When Ideas Became Cheap
For a long time, science followed a familiar script.
A human scientist noticed something strange, formed a hypothesis, designed an experiment, gathered evidence, and slowly refined a theory. Computers helped with calculation, but the core act of discovery still belonged to people: intuition first, verification later.
That is no longer quite true.
AI is not just making science faster. It is changing the shape of discovery itself. The biggest shift is simple: the cost of generating ideas has collapsed. What used to be scarce was hypotheses. What is scarce now is judgment.
Donald Knuth, Terence Tao, and Andrej Karpathy each reveal a different part of this change.
Knuthβs recent paper, Claudeβs Cycles, is one of the clearest signals that something real has shifted. In it, Knuth describes how Claude helped solve a combinatorial problem he had been thinking about and wrote that he needed to βreviseβ his opinions about generative AI. For someone as careful and skeptical as Knuth, that is not a throwaway remark. It suggests that AI is now capable of contributing to genuine exploratory work, not just summarizing, coding, or imitating style. οΏΌ
Tao points to the next consequence. In a recent interview, he said that current AI can be βa trustworthy co-author if used correctly.β He also noted that AI has made it much cheaper for him to produce code, plots, and numerical explorations, which changes the kinds of things he includes in papers. That is an important observation. AI does not just speed up the old workflow. It expands the space of what researchers can afford to explore. οΏΌ
Karpathy, as usual, turns the idea into a loop you can actually run. His autoresearch project gives an AI agent a small but real training setup, lets it modify code, run experiments, check results, and iterate overnight. The point is not just that AI can assist research. The point is that AI can now participate in the full experimental cycle: propose, run, evaluate, repeat. οΏΌ
Put those three together and the change becomes obvious.
Knuth shows that AI can help find ideas. Tao shows that this changes the researcherβs role. Karpathy shows what the new workflow looks like in practice.
Science used to be bottlenecked by idea generation. Now it is increasingly bottlenecked by filtering and verification.
And they are not alone in seeing this.
Demis Hassabis has described AI as opening a new era of scientific discovery. Jeff Clune and collaborators pushed this further with The AI Scientist, a system designed to generate hypotheses, run experiments, analyze results, and draft papers. James Zouβs βVirtual Labβ applies similar multi-agent ideas to scientific problem-solving, especially in biomedicine. Across these efforts, the pattern is the same: fewer manually driven loops, more machine-scale exploration. οΏΌ
Even the rhetoric from frontier labs reflects this shift. Sam Altman has argued that the world is building something like a βbrain for the world.β Dario Amodei has suggested that AI could compress decades of scientific progress into a much shorter period. The exact timelines may be debated, but the direction is clear: discovery is becoming more automated, more iterative, and more scalable. οΏΌ
Still, this does not mean βthe AI scientistβ replaces the human scientist.
If anything, the human role becomes more important at a higher level.
When ideas are cheap, taste matters more. When experiments are easy to run, choosing the right question matters more. When models can generate thousands of possibilities, the real skill is knowing what is worth believing.
That is why the best metaphor for the future scientist may not be the lone genius, but the conductor.
A conductor does not play every instrument. A conductor sets direction, hears what does not fit, decides what deserves emphasis, and turns many voices into one coherent movement. That increasingly looks like the job of the modern researcher too.
AI has not ended science as a human activity.
But it has changed the scientific method.
The old model was: human thinks, machine helps.
The new model is closer to: machine proposes, human judges, and both iterate together.
That is a very different world.
And we are already in it.