FARS: 100 Research Papers in 228 Hours and 11.4 Billion Tokens
The FARS system (Fully Automated Research System) has startled the research community: in 228 hours of continuous operation, the agent independently generated 100 scientific papers, consuming 11.4 billion tokens along the way. The experiment shows how autonomous AI agents are changing the pace of scientific research: tasks that take human researchers months can now be done in a matter of days. The question of the quality and reliability of automatically generated papers remains open.
AI-processed from Jiqizhixin (机器之心); edited by Hamidun News
One hundred scientific papers in nine and a half days. That is the result delivered by FARS (Fully Automated Research System), which did it without a single pause for sleep, coffee or a scientific argument with colleagues. The experiment, whose details have stirred the research community, puts a question to the academic world that can no longer be put off: what happens to science when a machine writes faster than a human can read?
In 228 hours of continuous work, FARS generated exactly 100 scientific papers and consumed a colossal 11.4 billion tokens, roughly what several tens of thousands of scientific books contain. The figures are impressive on their own, but their true scale only becomes clear in comparison: an experienced researcher spends anywhere from several weeks to several months preparing a single full-fledged paper. FARS averaged a little over two hours per paper. That is not a faster process; it is a qualitatively different one.
The emergence of such systems is no accident. The past two years have been marked by rapid progress in autonomous AI agents: software systems that can plan tasks, search for information, formulate hypotheses and write up results without constant human involvement. Early language models needed detailed instructions at every step, whereas modern agent architectures can string together long chains of actions, from posing a research question to final editing of the text. FARS is exactly this kind of closed system, with the entire research cycle automated.
The technical side of the experiment deserves separate attention. The 11.4 billion tokens are not just a measure of the computing resources consumed; they are an indirect indicator of how deep the system's work went. The agent did not merely paraphrase existing texts: it processed large volumes of data, built arguments, generated bibliographic references and structured the material to academic standards. Such high token consumption suggests the system really did go through multistep iterations rather than producing superficial summaries. Yet this is exactly where questions begin that have no clear answers so far.
The central problem of automated scientific generation is the quality and reproducibility of results. An academic paper is not only a formatted text but also verified knowledge: data that can be checked, methodology that can be repeated, conclusions that withstand reviewers' criticism. None of the experiment's organizers claims that all 100 papers have passed independent expert review. How well these papers meet the standards of peer-reviewed journals remains an open question, and it is arguably the most important one in this whole story. In science, speed without accuracy is not merely useless but dangerous: false results that enter the academic record can mislead entire research areas for years.
Nevertheless, the potential of such systems is hard to ignore. In some fields, including epidemiology, climatology and materials science, data processing speed is critical, and researchers literally cannot keep up with the incoming flow of information. An autonomous agent that can process and organize in a day a body of data that would take a human team a quarter is a real tool for accelerating knowledge, not just a display of computing power. The key question is how to build such systems into existing verification mechanisms without losing accuracy or scientific integrity along the way.
FARS is not a finish line but the first mark on a long scale. The academic community, publishers and regulators must develop new standards: how to label work with a high share of automated generation, how to adapt peer review to a different pace of text production, and how to tell machine-driven deep research from machine imitation of depth. The experiment has shown that the speed barrier has already been crossed. The next barrier is trust. And its height is not measured in tokens.
- Nov 21, 202615 Claude Code Subagents Built a Shop-Floor System, but a Release Gate Agent Without Edit Rights Approves It
- Nov 21, 2026Coding Agents Built an LLM Engine for AWS FPGAs: 213 Tokens/s and 38% Cheaper Than CPU
- Nov 21, 2026Empatra Studio 0.29 Cuts Frontend Agent Costs by 59% and Runs Twice as Fast
- Nov 21, 2026Ego lite Gives Each AI Agent Its Own Workspace Space in the Browser
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.