The non-profit organization FutureHouse has released the AI research system Kosmos. Official tests show that it can read 1,500 papers, generate 42,000 lines of executable code, and produce a complete reference report within 12 hours. The workload is equivalent to that of a human team working for six months, with a comprehensive accuracy rate of 79.4%.

image.png

Kosmos adopts a structured world model, parallelly scheduling literature retrieval, data analysis, and hypothesis verification modules. It averages 166 rounds of data processing and 36 literature reviews per single run, with conclusions traceable back to original code snippets. The system has independently replicated seven cutting-edge discoveries, including confirming nucleotide metabolism pathways in low-temperature brain processing and the 60g/m³ humidity failure threshold of perovskite solar cells, four of which are reported for the first time.

image.png

The platform charges approximately $200 per use, and academic users can apply for free quotas. It supports datasets under 5GB and browser interaction. FutureHouse plans to open its API this year and integrate lab automation equipment, forming a complete "hypothesis-experiment-analysis" loop. Industry insiders point out that Kosmos compresses the research cycle by an order of magnitude, but users need to master structured prompting techniques, and it requires high computing power. In the short term, it still cannot completely replace human creativity.