After Eureka

Frontier AI systems can now transform vast quantities of compute and accumulated human knowledge into successful scientific search. That capability is real. What is not yet real is an institutional vocabulary for distinguishing cognitive achievement from computational expenditure. Contemporary prizes, benchmarks, and media narratives largely report the endpoint — the theorem, the structure, the solved instance — while the search that produced it remains off-ledger: how many attempts, at what depth, against what opportunity cost, under whose allocation rule.

This note is about that ledger. We are not arguing that hard problems should go unsolved. We are arguing that humanity has invented the ability to manufacture millions of reasonably intelligent research attempts, and that we have almost no institutions for deciding where those attempts should be spent.

Eureka was never the whole story

Classical scientific credit evolved in a world where serious attempts were scarce. A single mind, or a small collaboration, could burn years on a conjecture. Failure was expensive in human time. Success was therefore legible as achievement under constraint: insight under scarcity of attention, skill, and lifespan.

Industrialized reasoning changes the economics without yet changing the mythology. Systems can propose lemmas, search literature, rewrite proofs, generate counterexamples, and retry at volumes no laboratory could staff. The endpoint may still look like a eureka moment — a clean paper, a named result, a prize announcement. The process may have been closer to industrial search: ten million tries, filtered by verification, compressed into a narrative of discovery.

That is not a scandal about machines. It is a governance lag. Credit systems that only see the endpoint systematically misprice the resource that made the endpoint possible.

We wrote earlier that evaluation under saturation must measure lineage, not only polish; that multi-step reasoning is metered in energy as well as tokens; that verification capacity is a scarce institutional good. Those threads meet here as a single question: when reasoning scale becomes manufacturable, what counts as discovery — and who decides the spend?

The Clay problems as case study, not target

The Clay Mathematics Institute’s Millennium Prize Problems are a useful stress test precisely because their scientific value is hard to dismiss. Navier–Stokes existence and smoothness, P versus NP, the Riemann hypothesis — these are not fashion problems. A correct resolution of any of them would be an achievement of historical magnitude. Clay’s own framing is explicit: elevate awareness that mathematics still has a frontier, and recognize work of lasting importance.

None of that is in dispute here.

What the prizes — and peer institutions like them — were not designed to adjudicate is the new production function of machine-assisted search. A million-dollar purse and a multi-year community scrutiny process still evaluate a proposed solution. They do not ask, as first-order questions:

• How much compute was consumed relative to expected knowledge gain?
• What adjacent problems of comparable difficulty were deferred by that allocation?
• Can the search path be reconstructed, audited, and reproduced — or only the final proof object?
• Does the credit narrative report a cognitive leap, an industrial campaign, or both — and do readers know which?

Treat the Clay problems as a case study in institutional lag, not as a campaign target. The real issue is not whether Navier–Stokes deserves solving. It unquestionably does. The issue is that prize culture, media culture, and much of benchmark culture still speak as if the scarce input were genius alone — when the newly scarce inputs are verified reasoning depth, energy, expert attention for checking, and the opportunity cost of not spending those on something else.

Ten million tries is not a eureka moment. It may be excellent science. It may even be the only viable path to a result. But calling it eureka without the ledger is a category error — and category errors, repeated, become policy.

What we mean by industrialized reasoning

By industrialized reasoning we mean large-scale machine search over scientific hypotheses, proofs, designs, or experimental plans — characterized by:

• High attempt volume relative to human-only research cycles
• Accumulated knowledge as feedstock — corpora, prior papers, formal libraries, simulation traces
• Metered depth — multi-step chains whose cost compounds with context and iteration (contextual depth)
• Verification as bottleneck — the expensive step is often checking, not proposing
• Narrative compression — public accounts that collapse search into insight

Industrialization does not make results illegitimate. Steel mills did not make bridges fake. It does make allocation and credit accounting first-class scientific problems. When proposals are cheap and verification is dear, institutions that only celebrate winners will systematically underfund the checking infrastructure — and over-allocate prestige to whoever can afford the largest search budget.

That is a political economy, whether or not anyone intends it.

Four allocation criteria for scarce machine reasoning

We suggest that scientific institutions — funders, prize bodies, national labs, university consortia — begin treating large-scale machine reasoning as scarce scientific infrastructure, allocated less by prestige magnetism and more by explicit criteria.

1. Expected knowledge gain

Not “is this problem famous?” but “what would a successful or failed campaign teach us about the adjacent space?” Prestige problems can score highly here. So can unglamorous ones with high branching value. The point is to name the expected gain rather than inherit it from century-old fame.

2. Resource proportionality

Match attempt budget — compute, energy, expert verifier time — to the claim’s stakes and to institutional capacity. A campaign that consumes a nation’s research-compute envelope to shave a leaderboard point is a different object from one that opens a field. Proportionality is how rationing becomes principled rather than accidental.

3. Reproducibility of the search path

Publish not only the endpoint artifact but enough of the attempt distribution, tooling, and failure modes that another group could contest the result. If the path is proprietary or unlogged, the scientific claim is thinner than the press release. Lineage applies to machine search as much as to agent-polished prose.

4. Social opportunity cost

Every megawatt-hour and every senior verifier-hour spent on one campaign is not spent on epidemic modelling, materials for energy transition, disaster warning, or education research. Opportunity cost is not an argument against deep mathematics. It is an argument against pretending the trade-off does not exist.

These four do not replace peer review. They precede and discipline the decision to industrialize a search in the first place.

What we are not claiming

We are not claiming that machine-assisted discovery is fraudulent. We are not claiming prizes should be abolished. We are not claiming that only “small” problems deserve compute. We are not asking mathematicians to apologize for ambition.

We are claiming that as machine reasoning becomes scalable, institutions must separate:

Report the endpoint	Also report the ledger
Solved / unsolved	Attempts, depth, energy band
Named credit	Tooling, search policy, verifier labour
Prize eligibility	Allocation rationale and alternatives foregone
Benchmark SOTA	Cost per verified conclusion

Without that separation, scientific culture will keep mistaking expenditure for insight — and will keep allocating the next ten million tries by habit rather than by design.

Open questions we are pursuing

• What minimal attempt ledger should accompany machine-assisted claims in mathematics, materials, and biology — short of dumping every token?
• Can funders score proposals on expected knowledge gain per verified reasoning-hour, not only on novelty language?
• How should credit be shared when the decisive step is a human conjecture amplified by industrial search — or the reverse?
• Where do moral crumple zones appear when a celebrated result cannot be checked by the community that must live with its consequences?
• Which civilian research agendas are currently under-attempted relative to their social return, because prestige allocation pulls compute elsewhere?

Working claim. As machine reasoning becomes scalable, scientific institutions must distinguish cognitive achievement from computational expenditure — and govern large-scale search as scarce infrastructure: allocated by expected knowledge gain, resource proportionality, reproducibility, and social opportunity cost, not by prestige alone.

The Clay problems illuminate the gap because they are worthy. That is exactly why they make the argument harder to dismiss. The capacity to manufacture intelligent attempts is new. The institutions that decide where those attempts go are not. Closing that gap is, we think, one of the genuinely important questions of this scientific decade.

If your mandate includes research allocation, scientific infrastructure, or evaluation design for machine-assisted discovery, start a conversation. We are looking for partners and funders who want the ledger — not only the eureka.