Sam Altman praises Kosmos, the AI scientist helping uncover disease clues

Sam Altman, the CEO of OpenAI, has expressed enthusiasm over a groundbreaking advancement in artificial intelligence research with the introduction of Kosmos, an innovative AI Scientist developed by Future House. In a recent statement, Altman described this development as "exciting," highlighting its potential to significantly influence scientific discovery in the coming years. Kosmos is touted as a major upgrade from Future House's earlier AI Scientist, Robin. Managed by Edison Scientific, a newly formed commercial entity, Kosmos aims to facilitate scientific progress while offering a generous free tier for academic use. A detailed blog post released last week revealed that traditional AI models often struggle with vast amounts of scientific data due to their limited memory capacity, which hampers their ability to navigate intricate chains of reasoning. In contrast, Kosmos employs structured world models to tackle these challenges, allowing it to leverage information from hundreds of agent runs while maintaining focus on a specific research objective, even when processing an extensive number of tokens. Each Kosmos run involves reviewing approximately 1,500 scientific papers and executing over 42,000 lines of analysis code, significantly exceeding the capabilities of previous research systems. Early users have estimated that Kosmos can accomplish in a single run what would typically take a human scientist six months to complete, with internal tests indicating that nearly 80% of its conclusions are accurate. During its testing phase, Kosmos made seven notable discoveries across various fields, including neuroscience, materials science, genetics, and aging research. Among its confirmations were unpublished findings regarding nucleotide metabolism changes in hypothermic mice, a materials science discovery related to perovskite solar cell performance, and the identification of common mathematical rules for neuronal connectivity across different species. Kosmos also generated four new insights: it proposed that elevated levels of the antioxidant enzyme SOD2 could mitigate heart fibrosis in humans, offered a new molecular explanation for a genetic variant linked to a lower risk of Type 2 diabetes, devised a method for analyzing tau protein accumulation in Alzheimer's disease, and identified age-related changes in gene expression among neurons in the entorhinal cortex that may render them more susceptible to microglial attacks. Edison Scientific is currently validating several of these findings through wet lab experiments. A standout feature of Kosmos is its traceability; each conclusion can be traced back to specific lines of code or scientific texts that contributed to it, thus addressing the common "black box" issue seen in many AI solutions. Kosmos is now available for $200 per run, with academics receiving some complimentary usage. Early adopters can secure the current pricing before it increases. The developers emphasize that Kosmos functions more like a scientific reagent kit rather than a chatbot, addressing high-value research questions. While it can produce dead ends or yield statistically intriguing yet scientifically irrelevant patterns, it may require multiple runs for certain projects. The assertion that Kosmos can equate to six months of scientific work even surprised its creators, as it was based on comparisons made by beta users against their manual research duration. Given the volume of papers read and analysis steps undertaken, Edison Scientific argues that the time savings align with several months of full-time research effort.