Kosmos is an autonomous AI scientist built by Future House (now Edison Scientific), released in early November 2025.
Kosmos is designed to operate like a full research team, independently doing: hypothesis generation data analysis across multiple scientific domains writing & executing code proposing testable insights
Kosmos processes ~1,500 scientific papers, 42,000 lines of code, and months of human PhD-level work in a single 24-hour cycle. Uses a structured world model to integrate cross-field data (neuroscience, genetics, materials science, etc.). Has already made 7 novel scientific discoveries with ~80% reproducibility after human review.
Considered a breakthrough in AI-driven scientific discovery, similar to tools like:
- Sakana AI’s Science Agent
- DeepMind’s AlphaFold
- OpenAI’s O1-series reasoning tools
- BenchSci’s ASCEND (drug discovery AI)
Other Notable Discoveries
- Aging-related flippase decline (separate aging mechanism).
- Mechanistic explanation of type 2 diabetes risk variants.
- Scaling laws in connectomics for materials/brain efficiency.
- Refined statistical genetics workflows.
These results show Kosmos behaves like a multidisciplinary research collaborator, similar to a “generalist AI lab analyst.”
SOD2 Reduces Heart Fibrosis
- Used genetic datasets to show high SOD2 levels causally protect against heart fibrosis.
- Found SOD2 reduces oxidative stress → prevents fibroblast activation → lowers tissue stiffness.
- Provides the first strong human genetic causal evidence, beyond prior mouse studies.
- Potential applications in precision cardiology and gene therapy.