Zach Weinberg on why AI won't cure disease, the NIH funding crisis, and pharma tariffs

Summary

Zach Weinberg, founder of Curie, makes the case that most excitement around AI and drug discovery is misplaced, and that the NIH funding crisis poses a more serious threat to American health outcomes than almost anything else being debated in Washington.

Biology vs. drugs

Understanding how the body works—cell signaling, disease mechanisms, what actually causes Alzheimer's—is biology. Creating a molecule to intervene in that process is drug development. You cannot patent biology, so there is no private-sector business model for funding it. Government money is structurally irreplaceable: without NIH-funded basic science, there is no target for any drug program to aim at.

Alzheimer's illustrates the problem. Science has identified tau and beta-amyloid as abnormalities in patient brains, but nobody knows whether those proteins drive the disease or are simply passengers. Without knowing the cause, there is nothing meaningful to drug. The gap between understanding biology and having a druggable target—not manufacturing capacity, not regulatory speed—is the real bottleneck.

The NIH budget is currently around $30 billion. If the US economy were ten times larger, that budget could reach $500 billion, and the pace of biological discovery would scale with it. Weinberg argues that economic growth matters primarily because tax revenue funds science.

Why AI won't cure disease

AI is a useful tool for a narrow slice of the problem, not a platform shift. Drug development is a roughly 40-step process that cycles through physical iteration: make the molecule, test it in a petri dish, tweak it, test it in a mouse, then a dog or monkey, then humans. The computer's suggestion still has to be synthesized and tested in the physical world at every stage. It is atoms, not bits.

The data problem is severe. Powerful models require massive labeled datasets with fast feedback loops. Biotech has neither. The training data available is perhaps one-billionth the size of what exists in the digital world. Most of it comes from petri dishes, not from cells behaving inside real biological systems, where the dynamics are different in ways science cannot yet fully characterize.

AlphaFold solving protein structure is genuinely useful for small-molecule design. But the leap from "AI can predict protein folding" to "AI will design drugs" skips over the biology layer entirely. If you do not know what causes a disease, there is nothing for the model to optimize toward.

Pharma tariffs

The strategic value in American pharma sits in drug discovery and development—the high-risk, high-cost phase before approval. Manufacturing the approved pill is the last mile. It employs relatively few people, involves largely quality-control processes, and can be done more cheaply elsewhere without meaningful national-security implications. Tariffing that final manufacturing step raises drug prices for Americans while doing nothing to strengthen the parts of the pipeline that actually matter.

Where AI in healthcare does apply

Administrative paperwork—prior authorizations, scheduling, insurance coordination—is well-suited to language models, and improvement there would be real if unglamorous. Differential diagnosis is the more interesting case: synthesizing patient symptoms against the medical literature to produce probability-weighted assessments is pattern-matching work that models should get better at over time.

The biomarker consumer boom

Companies like Function and Superpower selling broad biomarker panels to health-conscious consumers face a fundamental problem: data by itself is not valuable, and many biomarkers are poorly validated. Some fluctuate with something as mundane as morning coffee. Wide blood-based screening will mostly make healthy, neurotic tech workers slightly more anxious rather than catching meaningful disease. The more important healthcare problems are chronic disease and aging, which disproportionately affect people who are not the customer base for these products.

There is also a real clinical risk. Finding an ambiguous result on a scan or panel creates pressure to investigate further, and investigation carries harm. A biopsy—especially in the GI tract, especially in older patients—is not risk-free. More healthcare is not always better, and that is the part of the equation the quantified-self industry consistently underweights.