Kosmos: The Next-Generation Autonomous Research AI Scientist

AI for Science is easily one of the most astonishing areas of progress,

and the era of AI-accelerated research is not only here, it is steadily unfolding.

I read every single one of my friend Renee's AI observations,

and with her permission, I'm sharing this one with you too. Here goes — enjoy:

FutureHouse is an organization focused on AI + scientific research, co-founded by researchers with deep backgrounds in AI, biology, neuroscience, materials science and other fields. It positions itself as "an AI research lab for scientists + a research tools platform."

FutureHouse first caught the research community's attention with Robin (an AI Scientist prototype). Kosmos is Robin's next-generation upgrade, released by Edison Scientific, a new company spun out of FutureHouse, with the paper published at the same time: https://arxiv.org/pdf/2511.02824. Compared with the previous-generation Robin, Kosmos is a major leap in both capability and scale. A new paradigm for AI scientists

Earlier AI Scientists (including Robin) were mainly constrained by the limited context length of language models → they tended to "run out of steam" when reading and integrating large amounts of information. As a result, the depth of reasoning and the complexity of research they could handle were limited too.

Kosmos's core innovation: Structured World Models, which can efficiently integrate information across hundreds of agent trajectories and maintain a coherent research goal at the scale of tens of millions of tokens.

A typical full Kosmos run includes: reading 1,500 scientific papers; executing 42,000 lines of data-analysis code

That scale far exceeds any agent system known to date. One day of Kosmos, half a year for humans

Compared with Robin, Kosmos can carry out more complex and fine-grained analysis. Beta users estimate: one day of Kosmos work ≈ 6 months for a human researcher. Its conclusions are 79.4% accurate.

Kosmos has already made seven scientific discoveries spanning neuroscience, materials science, metabolomics, statistical genetics and other fields, several of them validated by outside scholars. The era of AI-accelerated research is truly arriving.

User feedback: a single Kosmos run ≈ 6 months of PhD/postdoc research time. And a rare inference-time scaling law for depth of scientific reasoning has emerged.

This estimate comes from a survey of 7 Beta users. A Kosmos run at a depth of 20 steps yielded an average of 6.14 months of human-equivalent research time.

More objective metrics back this up too: for three studies that had previously taken humans 4 months, Kosmos reproduced the core findings in a single run each.

Another independent estimate: reading 1,500 papers (15 minutes each); executing a number of analysis trajectories (2 hours each); totals roughly 4.1 months of human work. Auditable and reproducible

Kosmos is designed to follow scientific norms — every conclusion can be traced back to: the specific code used; the specific passage in the scientific literature

Research must be auditable and reproducible, and Kosmos guarantees this at the system level.

In addition, researchers used Kosmos to: reproduce unpublished results; produce genuinely new discoveries

All of which are included in the technical report.

The technical report presents seven of Kosmos's discoveries. Three of them independently reproduced earlier findings by human researchers: 1. Metabolomics: the brains of hypothermic mice

Using metabolomic data, Kosmos reproduced a study that had not yet been published at the time: nucleotide metabolism is the most significantly altered pathway in the mouse brain under hypothermia. The original study's preprint was uploaded to BioRxiv only after the Kosmos run. 2. Materials science: perovskite solar cells

Kosmos reproduced another preprint (published after the model's training-data cutoff), which showed that absolute humidity during thermal annealing determines cell efficiency, and identified the critical threshold (above roughly 60 g/m³, devices fail). 3. Neuroscience: cross-species rules of neural connectivity

Kosmos identified the mathematical law of cross-species neuronal connectivity reported by Piazza et al. Although that preprint was not accessed during the run, it predated the model's training-data cutoff. Four new contributions

In the other four studies, Kosmos produced entirely new contributions: 4. Myocardial fibrosis: the causal role of SOD2

Using public GWAS and pQTL data, Kosmos provided evidence that high levels of SOD2 may causally lower myocardial T1 times and the risk of myocardial fibrosis. This mechanism had previously been validated mainly in mice. 5. Type 2 diabetes: a new molecular mechanism

Combining multi-omics and statistical-genetics data, Kosmos proposed a potential molecular pathway by which a specific SNP lowers T2D risk. 6. Alzheimer's disease: the mechanism of tau accumulation

Kosmos developed a new analytical method to reconstruct, from AD patients' proteomic data, the sequence of molecular events by which tau accumulates in neurons. 7. Neurodegeneration and aging: a new mechanism (clinically relevant)

Task: understand neuronal vulnerability in aging. Kosmos ran a large-scale unbiased analysis of single-nucleus transcriptomes from young and old subjects and found:

In entorhinal cortex neurons, flippase gene expression declines with age → increased external exposure of the PS "eat-me" signal on the membrane → triggering phagocytosis by microglia.

The researchers validated this in an independent human AD dataset: in entorhinal cortex neurons at Braak stage II, flippase expression does indeed decline, coinciding with the onset of tau pathology.

These findings are now being further validated in wet labs. Using it

Kosmos is live, but it is completely different from an ordinary chatbot; it is closer to a "Deep Research Tool." https://platform.edisonscientific.com/ You need to learn some prompting techniques; pricing is $200 per run (200 credits); academic users get a certain free allowance

Kosmos is more like a high-value research reagent kit than a chat model. Significance and future directions

This suggests that the way METR and other organizations currently evaluate AI by "human-equivalent task duration" may be too simplistic: the "sustainable reasoning time" varies enormously across task types. For example: PaperQA2 can write high-quality Wikipedia articles (far beyond 4–15 minutes); a single math proof generated by GPT-5 also clearly exceeds 4 hours of human-equivalent time

So the real-world capacity of AI to accelerate research is far more complex.

Moreover, the deeper a Kosmos run goes, the more likely it is to veer off course (for example, chasing spurious correlations). Depth does not deliver unlimited positive gains. Stronger models will still be needed to solve this problem.