FutureHouse, a nonprofit organization focused on developing artificial intelligence (AI) for science, recently announced the launch of Edison Scientific. This commercial spinout is tasked with developing and deploying AI systems for practical applications. Founders Sam Rodriques and Andrew White lead both entities, with FutureHouse focusing on fundamental research in biology. Edison Scientific was created in response to growing interest from pharmaceutical companies, including six of the ten largest, which are seeking higher limits and additional features. The company promises a generous free plan for the scientific community, while paying users will gain access to advanced capabilities.
Kosmos Is Like an Advanced AI Scientist
Kosmos is a new generation of AI scientist that builds on FutureHouse's previous Robin system. This tool uses structured world models, allowing it to process information from hundreds of agent trajectories and maintain consistency across tens of millions of tokens. During a single session, Kosmos reads 1,500 scientific papers and runs 42,000 lines of analytical code. That is far more than other known agents. Beta users estimate that Kosmos can complete in one day work that would take a human six months. The accuracy of its conclusions reaches 79.4%.
Discoveries Made by Kosmos
Kosmos has already made seven significant discoveries in fields such as neuroscience, materials science, and statistical genetics. For example, in one case it analyzed metabolomic data and identified nucleotide metabolism as the primary altered pathway in the brains of hypothermic mice—confirming a study that was unpublished at the time. Another discovery involved perovskite solar cells, where Kosmos identified absolute humidity during thermal annealing as a key factor, with a critical threshold above 60 g/m³ at which devices fail. In genetics, it proposed a mechanism by which a single-nucleotide polymorphism reduces the risk of type 2 diabetes.
Kosmos also discovered a sequence of molecular events leading to tau accumulation in neurons of patients with Alzheimer's disease using proteomic data. In another case, it analyzed transcriptomic data from mice and found reduced expression of flippase genes in the entorhinal cortex with age, which could lead to neuronal damage. This finding was validated using data from human Alzheimer's disease cases.
How Kosmos Works in Practice
Kosmos emphasizes transparency—every conclusion can be traced back to the original lines of code or passages from the literature. It is available on the Edison Scientific platform for $200 per session, with 200 credits priced at $1 each. Academics have free access. The tool is not like a conventional chatbot; it requires careful formulation of objectives and may occasionally pursue dead ends, so multiple runs are recommended.
