Summary

Bhattacharya’s core operating thesis is that NIH should be managed like a venture portfolio: tolerate productive failures so one field-changing win can make the portfolio a success despite them. He says recent NIH culture favored six-to-eight-year-old ideas over zero-to-two-year-old ones and produced “fewer advances per dollar”; institute directors will be judged on portfolio-wide health and knowledge gains, not whether every grant succeeds.
The immediate policy package pairs a $50 million autism data-science program—13 teams to receive grants from 250 applicants—with action on an old drug and a pregnancy-use caution. Leucovorin may restore speech in “20%…I think” and improve up to 60% only among autistic children with a relevant brain-folate problem; acetaminophen evidence remains correlational and controversial, so Bhattacharya frames forthcoming FDA guidance as prudence, “not…panic,” while CMS changes payment for leucovorin.
His reform bottleneck is not simply a shortage of research dollars but a system that rewards publication and conservatism over replication and fresh ideas. A peer-reviewed paper is an investigator’s belief, “not…truth”; independent replication should be the standard, while centralized review, auditable foreign grants, and permission to publish productive failures are intended to improve accountability and learning.
The NIH’s capital-allocation split is explicitly democratic at the top and expert-driven inside each disease portfolio. Congress and the president should reflect public needs across cancer, HIV, heart disease, diabetes and pediatrics, while scientists select promising research opportunities; Bhattacharya rejects a “philosopher king” and points to flat U.S. life expectancy as evidence that science is not translating enough into health, while noting that NIH is not the only cause or answer.
The talent pipeline now delays a first major NIH grant from a median age of about 35 in the 1980s to the mid-40s, systematically aging the idea pool. Because his study found ordinary scientists’ ideas age one year per chronological year—versus one per two years for Nobel winners—he wants institute directors rewarded for early-career funding, strategic-plan coverage, and senior investigators’ mentorship.
Public-health trust will require officials to match confidence to evidence, say “I don’t know” when warranted, and treat citizens as partners rather than subjects. Bhattacharya blames unsupported pandemic measures and overconfidence, but resists false humility: he cites roughly 95% MMR uptake versus about 13% COVID vaccination among children as evidence that “the American people are not stupid” and can distinguish stronger evidence from weaker evidence. He says rebuilding trust will take a long time.
AI is an augmentation thesis, not a scientist-replacement thesis: it can narrow protein targets, improve radiology, and return doctors’ attention to patients, but its uses require research, including to address hallucinations. Bhattacharya says AlphaFold has “turbocharg[ed]” drug development and NIH is building privacy-protected tools, yet people writing as many as 60 visibly AI-generated applications have already overwhelmed review; a new limit is described as “like, 6 a year—6 a cycle or something.” Discovery still depends on scientists who “keep knocking on the door.”

Deep dive

1. The autism initiative couples research capital with near-term policy

Six months after Secretary Kennedy asked for answers for families, Bhattacharya launched a $50 million autism data-science initiative. With prevalence tentatively cited as the CDC’s latest one-in-31 figure, behavioral therapies that do not work very well for many children, and no clear causal or preventive answers, 250 teams applied and 13 will receive large grants intended to produce findings over the next few years.
One near-term treatment announcement centers on leucovorin, an old folinic-acid drug that can help deliver folate to the brain. Bhattacharya says doctors have seen benefits among autistic children with the relevant folate-processing deficiency: “20% of the kids, I think, restore speech,” while up to 60% improve—but he repeatedly cautions that this does not apply to every autistic child.
The acetaminophen announcement is deliberately narrower: a recent study highlighted by the dean of Harvard’s T.H. Chan School of Public Health found an association between use during pregnancy and later autism diagnoses. Bhattacharya calls the literature controversial and the message prudential—use it when genuinely needed, such as for high fever—because “that’s not the kind of result that should panic anybody.”
FDA Commissioner Marty Makary is working on revised pregnancy guidance, while CMS will change payment for leucovorin. Bhattacharya casts the package as cross-agency execution; a parallel preterm-birth effort targets another unresolved gap, with U.S. outcomes worse than Europe’s and prenatal-care access important but “not the whole answer.”

2. Replication—not publication—is the proposed standard of truth

Bhattacharya’s line in the sand: “The standard for truth in science ought to be replication.” A journal article records a scientist’s evidence-backed belief, but “the fact that it’s published in a journal doesn’t mean it’s right”; confidence should rise only when independent teams reach the same result.
The replication crisis is partly structural. Science has grown too specialized for researchers to routinely check one another, and a career spent reproducing other people’s work offers little path to “a professorship at a fancy university.” Negative replications often go unreported, leaving the original claim standing without the scrutiny that would expose scientists’ natural ability to convince themselves they are right.
His eggs example captures the communication cost: evidence in 1985 left him afraid to eat eggs, only for later science to reverse the message. Because science is intrinsically hard, NIH must invest in replication, demand reproducibility, and replace appeals to “high authority” with evidence that survives independent testing.

3. NIH needs venture-style tolerance for productive failure

Bhattacharya’s portfolio analogy is explicit: if a16z funds 50 projects, 49 fail, and the 50th becomes “Google or something,” the portfolio is a tremendous success. Scientists whose experiments fail productively should likewise not be punished and should have a venue to publish “what they learned from it.”
The historical comparison suggests NIH has moved away from that model. In the 1980s and 1990s, funded ideas were often zero, one, or two years old; by the 2000s and 2010s, the typical project rested on ideas six, seven, or eight years old. His conclusion: “We just became too scared of trying new ideas out.”
Peer review compounds the conservatism. An established reviewer sees a proposal challenging decades of personal work, says “there’s no way it can work,” and can quickly create panel consensus. The hosts recognize the venture analogue: partnerships seeking unanimous consent lose bold bets because one skeptic can veto the unfamiliar.
Execution reform also includes auditable foreign collaborations and uniform review. After NIH could not audit money sent to the Wuhan lab, Bhattacharya introduced tracking intended to preserve—not end—international science; he also moved parallel institute-level reviews into the Center for Scientific Review so the NIH’s 27 institutes use a common process.

4. Disease allocation is political; scientific selection happens downstream

Agarwala’s framework divides NIH stewardship into allocation and execution: determine how more than $35 billion is divided among disease areas, then choose investigators, monitor data, manage partners, and sustain risk-taking. Bhattacharya accepts the split but stresses that these layers answer to different kinds of judgment.
Congress and the president should set broad disease allocations because public money should address public needs; scientists then assess opportunities and build portfolios within those mandates. “There’s no philosopher king” who can objectively dictate the correct split among HIV, cancer, pediatric conditions, heart disease, or autism.
The hosts’ pushback—worth keeping—is that citizens may have deliberately delegated biomedical judgment to experts and cannot evaluate obscure diseases or scientific tractability. Bhattacharya answers with early-1980s HIV: the initial NIH response was inadequate, and patient political organization forced the institution to recognize the threat. Scientific expertise alone did not mediate fairly among populations.
His outcome test is practical health. U.S. life expectancy has not increased in roughly a decade and a half, and chronic burdens include heart disease, kidney failure, type 1 and type 2 diabetes, autism, and rising cancer incidence. HIV’s advances deserve celebration, but with about 40,000 people getting HIV “last year,” success there neither ends the need for investment nor excuses neglect elsewhere. He also cautions that NIH science is not the only explanation for these outcomes.

5. The aging grant pipeline is also aging the ideas

In the 1980s, the median age for receiving a first large NIH grant was about 35; now the milestone arrives in the mid-40s. Meanwhile, biomedical trainees face one, two, or three postdocs before an assistant-professor opportunity, causing promising researchers to leave before they can test independent ideas.
Bhattacharya’s own study found that the ideas in a typical scientist’s published work age one year for every year the scientist ages. The best researchers fight that drift: among Nobel Prize winners, ideas age roughly one year per two years of chronological age. “If you want the newest ideas, you have to let the young people have their try.”
His institutional lever is to give institute directors more portfolio authority and judge them on health impact, major biological advances, and alignment with strategic plans—not individual-grant perfection. Otherwise, peer-review scores can fund 10 proposals in one strategic area while leaving another major priority empty.
He does not identify the training pipeline—including predoctoral, MD-PhD, and postdoctoral support—as the weakest link; the failure comes at independence, where K awards and the transition to assistant professorship remain difficult. Senior investigators will therefore be evaluated partly on whether they mentor and advance early-career colleagues, rather than merely extending their own established programs.

6. Academic freedom and publishing economics are part of research infrastructure

Internal NIH scientists previously needed supervisors’ permission to publish. Bhattacharya ended that requirement: “People are going to publish research that I don’t agree with. That’s wonderful.” Universities, in his view, likewise cannot provide excellent research environments without protecting scientists’ ability to explore and dissent.
He distinguishes that commitment from the administration’s pressure on universities over issues including antisemitism. Whatever the surrounding conflict, the scientific requirement remains categorical: researchers must be free to say what they think and follow evidence without institutional preclearance.
Publishing concentration creates another choke point. A very small number of for-profit journal companies can charge around $10,000 to publish research funded by Americans, then ask readers for another $50 to $100 at the paywall. NIH has removed that reader paywall for NIH-funded work, though Bhattacharya says more openness in scientific publishing still requires further policy.

7. Trust depends on admitting uncertainty without performing helplessness

Bhattacharya roots today’s mistrust in pandemic practices: ubiquitous plexiglass, masks required while entering restaurants but removed when seated, and school closures imposed on weak evidence. Children remain “years behind” educationally, he argues, so distrust is an understandable response to officials who expressed certainty while changing lives for the worse.
His first repair mechanism is “gold-standard science”: replication, unbiased peer review, reproducibility, and candor about limitations. The second is posture. Public-health officials are “servants of the people,” yet during the pandemic they often appeared to sit above citizens, attaching vaccination decisions to jobs and access rather than acting as partners in scientific investigation and public health.
The third-year-medical-student analogy supplies the operating rule. A white coat tempts the inexperienced student to “start freelancing” when a patient wants an answer; the responsible response is, “I don’t know. I’m going to look it up.” In a new pandemic, officials should state uncertainty, explain how they are seeking answers, and consult people who know more.
The hosts press the other side: humility about a novel virus must not dissolve confidence in evidence-backed tools such as longstanding childhood vaccines. Bhattacharya agrees—MMR prevents a potentially deadly disease and does not require false humility—while preserving room for dissent. Roughly 95% MMR uptake versus about 13% COVID vaccination among children suggests to him that citizens respond rationally to differences in evidence.

8. AI can focus scientific labor, but cannot manufacture conviction

On chronic disease, Bhattacharya highlights observational work involving Zostavax, an old shingles vaccine. A Stanford colleague found it associated with up to a 20% or 30% reduction in cognitive decline or Alzheimer’s risk; despite the vaccine no longer being used because it performed poorly against shingles, Bhattacharya sees a potentially cheap, safe way to delay or prevent many cases—an idea needing “a little bit of scientific love.”
AlphaFold demonstrates AI’s research leverage: predicted protein structures can narrow likely target sites and drug candidates before expensive laboratory work. The lab work remains necessary, but computation focuses it. In care delivery, AI might help radiologists catch missed findings or draft electronic records so physicians look at patients instead of screens.
Bhattacharya rejects a future where AI writes institute roadmaps, submits grants, and reviews grants. Models summarize existing knowledge better than they challenge paradigms; some people are writing as many as 60 visibly AI-generated applications, flooding review with noise and prompting a limit he described as “like, 6 a year—6 a cycle or something.” NIH’s privacy-protected systems are therefore “an augmentation of capacity rather than a substitution.”
His closing portfolio answer is “yes, yes, yes” to better disease management, new molecules, and changes in how people live because he cannot predict the winner. He points to GLP-1s—specifically a GIP/GLP-1 molecule—and the first decline in average U.S. body weight in decades “last year,” then invokes Max Perutz spending a decade on myoglobin despite professors urging an easier problem: NIH should leave room for scientists who “keep knocking on the door” until a field changes.

Replit CEO on The Career of Coding, AGI, and Vibe Coding Wars w/ Amjad Masad, Dave B & Salim Vol.74 Sports May Be Humanity’s Last Front Against AI — Crossover Episode, First and Second Half