The data scientist’s job was built on three skills: statistics, algorithms and software engineering. A new generation of AI systems is now performing large parts of that work, from cleaning raw data and engineering features to finding patterns in complex datasets, and the most ambitious of them attempt to run the entire research cycle on their own. Kosmos, an “AI scientist” announced in November 2025 by Edison Scientific, a company spun out of the nonprofit research lab FutureHouse, is the clearest example so far of where this is heading.
Given a dataset and an open-ended research objective, Kosmos works for up to 12 hours without human input, according to the developers’ paper posted on arXiv. During that time it reads scientific literature, writes and runs analysis code, builds models and produces a report with figures, code and citations.
How it works
Kosmos is not a single model. The developers describe it as a system of many specialized agents: some search and read papers, some analyze data, some write code and others check for errors and correct them. The agents share a structured “world model,” a common memory that lets the system stay focused on one objective across a long run rather than losing the thread after a few steps, a known weakness of earlier agent systems.
In a single run, the developers say, Kosmos can execute hundreds of analyses, read roughly 1,500 papers and generate tens of thousands of lines of code. It starts by generating many candidate research directions from the data, selects the most promising, and builds a plan around them.
Early results
Edison Scientific reports several findings from early trials. In one, the system identified a mechanism by which brain cells conserve energy under cooling, which the company says was consistent with later human research. In another it proposed a mechanism for humidity-driven degradation in perovskite solar cells that human researchers then confirmed experimentally. In a third, it found a mathematical relationship in neuron connectivity data across species that the developers say had not been previously noted.
When independent scientists reviewed Kosmos reports, about 79 percent of the statements in them were judged scientifically accurate, according to the paper. The developers estimate that a single 12-hour run performs work equivalent to about six months of effort by a junior researcher. New Scientist, reporting on the launch, noted that the claim is the company’s own and that the system’s output still requires expert review.
Limits
The developers acknowledge several constraints. Kosmos does not handle unstructured or unlabeled data well, cannot work with very large files, and does not accept new instructions once a run has started. Roughly one in five of its statements was inaccurate in the evaluation, which means every report needs checking before anyone acts on it.
The most important limit is judgment. Kosmos can generate findings, but deciding which ones matter, which are coincidental and which deserve follow-up experiments remains a human task. Edison Scientific positions the system as a collaborator that accelerates research rather than a replacement for scientists.
What it means for the profession
For data scientists, Kosmos and systems like it point toward a shift in where human skill adds value. Routine analysis, code and literature review are increasingly automatable. Framing the right question, judging the significance of a result and taking responsibility for conclusions are not, at least not yet. Research is no longer bound to the pace at which humans can run experiments, and the people who direct these systems well will have an advantage over those who compete with them on throughput.
