Why the American Science Machine Takes Twenty Months to Say Yes

A researcher sits down to write a grant application. Two to four months of drafting, plus preliminary data for the very project the money is meant to fund, plus her university’s internal review. Submit to the National Institutes of Health in February and she will be fortunate to hear by December. Some awards run twenty months from submission to decision, roughly how long Boeing needed to take the 747 from drawing board to production line.

That comparison is not mine. It appears in a report the White House Office of Science and Technology Policy published on 21 July, written by presidential science adviser Michael Kratsios and framed as successor to Vannevar Bush’s 1945 Science: The Endless Frontier. Bush’s document gave America the National Science Foundation and eight decades of scientific architecture. This one argues the architecture has seized up.

A sitting administration wrote this document, and stretches read that way; chapters on merit and on foreign doctoral students carry political freight. But the diagnosis of the funding machinery rests on published metascience, which existed before the report and will outlast it. Start with drugs, where the arithmetic is nearly comic in its bleakness. Researchers have coined “Eroom’s law”, Moore’s law spelled backwards, for pharmaceutical discovery: new approvals per inflation-adjusted billion dollars have fallen roughly eightyfold since 1950, halving about every nine years. The NIH budget more than doubled since the 1990s. Breakthrough treatments did not follow.

The Ladder, Not the Tree

The pattern turns up elsewhere. Sustaining Moore’s law itself now takes more than eighteen times as many researchers per doubling of transistor density as in the early 1970s, implying a roughly 7% annual decline in ideas productivity. Years of life saved per clinical trial peaked in the mid-1980s, then fell away. One obvious reading: this is just what maturity looks like, low-hanging fruit picked, ladder getting longer. The report will not have it, and reaches for a good historical needle. Max Planck was told by a professor that physics was more or less finished, a mature field on the order of geometry, not long before Planck went off and dismantled the whole picture of physical reality; a nineteenth-century surgeon wrote that there could not always be fresh fields for conquest by the knife, and then came research hospitals, modern medical education, a century of conquest. The tree only looked bare. Nobody had brought the right ladder.

Between 1980 and 2008 the average age of NIH principal investigators climbed from 39 to 51. Average age at first grant rose from 36 to 42, which now exceeds the typical age at which Nobel-winning work in comparable fields was actually done. Our young researcher adjusts her ambitions accordingly. She learns that a long shot at an established paradigm may cost her tenure, and that by the time tenure arrives her most creative years will largely be spent; she learns that slicing work into the smallest publishable units (salami-slicing, they call it) pays better than swinging hard.

Forty-Eight Hours Against Twenty Months

A grim footnote: when a star biomedical scientist dies unexpectedly, outsider contributions surge into the space their network had dominated, and those papers are more likely to become highly cited. Planck’s much-mangled line about science advancing one funeral at a time turns out to be measurable. Peer review compounds it. Reviewers assessing the same NIH proposal often reach contradictory conclusions about the credibility of the science. When an agency funds one application in ten, noise drowns signal, and consensus panels drift toward the least divisive idea rather than the most promising. The report is sharp on why nobody fixes this: the panel works well as a liability shield, since decisions get attributed to the scientific community rather than to anyone who might answer for a bad call. A mechanism built to hedge risk ends up precluding the risk-taking that breakthrough science needs.

The remedies come mostly from metascience, and several have data behind them. The Howard Hughes Medical Institute funds people rather than projects, roughly $10 million over seven years, minimal reporting, tolerance for early failure; against similarly accomplished federally funded peers, its investigators produced high-impact papers at nearly double the rate and were far likelier to explore novel lines. A Danish foundation has trialled “golden tickets” letting one reviewer champion a proposal consensus would bin. And a privately funded American programme made funding decisions in 48 hours rather than six to nine months, on applications taking half an hour to prepare, with no measurable loss of quality. Forty-eight hours. Against twenty months.

The Verification Problem

Then comes AI, handled with a caution the surrounding rhetoric does not prepare you for. In September 2025 an AI-assisted effort formalised the Prime Number Theorem in three weeks, after twenty-odd mathematicians had spent 18 months on it and stayed stuck on complex analysis. But the report reaches for forest fires. A century of suppressing small burns looked like a run of obvious successes, right up until the accumulated fuel produced infernos nobody could contain, and AI could do much the same to science, making each researcher more productive while dysfunction piles up underneath.

Because the knowledge base is shaky. Fewer than 40 of 100 psychology studies replicated in one large attempt; irreproducible preclinical biomedical work may misdirect $28 billion a year. A 2009 Alzheimer’s paper in a top journal was shown irreproducible by 2012 and drug development was terminated on that basis, yet it gathered more than 800 citations and steered priorities for another decade. Its lead author became a university president. Retraction came fifteen years after publication. Train a model on that corpus and you do not repair the corpus. You extend it, faster. Generation has got exponentially cheaper; verification has not, and the answer proposed is verification infrastructure built at the generator’s scale: machine-auditable replication packages, open standards, prizes for disproving influential work. Mathematics offers the proof of concept, since checking a proof is easier than finding one. Whether that asymmetry holds outside mathematics is the question the document poses and cannot answer.

What the report cannot supply is anyone outside the administration saying whether it is right. No independent scientists are quoted, and the contested chapters are not presented as contested. The section on foreign doctoral students is the clearest case: graduate school is America’s principal pipeline for recruiting the world’s scientists, and a Peterson Institute brief puts the economic cost of narrowing that door at up to $481 billion a year — a figure the report does not engage with. The same applies to its claims about diversity criteria and pandemic school closures, all live disputes, none flagged as such. Read it as a policy brief with a good literature review attached.

The strangest passages imagine what remains once you take the machinery seriously: a funder posting a bounty for a validated therapeutic target, an agent noticing a lead and posting a smaller bounty for replication, others bidding, contracting a cloud laboratory over the internet. None of it exists in mature form and may never assemble that way. But the underlying observation stands: the journal system was built in the seventeenth century for a few hundred correspondents writing letters, and now serves nine million researchers publishing millions of papers a year. Nobody would design it this way now.

Meanwhile our researcher is still filling in forms. She will be at it until roughly December.

Frequently Asked Questions