FutureHouse’s Kosmos, now operated by Edison Scientific, marks a breakthrough in AI-assisted science—an auditable, deep-research system capable of reading thousands of papers, performing tens of thousands of analyses, and producing verified discoveries across disciplines at a speed equivalent to six months of human research in a single day. (Source: Image by RR)

Kosmos System Achieves Nearly 80 Percent Accuracy Across Thousands of Analytical Steps

FutureHouse has unveiled Kosmos, a groundbreaking next-generation AI Scientist capable of synthesizing vast amounts of scientific data, performing large-scale analyses, and making new discoveries across disciplines. Developed as a successor to Robin, FutureHouse’s earlier AI research system, Kosmos represents a leap in computational science. The system can read over 1,500 research papers, run 42,000 lines of analysis code, and maintain coherence over tens of millions of tokens — all in a single research run. Its results are striking: Kosmos can reportedly accomplish six months of human scientific work in one day with nearly 80% accuracy.

Kosmos’s launch coincides with the transition of FutureHouse’s platform management to Edison Scientific, a commercial spinout tasked with maintaining a free academic tier while introducing paid options for advanced users. Powered by structured world models, Kosmos connects hundreds of agent trajectories to pursue specific research objectives, enabling it to uncover patterns and relationships across complex data. Unlike typical chatbots, Kosmos functions as a deep research agent, not a conversational AI. FutureHouse, as noted in edisonscientific.com, emphasizes transparency: every finding can be traced back to its data source or code line, ensuring auditable, verifiable science.

In collaboration with academic beta testers, Kosmos made seven major discoveries in fields from neuroscience and materials science to genetics. Three replicated unpublished or recently published human findings, such as confirming the effect of humidity on perovskite solar cell efficiency and the cross-species mathematical rules of neuronal connectivity. Four others were entirely novel, including a genetic mechanism that may reduce Type 2 diabetes risk and a potential link between SOD2 protein levels and heart fibrosis reduction. Perhaps most notably, Kosmos discovered that reduced flippase gene expression may trigger early Alzheimer’s-related neuron loss, a finding later validated in human RNA-seq data.

While the system’s output is extraordinary, FutureHouse admits it isn’t flawless — Kosmos can pursue false correlations or unpromising “rabbit holes.” Runs cost $200 each, with discounted pricing for founding subscribers. FutureHouse acknowledges Kosmos is “not a chatbot” but a scientific reagent—a tool for serious inquiry requiring expert handling. Despite imperfections, Kosmos signals the arrival of AI-accelerated discovery, where intelligent systems act as autonomous collaborators rather than assistants.

read more at edisonscientific.com