Toronto-based Transformer Lab has launched Primus, an AI system that can produce academic journal-style papers from a single prompt. The company claims its technology is further along than better-funded competitors founded by star AI researchers, and that it can produce research results for less than $3,000 per paper.
Primus works by using AI agents to scan relevant scientific literature, develop a theory, run digital experiments to test it, and write up findings. The initial focus is on machine learning research, but the company plans to expand into other scientific fields.
"Every time it does an experiment, it learns about what it could have done better," said Transformer Lab CEO Ali Asaria. "And then it can improve itself."
Transformer Lab has already used Primus to generate 30 papers on subjects including improving audio model output, making AI cite legal provisions accurately, and analyzing ChatGPT's underlying values. Google DeepMind recently cited a paper written by Primus in a technical report about one of its new models.
The four-person startup, co-founded in September 2024 by Asaria and former colleague Tony Salomone, has raised a pre-seed round from Garage Capital, Ripple Ventures, and Mozilla's venture arm. It is competing against high-profile rivals including Discovery Loop, founded by Google chief scientist Jeff Dean and three colleagues last week, and other companies founded by researchers such as University of British Columbia professor Jeff Clune and University of Alberta alumnus David Silver.
Transformer Lab sells its own library of research papers weighted by quality to help Primus avoid hallucinations and produce accurate citations. The system must demonstrate real data to back up its results and address critiques from inbuilt AI agent reviewers.
Rubric Labs, a Toronto software development agency, used Primus to explore ways to condense data fed into AI models' working memory. "Ultimately, it's boiling down to, 'What ideas do you have?'" said CEO Sarim Malik.
Transformer Lab plans to charge a fraction of the US$1 million annual cost of employing a human machine learning researcher. "You now have access to as many AI engineers as you want," Asaria said. "The big bets used by folks with a resume. You can just log in and start writing your own papers."
Competition heats up
Primus is up against similar projects from the world's biggest tech companies and some of the most funded AI startups. They are pursuing self-advancement, the stage at which each system can design and develop better versions of itself. Anthropic projects such a tipping point could be close.
AI for Research Scientists offers training for science scholars exploring AI tools, directly matching the needs of the target audience here.
Why this matters for Science & Research professionals
This shifts the cost and access of AI-driven research. Professionals in finance, pharmaceuticals, and other fields can now test hypotheses without buying expensive hardware or hiring machine learning experts. Primus suggests AI-generated publications can reach institutional validation, as demonstrated by Google DeepMind's citation of its work.
AI for Science & Research covers how tools like Primus are changing the research workflow for working professionals.
Your membership also unlocks:
