When Hypotheses Become Cheap
Short summary
AI has made generating hypotheses and running simulations cheap, but the actual experimental verification remains slow and expensive — a gap that will determine how much AI truly accelerates scientific discovery. The bottleneck isn't computing power; it's the physical work of building things, measuring them, and getting reproducible results.
- In formal math and software, verification is nearly free once problems are machine-checkable (Lean proofs, AlphaEvolve), so the limiting factor becomes compute. But most science requires physical experiments that can't be sped up by better algorithms.
- Amdahl's law shows the hard limit: if 80% of discovery time is lab work, infinitely fast hypothesis generation only delivers 1.25× speedup. Better models widen the verification gap rather than close it.
- Cheap compute creates "orphaned hypotheses" — promising candidates generated by grad students and small teams that go untested because experimental capacity is concentrated in a few rich institutions and countries. Accessible, affordable lab tools could democratize who can run the full loop of science.
- Surrogates like AlphaFold work only when built on high-quality experimental data and validated with real measurements. A country or company that scales compute without matching lab infrastructure and data capacity will have fast models that are confidently wrong.
Full summary
AI has made generating hypotheses, reading papers, and running simulations dramatically faster, but it's barely touched the hard part: actually building things and measuring them. Math and software show what happens when verification becomes nearly free—AlphaEvolve found a scheduling heuristic that recovers 0.7% of Google's global compute, and AXLE solved all 12 Putnam problems by treating proof-checking as a machine problem. But most of science doesn't work that way. A drug that looks perfect in simulation can be toxic in mice. A predicted battery cathode can fail after two hundred cycles. You need physical reality to cooperate, and no amount of compute changes that. The author calls this "verification elasticity"—how much more checking you can buy when compute gets cheaper. Software has extremely high elasticity. Wet chemistry and biology have low elasticity because every check requires something physical to happen. When 80% of discovery time is bench work, making hypothesis generation infinitely fast buys you only a 1.25× speedup overall.
The gap is already creating visible problems. Anthropic's 950 Claude agents spent 21 hours sifting through 200,000 DNA sequences and flagged a potentially new reverse transcriptase system. Now humans need months of lab work to figure out what it actually does. DeepMind's GNoME predicted 2.2 million crystal structures, but Berkeley's A-Lab discovered that interpreting measurements with confidence was harder than making materials quickly—a 2026 correction narrowed their novelty claims and marked some results inconclusive. The pattern repeats: search collapses to hours or days, everything after that stays at human and physical speed.
This creates a weird new inequality. A materials grad student can now screen hundreds of thousands of battery compounds in a week using open models and rented GPUs. A two-person startup can design proteins on cloud compute. Neither can afford to test their best ideas. The author calls these "orphaned hypotheses"—good ideas trapped behind experimental capacity that's concentrated in a few rich countries and institutions. A grad student in Lagos with three promising battery candidates has no affordable way to verify them, while one in Boston can. The real infrastructure gap isn't compute. It's benchtop instruments that cost thousands instead of millions, 3D-printed reactors, open-source lab robots, and cloud labs priced per experiment. Personal computers democratized computing. Something similar needs to happen to experimentation, or the speed of AI-generated ideas will just widen the gap between who can ask questions and who can answer them.