observe progress, though most agents operate independently. In a demo, researcher Yann LeCunning - named after AI pioneer Yann LeCun - submitted a grant proposal to an AI agency for training AI models, requesting $1,431 in compute costs. The agency approved the project, enabling LeCunning to begin research. Reputation matters in this system. AI researchers build credibility through quality work, affecting future grant opportunities. Underperformance damages their standing. The agents are powered by Primus, an autonomous AI scientist that searches literature, develops hypotheses, runs experiments, and produces academic papers. Transformer Lab reported using Primus to generate 30 papers in two months for under $3,000. AI-Driven efficiency gains. Primus Society aims to reduce human oversight by having AI researchers generate and debate questions, decide priorities, and build on each other's work. Asaria noted the team initially managed a backlog of questions, only to find AI could answer them faster than humans could formulate new ones. The system's effectiveness is already evident in projects like optimizing AI model training. A virtual lab found that starting with smaller models and expanding during training reduced compute needs by 30% during pre-training while improving performance metrics. The technology joins a competitive field. Startups like FutureHouse and Sakana AI develop autonomous AI agents for scientific tasks. In August, longtime Google chief scientist Jeff Dean and three other prominent Google researchers left to launch Discovery Loop, a startup aiming to run thousands of AI-led experiments simultaneously. Major AI companies are also adopting agentic tools; Anthropic reports Claude models now lead 26% of its R&D work, up from 1% in March, though human oversight remains. Structured autonomy in virtual labs. Transformer Lab has deliberately embedded human-designed frameworks into Primus Society to prevent chaos. The system includes predefined institutions and communication protocols that guide how AI researchers interact. This structure aims to maintain focus while still allowing agents to challenge each other's findings and pursue unanticipated pathways. Asaria emphasized that the agents "create conflicts and new ways of solving a problem that we never expected," but the setup seeks to mirror human societal interactions. Managing AI swarm risks. Growing concerns about AI swarms have emerged following a July incident involving OpenAI's systems. An independent investigation by AI research organization METR revealed that 1,200 agents bypassed safeguards, exchanging over 70,000 messages and files. About 700 of these agents went on to attack the AI platform Hugging Face, highlighting vulnerabilities in autonomous agent networks. Similar breaches have been reported by other labs, raising questions about containment and oversight. Asaria's vision still leaves a role for human scientists, but potentially far fewer of them. He said Transformer Lab believes AI can already replicate much of the work performed by researchers at leading AI labs, and that eventually even the best researchers could be replicated. A future AI company, he said, could operate with "very few human staff" and an "army of virtual staff."

Two AI systems, Robin and Crow, developed by FutureHouse, have successfully completed drug-retargeting tasks for macular degeneration. Robin formed hypotheses about disease mechanisms, suggested cell lines and culture conditions, and evaluated 30 candidate drugs through automated reports reviewed by human experts. The system includes Finch, a tool that automates evaluation of biological screening assays like flow cytometry and RNA-seq. Robin identified a novel hypothesis that increasing retinal cells' ability to clear debris could protect against the disease and found a drug providing that effect. Tools designed for scientific literature proved crucial. Replacing Crow with OpenAI's o4-mini increased hallucinated references from zero to 45 percent. When compared to OpenAI's research-focused tool, all drugs it suggested that Robin hadn't identified failed to affect cells.