When AI Becomes a Scientist: FARS and the Industrial Revolution of Research
*A story about 100 papers, 228 hours, and boundless curiosity*
Have you ever wondered how scientific discovery actually happens?
Picture a typical scenario: a graduate student sits in a lab, staring at a pile of data. She proposes a hypothesis, designs an experiment, spends three months collecting data, and then discovers—the hypothesis is wrong.
Three months. Wrong.
In traditional research, this is normal. We call it "trial and error." But trial and error carries enormous costs: time, money, and countless wasted talents.
What if a machine could trial-and-error tirelessly, completing a full research cycle every two hours?
This isn't science fiction. This is FARS.
---
1. The Parcel Station Insight
Start with a simple analogy. Traditional research is like standing in line at a parcel pickup station, shuffling forward one step at a time. You might wait a long time, and if your pickup code turns out to be wrong, you have to start over.
That's the current state of human research:
- Propose an idea (queue up)
- Apply for funding (keep waiting)
- Run experiments (finally your turn)
- Find out it's wrong (back of the line)
- We only publish "successful" experiments (so the same failures get repeated countless times)
- Every paper must be a "complete story" (causing massive redundant work)
- Researchers need sleep, meals, and the occasional social media break (hurting efficiency)
- ICLR 2026 human submissions average: 4.21
- ICLR 2026 accepted papers average: 5.39
- Runtime: 228 hours 28 minutes 33 seconds
- Hypotheses generated: 244
- Papers completed: 100
- Tokens consumed: 11.4 billion
- Total cost: ~$104,000 (≈750,000 RMB)
- Average cost per paper: ~$1,000
- AI Scientist (Sakana AI)
- CycleResearcher
- Zochi
- AI Scientist v2
- DeepScientist
- AI researchers: large-scale hypothesis generation and verification
- Human researchers: grand visions, value judgments, meaning-making
- Analemma. (2026). *Introducing FARS*. https://analemma.ai/blog/introducing-fars
- 36Kr. (2026). *228 hours without sleep: AI scientist FARS produces 100 papers*. https://36kr.com
- GitHub. *FARS-Analemma*. https://github.com/fars-analemma
- Sakana AI. *The AI Scientist*. https://sakana.ai/ai-scientist
FARS says: stop queueing.
It creates a never-stopping research pipeline. Like a modern logistics hub, packages flow along the conveyor belt, each station does one thing, and the whole system runs 24/7.
The result: 228 hours, 100 papers.
One paper every 2 hours 17 minutes on average—equivalent to 3–6 months of work for a human researcher.
---
2. Four "People" Collaborating
FARS is not one big model. It is four specialized agents working together like a research team.
1. Ideation
The team's "creative director." It reads literature nonstop, hunts for research gaps, and proposes hypotheses. Imagine a person who can read every new arXiv paper around the clock, remember everything, and notice "hey, there's a pattern nobody has spotted." During FARS's live experiment, Ideation generated 244 research hypotheses.2. Planning
Once a hypothesis exists, what next? The Planning agent designs the experiment: what data is needed, which models, how to evaluate. Like an experimental designer that can weigh dozens of candidate designs in seconds and pick the optimal one.3. Experiment
The hardest part. The Experiment agent writes code, runs experiments, and analyzes results—backed by a 160-GPU cluster it can use freely. Imagine a programmer running different experiments simultaneously on 160 machines, never tired, never sloppy.4. Writing
Finally, the Writing agent assembles everything into a paper. But there's a key difference: FARS's papers are short papers focused on a single contribution. And it reports negative results. In traditional research, negative results often disappear into a drawer forever. FARS says: no—negative results are knowledge too.---
3. First Principles: What Is Research?
The FARS design team asked a fundamental question:
"If we designed a research system from scratch, free from human limitations, what would it look like?"
Human research has odd constraints:
FARS's design philosophy: return to the essence of research.
What is that essence?
A clear hypothesis + reliable verification of it.
Whether the result is positive or negative, it's knowledge. FARS's output is this "minimal unit of knowledge": short papers, single contribution, honest reporting.
---
4. The Quality Question: Can AI Do Good Science?
The question on everyone's mind: how good are FARS's 100 papers?
The team evaluated them with Stanford's Agentic Reviewer system (an AI that simulates human peer reviewers).
Result: average score 5.05 out of 7.
For reference:
FARS outperforms the human submission average but falls slightly below the acceptance line.
What does this mean? FARS is a stable mid-tier production machine. It won't suddenly produce Nobel-level breakthroughs, but it can steadily produce meaningful, academically valuable research. Given that it is a fully unattended automated system, this is already remarkable.
---
5. The Power of Scale
FARS chose to live-produce 100 papers for a deep reason: scale is the only way to evaluate an automated research system. A few cherry-picked examples prove nothing. Only sustained mass output exposes a system's true capabilities and limits—like testing a car on a long road trip, not one lap.
FARS statistics:
Each paper consumed roughly 114 million tokens, far above typical writing tasks. FARS is still in a "trading compute for intelligence" phase, with algorithmic efficiency room to improve.
---
6. Limitations and Boundaries
FARS is not omnipotent. It has clear limits:
1. Domain limitation: currently only AI-domain research (AI4AI)—a pragmatic choice, since AI experiments run entirely on computers without physical equipment. 2. Compute dependency: requires a 160-GPU cluster, beyond the budget of individuals or small teams. 3. No human experiments: cannot do human factors research, psychology experiments, or manually annotated datasets. 4. Quality variance: while average quality is solid, some papers may be incremental and lack deep insight.
---
7. Philosophical Questions About the Future of Science
What is "good" research? Traditional standards: top venues, many citations. FARS standards: a clear hypothesis plus reliable verification, regardless of outcome. Which is better?
What is the human researcher's role? Possible answers: posing grand, interdisciplinary questions; making value judgments about what deserves study; interpreting and communicating findings; doing physical-world experiments.
Will discovery accelerate? If systems like FARS proliferate, scientific progress may grow exponentially—but new problems may emerge: information overload, quality control, research integrity.
---
8. FARS Is Not the Only Player
FARS's predecessors include:
What sets FARS apart: 1. End-to-end automation: from idea to paper, fully unattended 2. Scale: 100 papers produced live, not a few curated examples 3. First-principles design: no pandering to traditional academic formats, focus on knowledge itself 4. Transparency: all code and papers publicly released in real time
---
9. Conclusion: The Industrialization of Science
FARS represents a trend: the industrialization of science. Just as manufacturing moved from workshops to assembly lines, research may move from "individual genius's flash of insight" to "systematic, scaled knowledge production."
This doesn't mean human researchers will be replaced. Industrialized manufacturing still needs designers, engineers, and quality controllers. The future research ecosystem may need:
FARS is a beginning. It proves fully automated, end-to-end research is possible. Many limitations remain, but the concept holds.
228 hours, 100 papers. This is just the start.
---