Nonprofit research institution FutureHouse has released the AI research system Kosmos, which can read 1,500 papers and generate and execute 42,000 lines of analysis code in a single 12-hour session, producing traceable reports. The workload is equivalent to that of a human team for half a year. The system is based on a structured world model, maintaining logical coherence within a context of over 10 million tokens. It has completed seven discoveries in fields such as neuroscience and materials science, four of which are reported for the first time.
Kosmos uses an autonomous loop architecture: parallelly initiating literature retrieval and data analysis tasks → updating the internal knowledge graph → then planning the next round of exploration. On average, each run iterates 166 data analyses and 36 literature reviews. The conclusions cite precise code segments or original texts and support full auditing. In tests, Kosmos independently replicated three unpublished studies, including confirming nucleotide metabolism as a key pathway for low-temperature brain processing and discovering an absolute humidity failure threshold of over 60g/m³ for perovskite solar cells.
The platform is now commercialized, with a cost of approximately $200 per run, and academic users receive free credits. The system performs best with datasets under 5GB, with cross-domain comprehensive reasoning accuracy of about 58%. The research team is working on integrating lab automation equipment to form a full "hypothesis-experiment-analysis" closed loop.