Kosmos AI reads 1,500 papers and executes 42,000 lines of code in 12 hours, matching 6 months of human research.
The nonprofit FutureHouse released Kosmos, an AI system that can autonomously process 1,500 research papers and generate 42,000 lines of analysis code within a single 12-hour session, equating to six months of work by a human research team. Kosmos maintains logical coherence over 10 million tokens using a structured world model and employs an autonomous loop architecture that iterates through literature retrieval, data analysis, knowledge graph updates, and planning the next exploration phase. It averages 166 data analyses and 36 literature reviews per run, producing fully auditable, traceable reports with precise citations.
Kosmos has independently replicated three unpublished studies, including novel insights on nucleotide metabolism in low-temperature brain processing and humidity thresholds affecting perovskite solar cells. The platform costs about $200 per run commercially, with academic users receiving free credits. It is optimized for datasets under 5GB and achieves approximately 58% cross-domain reasoning accuracy. Future developments aim to integrate lab automation for a closed-loop hypothesis-experiment-analysis cycle, pushing autonomous scientific discovery capabilities forward.
Source: aibase