FARS: 100 research papers in 228 hours and 11.4 billion tokens

Hundred scientific papers in nine and a half days. That is the result demonstrated by the FARS system — Fully Automated Research System — without taking a single pause for sleep, coffee, or scientific debate with colleagues. The experiment, whose details shook the research community, raises a question for the academic world that can no longer be postponed: what happens to science when a machine writes faster than a human can read?

In 228 hours of continuous operation, FARS generated exactly 100 scientific papers, consuming a colossal 11.4 billion tokens — roughly equivalent to the content of several tens of thousands of scientific books. The numbers themselves are impressive, but their true scale becomes evident only in comparison: an experienced researcher spends from several weeks to several months preparing one complete article. FARS averaged just over two hours per paper. This is not an acceleration of the process — it is its qualitative transformation.

The emergence of such systems at this moment is no coincidence. The last two years have been marked by rapid development of autonomous AI agents — software systems capable of independently planning tasks, searching for information, formulating hypotheses, and presenting results without constant human involvement. While early language models required detailed instructions at each step, modern agent architectures can build long chains of actions: from formulating a research question to final text editing. FARS represents exactly such a closed system, where the research cycle is fully automated.

The technical side of the experiment deserves separate attention. 11.4 billion tokens is not merely a measure of consumed computational resources; it is an indirect indicator of the depth of the system's work. The agent did not simply paraphrase existing texts: it processed data arrays, built argumentation, generated bibliographic references, and structured material according to academic standards. Such high token consumption suggests that the system truly went through multi-step iterations rather than generating superficial summaries. Yet this is precisely where questions arise for which there are not yet definitive answers.

The central problem of automatic scientific generation is the quality and reproducibility of results. An academic paper is not only formatted text but verified knowledge: data that can be checked, methodology that can be repeated, conclusions that will withstand peer review. None of the experiment's organizers claim that all 100 papers passed independent expert review. The question of how well these papers meet the standards of peer-reviewed journals remains open — and this is perhaps the most important question in this entire story. Speed without accuracy in science is not merely useless — it is dangerous: false results that enter academic circulation can mislead entire research directions for years.

Nevertheless, the potential of such systems is difficult to ignore. There are areas where the speed of data processing is critically important — epidemiology, climatology, materials science — where researchers simply cannot keep up with analyzing incoming information flows. An autonomous agent capable of processing and systematizing in a day what a human team would take a quarter to do — this is a real tool for accelerating knowledge, not merely a demonstration of computational power. The key question is how to integrate such systems into existing verification mechanisms without losing either accuracy or scientific integrity.

FARS is not a final point but a first milestone on a long scale. The academic community, publishers, and regulators face the necessity of developing new standards: how to label works with a high proportion of automatic generation, how to adapt peer review to a different pace of text production, how to distinguish machine deep research from machine simulation of depth. The experiment clearly showed that the speed barrier has already been overcome. The next barrier is trust. And its height is measured not in tokens.