AI Scientists Debut: Achieve in 12 Hours What Human Scientists Take Half a Year, with 7 Major Achievements

AI scientist Kosmos makes its debut, reading 1,500 research papers in 12 hours and generating traceable reports. Human-machine collaboration reshapes scientific research.

[Introduction] Altman says GPT-6 may make "AI creating new science" a reality. In response, the "AI scientist" Kosmos has emerged: it can read 1,500 research papers and run 42,000 lines of code in 12 hours, generate traceable reports, and make new discoveries in fields such as materials science. It can autonomously plan with continuous memory and is evolving from a tool to a collaborator. However, restricted by data sources and reproducibility, about 20% of its conclusions still require human judgment. Human-machine collaboration may reshape scientific research, accelerating the evolution of scientific research paradigms, and the future looks promising.

In a YouTube interview released last night, Altman, the CEO of OpenAI, mentioned his expectations for the future GPT-6.

https://www.youtube.com/watch?v=cuSDy0Rmdks

He said that GPT-5 has allowed him to see the dawn of "AI creating brand-new science" for the first time, and GPT-6 may truly achieve this.

These remarks have sparked the public's infinite curiosity: Can AI really autonomously create new scientific knowledge?

Just when everyone was full of doubts, an "AI scientist" named Kosmos suddenly appeared and gave a preliminary answer with amazing results.

https://arxiv.org/abs/2511.02824

Lightning-fast scientific research speed

A tireless research assistant reads thousands of papers and writes tens of thousands of lines of code in just half a day, and finally produces a well - reasoned scientific report.

This is exactly the scenario that Kosmos has achieved.

In one run, Kosmos completed the research workload equivalent to what an average human scientist would take six months to finish in less than 12 hours!

In this single run, Kosmos executed a vast number of tasks in parallel, reading approximately 1,500 research papers and executing about 42,000 lines of code, which is almost equivalent to the efforts of a research team for half a year.

What's even more surprising is that Kosmos noted the source for each conclusion in the report, either by citing the corresponding code output or the original literature source, ensuring that all inferences are traceable.

It's really astonishing that an AI system can complete such a heavy scientific research task so efficiently and rigorously.

The emergence of Kosmos has let people see multiple advantages of AI in assisting scientific research:

Acceleration: Kosmos can operate autonomously for a long time, quickly completing a large amount of reading and data analysis work, significantly reducing the scientific research time.
Breadth and depth: It conducts research across different disciplinary fields and can always focus on the established goal in hundreds of steps without going off - topic or losing clues.
Transparency: Each conclusion is supported by code or literature, and the reasoning chain is clear and transparent, making it convenient for others to verify. Compared with human's sometimes intuitive and leap - thinking, Kosmos' report is more like an "auditable" scientific research diary.
Scalability: As the operation cycle increases, the number of valuable discoveries made by Kosmos also increases linearly. Investing more computing time can yield more results, and scientific research output is no longer strictly limited by human energy.

Such a remarkable improvement in efficiency indicates that scientific research is about to experience a quantitative leap.

Experiments that used to take months to run, literature to read, and hypotheses to verify can now be completed by AI within a single day.

The impact of this "fast - forward" science is no less than the subversion of manual labor by mass machine production during the Industrial Revolution.

From a tool to a scientific research collaborator

Kosmos is not a simple automated tool that passively executes instructions, but more like an active scientific research collaborator.

Its usage is quite special. Human scientists only need to provide an open - ended research goal and the corresponding dataset, and Kosmos will take over and independently carry out the subsequent research process.

During a running time of up to about 12 hours, it will continuously cycle through the following process:

Plan several small tasks (such as analyzing data or searching relevant materials in the literature library), execute these tasks in parallel, update the obtained discoveries to the shared "world model", and then plan the next round of exploration based on the latest knowledge.

This "world model" can be understood as a structured scientific research notebook inside Kosmos, recording all intermediate results and clues, ensuring that it remains clear - headed and does not deviate from the research direction even after 200 consecutive action steps.

It is precisely because of this continuous memory and context - sharing ability that Kosmos can independently form an exploration path, generate and test scientific hypotheses, and complete a series of coherent experiments and analyses without human intervention.

In contrast, previous scientific research AI tools often could only execute preset pipelines and would easily become "confused" and lose coherence once the steps were too numerous.

The emergence of Kosmos marks that AI is evolving from a scientific research tool to a scientific research partner.

It is no longer just an advanced toolbox for humans, but to some extent, it has the initiative and strategic thinking for scientific exploration.

As a commentator said, in the past, our tools "were just tools" used to complete the tasks we assigned;

now, AI can join our reasoning process, help us think about new problems, discover patterns that we ourselves haven't noticed, and sometimes even give us innovative ideas that puzzle us.

This transformation means that in the laboratory, AI has begun to fight side by side with us as a collaborator.

Of course, calling Kosmos an "AI scientist" does not mean that it can completely replace the creativity and judgment of human scientists.

It is more like a tireless collaborator that can put forward endless ideas and complete complicated verifications, but still needs human wisdom to supervise the direction and judge the importance of the results.

As mathematician Sir William Timothy Gowers sighed after using GPT - 5 to assist in proof: "We have entered such an interesting period: research has been greatly accelerated by AI, but AI still needs us."

AI is increasingly emerging as a powerful force in scientific research, rather than trying to replace humans alone.

Similarly, in the Kosmos project, researchers also emphasize that humans still play a crucial role as judges.

In the report given by Kosmos, about 20% of the content is inaccurate or debatable, and the insights of the "AI scientist" also need to be screened for authenticity and value by experienced researchers.

The combination of the rigor and speed of machines and the insight and creativity of humans may be the ideal model for future scientific research.

Independent discoveries and original achievements

One of the most commendable aspects of Kosmos is its potential to independently produce new scientific research discoveries.

The paper's author team specifically listed seven representative scientific research achievements made by Kosmos in different industries, some of which are quite dramatic.

Neuroprotection

In a study on neuroprotection mechanisms (researching how low temperature protects the mouse brain), after analyzing a large amount of metabolomics data, Kosmos pointed out that the "nucleotide regeneration pathway" in cells is significantly activated, which is a protective mechanism for cells to save energy at low temperatures.

Surprisingly, this conclusion is highly consistent with the results of an unpublished paper by human scientists, and Kosmos did not have access to this paper during its operation.

That is to say, AI completely independently reached the same insight as human researchers based on the data.

Materials science

In the field of materials science, Kosmos was given the goal of improving the manufacturing efficiency of perovskite solar cells.

Through machine learning to analyze experimental data, it discovered a key factor: in the battery manufacturing process, the environmental humidity during the "thermal annealing" step is crucial - a slightly higher humidity may lead to a serious decline in device performance.

Even more impressively, Kosmos further proposed a simple quantitative relationship: during the spin - coating process, the higher the vapor pressure of the solvent DMF, the linearly lower the short - circuit current performance of the battery.

This new law was later experimentally verified by human researchers.

This means that Kosmos not only repeated and confirmed the known experience but also independently proposed a new engineering guidance law.

Among Kosmos' seven discoveries, three were later confirmed to coincide with the unpublished results independently obtained by human research teams - AI "independently repeated" those human discoveries that had not yet been made public;

while the other four were completely original contributions, advancing the knowledge frontier in the relevant fields.

For example, in the study of brain connectomics, Kosmos found that the number of neuronal connections in different species follows a log - normal distribution and proposed a possible generation mechanism, which is consistent with and goes further than the conclusion of a previous human pre - print;

Another example is in genetics, where it predicted a key protein (superoxide dismutase SOD2) against cardiac fibrosis and gave a speculated mechanism of action, providing new clues for medical research.

These cases show that Kosmos can really sniff out valuable new patterns and hypotheses from data and literature, some of which humans themselves have not even realized.

Kosmos has let us see the charm of AI as a scientific research "discoverer" for the first time.

The new knowledge and insights independently produced by it, although still need human confirmation, are sufficient to prove the great potential of artificial intelligence in creating knowledge.

In the past, human scientists had to read a vast amount of literature, try various analysis methods, and experience countless failures before they could occasionally have a flash - of - inspiration discovery;

now, an AI may be able to submit a list of amazing results after a long - night of computing.

This scenario was almost unimaginable a few years ago, but it has really happened.

The future of scientific research: challenges and speculations

Although Kosmos has achieved remarkable results, current AI scientists still have obvious limitations and challenges.

First, Kosmos does not collect new data on its own.

It requires researchers to provide a dataset in advance and conducts analysis within the scope of the given materials.

Even if Kosmos is very smart, it can't "make bricks without straw" if there is a lack of high - quality data.

It also cannot make on - the - spot decisions like humans to start new experiments or crawl the latest data on the Internet. This is both a technical limitation at present and a consideration for security and ethics.

Second, Kosmos is not yet able to directly process unstructured data such as raw images.

In this work, it mainly analyzed structured data such as tabular experimental results, gene sequences, and neural connection matrices, as well as paper texts.

If the research involves visual information such as microscope images and medical images, Kosmos is still powerless and needs the assistance of other image - recognition AIs for pre - processing.

This means that its scope of independent discovery is currently still limited to data types that it can "understand" and problems that it can "comprehend".

Third, there is