AI scientist claimed to do six months of research in just a few hours (www.newscientist.com)

Edison Scientific announced Kosmos, a multi-agent “AI scientist” designed to ingest datasets, search the literature and iteratively generate analyses and code. In a typical 12‑hour run Kosmos will scan roughly 1,500 papers and produce ~42,000 lines of code to interrogate data, outputting summaries, citations and a plan for further cycles. Edison says 20 such cycles are roughly equivalent to six months of human research and claims Kosmos produced seven externally validated discoveries (four labeled novel), including a method for timing cellular pathway failure in Alzheimer’s and an association between SOD2 levels and reduced cardiac scarring. The system highlights how autonomous agents can massively accelerate literature review, reproducible analysis pipelines and hypothesis generation, but important caveats remain. Independent PhD‑level reviewers judged 79.4% of 102 Kosmos statements as supported overall (85.5% for data‑analysis claims, 82.1% for literature claims), yet only 57.9% of its novel breakthrough claims were accurate. Critics point to methodological flaws, failed or ignored code runs, and heavily preprocessed data that may overstate automation. Edison acknowledges limitations and frames Kosmos as a collaborator, not a replacement — promising for scaling certain data‑driven discovery tasks, but still requiring human oversight to validate methods, catch failure modes and provide deeper creative insight.