TORONTO — Transformer Lab has built a virtual network of research institutions populated by thousands of AI scientists, in an effort to automate scientific discovery on a scale the company says would be impossible with human researchers alone.

The Toronto-based startup launched Primus Society on Thursday, an expansion of the autonomous AI scientist it unveiled this summer. The new tool is designed to co-ordinate masses of AI researchers to collectively pursue scientific problems with limited human direction. The AI agents can explore different ideas, challenge each other’s findings and compete for the computing power needed to run experiments. Transformer Lab ultimately hopes the approach will let a relatively small team of people oversee research operations far larger than would be possible with an all-human staff.

In the AI-populated world, agentic researchers inhabit simulated communities that go by names like Math Mountain and Large Language Model Town, where they work on various research projects across different labs. Researchers have their own workstations, where they exchange emails and Slack messages with colleagues, read research and plan experiments. Human users can enter the virtual world to see what individual AI scientists and labs are working on, but for the most part, the agents go about their business on their own.

In a demo of the tool shown to The Logic, one researcher called Yann LeCunning—a riff on renowned AI researcher Yann LeCun—can be seen working at the Bitter Institute. The scientist exchanges emails with an AI granting agency inviting proposals for a research project about training AI models. LeCunning responds with a pitch and a budget request of $1,431, representing the cost of compute required to do the research—real money Transformer spends to run the experiment. The granting agency accepts the pitch, letting LeCunning begin researching the topic.

As in the real world, agentic researchers build reputations based on the quality of their work, which can improve their chances of securing future grants. If they overpromise and underdeliver, their standing suffers.

The researchers are powered by Primus, the autonomous AI scientist Transformer Lab launched in August. Given a research question, Primus can search scientific literature, develop a hypothesis, run digital experiments and produce an academic-style paper. Transformer Lab said this summer that it had used Primus to produce 30 papers in two months for less than $3,000.

Primus Society is meant to take that a step further by reducing how often humans have to decide what the AI should investigate. Transformer Lab CEO Ali Asaria said his team initially kept a backlog of questions for Primus to work through, only to find that the AI could answer them faster than the team could think of new ones. Primus Society is designed to have AI researchers generate and debate their own questions, decide which ideas are worth pursuing and build on one another’s work, allowing research to continue with less human direction.

Asaria said the company is already seeing useful results from the new approach. In one project, a group of virtual labs investigated whether AI models could be trained more efficiently by starting with a smaller model and expanding it during training. Asaria said the researchers developed an approach that required about 30 per cent less compute during pre-training while producing a better result on one measure of model performance. 

Transformer Lab is entering a crowded field of companies trying to automate scientific research. Startups including FutureHouse and Tokyo-based Sakana AI have developed AI agents designed to autonomously carry out parts of the scientific process. 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. 

The technology is already being put to work inside major AI companies: Anthropic said this month that its Claude models were leading 26 per cent of its AI research and development work, up from one per cent in March, though the company said the work still involves human oversight.

Asaria said Transformer Lab’s edge in the sector will depend on how fast it can build its tools and get them in the hands of those who can use them. He said the company plans to use Primus Society for its own research, which will ideally lead to new breakthrough discoveries and companies. It will also provide access to researchers and companies tackling big problems in fields such as AI, quantum computing and mathematics. The company is initially concentrating on fields where experiments can be done entirely on computers. 

Transformer Lab has also deliberately imposed some human-designed structure onto Primus Society, in an attempt to maintain focus and order. Rather than allowing thousands of agents to devise their own ways of communicating and organizing themselves, the company created the institutions and set parameters for how to communicate and operate within them. The agents still have room to disagree, pursue unexpected approaches and develop new ways of solving problems, said Asaria.

The approach comes as concerns grow around the emergence of AI swarms, large groups of autonomous agents that communicate and work together. In July, roughly 1,200 agents deployed in OpenAI cybersecurity evaluations found an unauthorized way to communicate, exchanging more than 70,000 messages and files. About 700 went on to participate in an attack on AI platform Hugging Face, according to an independent investigation by AI research organization METR. Other AI labs have reported agents breaking out of intended testing environments.

“They still end up behaving in emergent ways,” Asaria said of Primus researchers. “They create conflicts and new ways of solving a problem that we never expected, but we’re trying to get them to communicate in a way that mimics human societies.”

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.”