(Shutterstock)
FutureHouse, the nonprofit research group known for a modular approach to agentic AI for science, has launched a for-profit spinout called Edison Scientific. The new company will commercialize FutureHouse’s AI research tools and develop new ones built for use in laboratories and research institutes. With the announcement, the group introduced Edison Scientific’s first product: Kosmos, an experimental AI co-scientist designed to support literature review, data analysis, and hypothesis generation.
Founded by neuroscientist Sam Rodriques and chemistry researcher Andrew White, FutureHouse has focused on building specialized AI agents optimized for different parts of the scientific process. Its earlier tools, such as Crow and Falcon, are built to help researchers search scientific papers and databases, while Phoenix and Owl are made for experimental design and finding prior work for a given subject area. Those tools were offered under a nonprofit model through a developer platform and API released earlier this year. Edison Scientific will now manage the commercial side of that work, providing sustainable funding and support for wider adoption.
In a statement, FutureHouse’s founders said the decision to create a commercial spinout is a response to growing demand from industry users. Since launching its platform in May 2025, the organization has seen rising interest from pharmaceutical and biotech companies, including executives from several of the world’s largest firms. Requests for expanded access and higher usage limits have exceeded what FutureHouse could support under its nonprofit structure. According to Rodriques and White, developing commercial infrastructure such as payment systems, customer support, and enterprise deployment would not be an appropriate use of philanthropic funds. Creating Edison Scientific allows the group to attract private investment to scale its technology, while keeping the nonprofit focused on basic research that cannot be funded otherwise, the founders said.
“Collaborators reported that a single 20-cycle Kosmos run performed the equivalent of 6 months of their own research time on average,” Edison Scientific claims
Edison Scientific’s flagship project is Kosmos, described in a technical paper this week as an AI scientist capable of conducting end-to-end research. Kosmos can read scientific literature, analyze data, form and test hypotheses, and produce reports that include citations and code. The system is built around a framework of multiple agents that operate in parallel, sharing information through what the authors call a structured “world model.” This shared representation allows Kosmos to connect patterns across datasets and disciplines rather than treating each task in isolation, with the authors claiming it can execute an average of 42,000 lines of code and read 1,500 papers per run.
In demonstrations, Kosmos has processed large datasets in fields ranging from neuroscience to materials science and generated reports with traceable claims tied to either data outputs or literature sources. “Independent scientists found 79.4% of statements in Kosmos reports to be accurate, and collaborators reported that a single 20-cycle Kosmos run performed the equivalent of 6 months of their own research time on average,” the paper claims.
“The most surprising part of our work on Kosmos — for us, at least — was our finding that a single Kosmos run can accomplish work equivalent to six months of a PhD or postdoctoral scientist,” wrote Sam Rodriques and Michaela Hinks in an Edison Scientific blog post about Kosmos.
The authors say the six months estimate comes from feedback gathered during Kosmos’s beta testing, when scientists who used the system were asked how long it would have taken them to reach similar conclusions on their own. The team also compared Kosmos’s output to real research timelines, noting that several of its reproduced findings had taken human researchers about four months to complete. A separate calculation, based on the time it would take a scientist to read the same number of papers and perform the same analyses, produced a similar result. While the company acknowledges that these estimates are rough, it argues that they suggest Kosmos can perform the equivalent of several months of human research effort in a single automated run. A more detailed discussion of their reasoning appears in the blog post.
Scientific AI tools have struggled to meet the precision standards of research, yet Kosmos offers evidence that the accuracy gap could be narrowing. The AI tool is not presented as a replacement for scientists. The company describes it as a system meant to support human research by freeing up time spent on repetitive analysis and document review. The developers emphasize that human oversight remains central, and that Kosmos’s findings are only meant to guide investigations. The system still depends on quality input data and performs best in domains with structured datasets and established methods.
While FutureHouse continues its nonprofit research on agentic AI systems for open science, Edison Scientific will focus on commercial deployment of products like Kosmos. It will be interesting to see if the two organizations can demonstrate how AI might one day significantly speed up the scientific research process while maintaining the standards of good science.
Related
