Two years after setting out to build an AI Scientist, the team behind FutureHouse is splitting its efforts. This month marked the launch of Edison Scientific, a for-profit spinout tasked with commercializing the group's AI research capabilities, while the original FutureHouse remains focused on non-profit, foundational biology research.
This move comes as the team unveils Kosmos, their next-generation AI Scientist, claiming it can accomplish the equivalent of six months of human research in a single day. Sam Altman christened their debut with praise on x.com.
This is exciting; I expect we are going to see a lot more things like this and it will be one of the most important impacts of AI. Congrats to the Future House team.https://t.co/Cxeh8UlWdk
, Sam Altman (@sama) November 16, 2025
The pivot to a commercial entity, Edison Scientific, is a direct response to overwhelming inbound interest. Founders Sam Rodriques and Andrew White noted they were inundated with requests from major pharma players following the initial platform launch in May 2025. Building out product features, payment systems, and customer support, they argue, is better suited for for-profit capital than their philanthropic funding base. Edison promises a "generous free tier" to maintain community access, but power users and enterprises will be paying for higher throughput.
Kosmos itself represents a significant leap over its predecessor, Robin. The key technical innovation cited is the use of "structured world models," allowing the agent to maintain coherence over massive information loads, reportedly processing 1,500 papers and 42,000 lines of analysis code in a single run. This overcomes the context length limitations that plagued earlier models. Beta user feedback suggests a 6-month human-equivalent output from a single Kosmos run, a figure the team validates through objective replication of prior scientific discoveries and independent time estimation based on task components.
The Six-Month Leap and Industry Implications
The claim that Kosmos achieves six months of human work in one day is the headline grabber, though the team is careful to qualify it. They acknowledge that the AI can still chase "rabbit holes" and that multiple runs are often necessary. However, the validation, reproducing unpublished findings and calculating time based on documented analysis steps, lends weight to the acceleration narrative.