In 2025, Andrej Karpathy coined a term that captured the zeitgeist of the AI era: vibe coding. The idea was simple — let an AI write code from a natural-language prompt, and then iterate on it in real time. No formal specs, no five-year plan. Just a vibe. It was messy, chaotic, and surprisingly effective. That same spirit could be the antidote to American science's worst habits.
American science is stuck in a century-old operating system. Federal R&D spending as a share of GDP has fallen from 1.9% in 1964 to just 0.6% in 2022, according to the AAAS. The NIH, the world's largest biomedical funder, takes an average of 18 months to process a grant. In that time, a vibe coder would have shipped 18 prototypes. The result? A structural inability to take risks, a replication crisis, and a 'valley of death' that kills promising discoveries before they reach the market.
Vibe coding isn't about abandoning rigor. It's about removing the friction between an idea and a test. In programming, that means a natural language prompt and a few calls to an AI API. In science, it means rethinking the institutional systems that govern how research gets funded, reviewed, and shared. The case for overhauling American science is not a call for chaos — it's a call to apply the principles of vibe coding to the laboratory.
The Diagnosis: A System Working Against Discovery
Consider the lifecycle of a typical idea in academic science. First, a researcher spends weeks writing a proposal. The proposal is reviewed by a panel of peers who are often competing with the applicant for the same funds. It then gets scored, revised, resubmitted, and finally funded — or more often, rejected. According to a study by the American Psychological Association, researchers report spending up to half their time on administrative duties rather than actual experiments.
Worse, the system punishes failure. A failed experiment, even one with high methodological quality, is rarely published. Funders see it as waste. As a result, scientists self-censor, choosing incremental questions over breakthroughs. This is the opposite of vibe coding, where a dead end is just a 'quick branch' and you move on.
What Vibe Coding Can Teach Science
The philosophy behind vibe coding is remarkably aligned with how science should work: hypothesis, experiment, iterate. The key difference is the speed and the cost of iteration. In coding, you can run a test in seconds. In science, an experiment might take years and cost millions. But that doesn't mean the system can't be made leaner.
One approach is to move to portfolio funding, as pioneered by DARPA. Instead of funding a single long-term project, DARPA funds a portfolio of short-term projects with clear go/no-go points. It's essentially the AI 'beam search' applied to research. This gives scientists the freedom to fail fast and pivot.
Another lesson from vibe coding is the value of low-key tools. For example, AI assistants can now draft sections of grant proposals, scan thousands of papers, and even design preliminary experiments. Yet most institutions still treat these tools with suspicion. That mindset needs to change.
A Five-Step Framework for the Overhaul
Here is a practical, step-by-step framework that could transform American science, inspired by the agile, AI-native approach of vibe coding.
Step 1: Flip the Funding Model
Replace the current grant cycle with a 'research account' system. Each lab receives a lump-sum annual budget, with no strings attached. They can spend it on whatever experiments they think are promising. This eliminates the enormous overhead of proposal writing and peer review. Everyone from the NSF to the NIH has tested versions of this, but none have made it the default.
Step 2: Let AI Triagulate Peer Review
Peer review is the gold standard, but the gold has tarnished. A 2019 study in Nature revealed that AI models could predict whether a grant would be funded with 80% accuracy — not because the science is good, but because the language mirrors existing work. That's a badge of conservatism. Instead, we should use AI to weed out mediocre proposals, and only then send the top 20% to human reviewers for deep analysis. This cuts review time dramatically.
Step 3: Open Everything
Open science is not an optional extra — it's the foundation. The current White House policy requires open data for federally funded research, but enforcement is spotty. We need to go further: open code, open notebooks, open negative results. GitHub is already the de facto platform for many data scientists. ASI Biont supports integration with GitHub via API — learn more at asibiont.com/courses. This kind of seamless toolchain would let scientists share and build on each other's work in real time.
Step 4: Create a 'Risky Research' Lottery
Every year, federal agencies should allocate 5% of their budget to high-risk, high-reward projects selected by a lottery. Not peer review — a lottery. Why? Because peer review is designed to predict whether a project will succeed, and by definition, high-risk projects have low predictive success. DARPA's most famous inventions, including the internet, came from ideas that initially seemed absurd. The MacArthur Foundation's 'genius grants' already work this way.
Step 5: Rotate Academia and Industry
Overhaul also means breaking the glass wall between academia and industry. Currently, a scientist in a university who spends more than a year in a company often loses their tenure-track position. That's insane. We should create a standardized 'career bridge' that allows scientists to flow back and forth. This aligns with vibe coding's culture of continuous learning and adaptation, and it would massively accelerate the validation of discoveries.
A Vibe-Coded Example: Analyzing the Replication Problem
To see how these principles could work, consider a simple Python script a scientist might use to expose replication issues. The code is intentionally minimal — the essence of vibe coding.
import pandas as pd
# Load data from the Center for Open Science's Replication Project
df = pd.read_csv("replication_studies.csv")
print(df.groupby("discipline")["replicated"].mean())
In real life, you'd replace the CSV with a live API, but the point is that in just a few lines, you can get a birds-eye view of which fields are failing to replicate. In the same spirit, a scientist can rapidly prototype an AI tool to screen their own papers before submission.
Of course, this is not a full analysis. But the mindset is the message: stop polishing your proposal, start testing your assumptions.
The Backlash and the Bottom Line
Skeptics will say that science needs slow, careful review, and that AI will introduce bias. They're right — but the current system has its own biases and is anything but fast. In a 2016 Nature survey, over 70% of researchers said the replication crisis is real, and more than half said they had failed to reproduce their own experiments. That's the ultimate 'code smell'.
The case for overhauling American science is not about discarding the scientific method. It's about removing the heavy, slow layers of bureaucracy that choke it. Vibe coding proved that even a chaotic, AI-driven workflow can produce working software. Applied to science, it might produce something even bigger: a system that is affordable, responsive, and genuinely built for discovery.
The next frontier isn't just AI for science — it's using the philosophy of vibe coding to redesign the entire engine. Let's stop writing 700-page grant proposals and start running experiments.
Comments