The Consumer Playbook That Put Medical AI in 100,000 Exam Rooms

Asked how many doctors will use OpenEvidence today, Daniel Nadler does not hedge. Probably more than 100,000 physicians in the United States, he tells host Pat Grady, and more globally. A year earlier the number would have been a thousand, or "fewer than zero." There are only about a million doctors in the country serving more than 340 million people, a shortage Nadler flags as a problem in its own right.

On a monthly basis, he says, about 100,000 of those physicians use the product. Roughly 300,000 to 400,000 monthly active users touch the system, including more than 200,000 who log in and ask questions. Depending on whether you count licensed physicians or those actually practicing, that is somewhere between 10 and 25 percent of U.S. doctors. Grady, investing from Sequoia's technology seat rather than a healthcare franchise, notes that 100,000 physicians overnight is not how health systems usually move.

Nadler says the refrain that healthcare is impossible to break into is usually earned, because almost everyone bangs on the same door. The standard path is to climb an integrated delivery network, book the CMO or CTO, and wait three or four months for meeting one even with good connections. That meeting goes well. The team high-fives. The follow-up is a Responsible AI committee another three months out.

By the time you present, committee members have changed, the hospital's AI posture has changed, and maybe the presidential administration has changed, with JD Vance saying the Responsible AI approach itself has to change. A year in, you are in meeting three or four, and no doctor along the way has actually used the product. Nadler had already sold Kensho, a successful AI company, and recognized the pattern as large organizations, not a healthcare peculiarity.

This second company was mission-driven. It would not have felt like much of an accomplishment, he says, to have nominally started an impact business in medicine and then have no doctors using it because the team was waiting for meeting six, seven, or eight. So OpenEvidence took the opposite path. "Doctors are people too. Doctors are consumers too." Make something profession-changing, put it on the App Store for free, and skip the marketing budget.

All of the growth, Nadler says, is doctor to doctor. He compares it to early Tesla: car commercials were once one of advertising's largest categories, and dogma held that you could not sell automobiles without enormous ad spend. Tesla made a car so much better that drivers told other people. OpenEvidence copied that loop. A year ago usage sat between zero and a thousand doctors; today it is 10 to 25 percent of active U.S. physicians, by Tesla logic rather than a hospital RFP.

What makes the app awesome, Nadler argues, is that it solves a pain people who are not doctors rarely see. Physicians are stretched too thin, but they are also expected to keep up with a fire hose of medical information. Two new medical papers are published every minute, 24 hours a day, seven days a week. A *Nature* study put the doubling time of medical knowledge at every 73 days; Nadler thinks that methodology, which counts all citations, is aggressive.

OpenEvidence ran its own conservative estimate, counting only what doctors really need to know, such as the top third of journals by impact factor. Even then, medical knowledge doubles every five years. In 1950 it doubled every 50 years, so the half-life of medical school roughly covered a career. Anatomy and cell biology still move slowly, but new treatments do not. Today, depending on specialty, half of what a doctor learned about new therapies can be out of date by the end of residency and fellowship. Medical school as an information channel now sustains a physician for only a couple of years.

He walks through a real pickle: a psoriasis patient who also has multiple sclerosis seeing a dermatologist. New biologics, IL-17 and IL-23 inhibitors, look promising for psoriasis in general, but IL-23 is much safer if the patient has MS. The dermatologist may keep up with dermatology journals; expecting that same doctor to read every neurology page, let alone the border region between specialties, is ludicrous. Google and PubMed return article titles. They do not answer a question this specific, and the comparison lives on page five, six, or seven of a 30-page paper.

Engineers, he says, will recognize the job immediately: everything is an edge case, everything is a corner case. The surface area of medicine is not truly infinite, but for a working doctor it might as well be. IL-23 inhibitors arrived in 2017, 2018, and 2019; even a physician Nadler's age could not have learned them in medical school. The feedback OpenEvidence hears over and over is that a doctor used it for a case they might see once or twice in a career, and then said the same thing about the next case, and the next. The product is search and discovery on a tail that, to a human brain, is infinitely long.

Grady asks Nadler to unpack the company name. Evidence, Nadler says, is peer-reviewed medical literature, and it matters most to say what it is not. Large tech companies produced egg-on-face medical demos because they retrieved across the public internet, including health blogs and, for a time, Twitter. Many health-blog writers are well-meaning people whose core skill is writing. They may also file travel pieces or recipes, and they do not know more about medicine than they know about planning a trip to Mexico. Early language models were not inventing craziness; they were regurgitating that training data.

OpenEvidence trained models with no connection to the public internet. What does go in is the *New England Journal of Medicine*, via a partnership Nadler believes is unique among AI companies. Other well-known labs showed up with enormous checks and were refused: the Massachusetts Medical Society is a nonprofit, and no commercial contract was going to buy a mercenary decision. OpenEvidence did not knock. Senior people on NEJM's editorial board were already power users and wanted their own content inside the tool they used. Had the company waited for hospital meeting number 17, those editors would never have fallen in love with it, and the full-text deal would not exist. Nadler says with certainty that no other AI on the market trains on NEJM full text.

Open, for Nadler, was a reminder not to build another enterprise SaaS company. Kensho had been that, and those businesses can be great; he did not want to repeat himself. Open meant going directly to doctors as prosumers, past the gatekeepers, at the daily pain of being overstretched and forced to drink from the information fire hose. The best-endowed U.S. hospital systems can buy every tool in a category and use almost none of them. If Elon Musk's efficiency shop audited those SaaS bills, Nadler jokes, it would have a field day. Meanwhile many physicians are small-business owners in practices of ten or fewer, with a secretary and no university-style endowment, and they cannot pay $10,000 a year for software.

A community oncologist in Albany, Georgia, wrote that OpenEvidence had become a lifeline for his cancer center. Wikipedia told Nadler the city is 75 percent African American, with a median household income of $43,000, and the doctor is likely the only oncologist in a 50-mile radius. A physician in Fairbanks, Alaska, wrote that the product was a game changer where subspecialists are scarce. The same free tool is used at Mayo Clinic, where OpenEvidence was partly incubated, at Cleveland Clinic, at Walter Reed, and inside the VA, which would otherwise take three or four years to procure anything. Nadler says they are probably still in year 1.5 of that process if they had not gone around it. Open means every doctor in the country.

Under the hood, OpenEvidence is not a wrapper on GPT-3 or GPT-4. Nadler assembled PhD-level scientists from the labs that were already at the frontier before ChatGPT. Co-founder Zachary Ziegler studied with Alexander Rush in Harvard's NLP lab; Evan Hernandez comes from Jacob Andreas's lab at MIT; Eric Lehman is from MIT as well, with Jonas Wolf and Micah Smith among the early senior hires. They needed that mix of academia and engineering because consumer models were already creating embarrassing moments in medicine, and the bar had not been cleared.

In 2022 the field was reading DeepMind's Chinchilla paper and scaling larger. OpenEvidence bet the other way: smaller, highly specialized models overtrained on in-domain data would outperform much larger models on those in-domain tasks. They were rigid. They would not write a poem, and they fell over the second you left medicine. Inside the domain, Nadler says, they were beautiful. The team published "Do We Still Need Clinical Language Models?", which won best paper at the leading machine learning and healthcare conference in 2023. Ilya Sutskever's JPEG-compression metaphor, from Jensen Huang's interview, is the frame Nadler likes: if you compress the public internet you get the internet, including those blogs. OpenEvidence compressed medicine, starting with public-domain FDA and CDC material that U.S. copyright law puts in the commons, then winning NEJM users once the product was good enough.

In most apps a hallucination is annoying; here it can be life-threatening. The JPEG they made is an ensemble of about half a dozen specialized retrieval and ranking models handing tasks to each other. You cannot reach OpenEvidence's accuracy, Nadler says, by training one large language model. The other half of the fight is transparency. In early 2023, before ChatGPT shipped similar features, every answer was grounded in references a physician could drill through to the source. That also sent tens of millions of visits to journal pages, including methodology-section traffic publishers would never have received, until medical societies started asking to be indexed so they could get the referrals too.

He pauses to tell other builders that hallucination is not always a bug. In Midjourney it is the feature; the judo move is to find a domain where the usual hesitation becomes the product. If he were back in finance, he would start there: not PE ratios, but black-swan risk, the 2008-level tail that a "wet brain" struggles to imagine. He has tried it on his own portfolio, feeding enough context that the model invents Nvidia scenarios he had never considered. In 2025 he thinks applications that treat riffing as an advantage are about 1 percent of the market, with 99 percent still up for grabs. None of that, he adds, is fair game in medicine.

Grady asks what a decade between Kensho and OpenEvidence changed in the stack. Kensho, founded in 2013, was pre-large-language-model, almost "pre-fire." What the two companies share is infrastructure. Even a team that only calls a frontier API will watch that dependency fall over if the product succeeds, and then the work is classical software engineering. Nadler wants listeners to hear that ChatGPT did not repeal the laws of physics. Technology has been getting better from 1982 to 1987 and from 1993 to 1997; the slope is steeper, Ray Kurzweil's nonlinearity is real, and it is still a continuum, like accelerating toward light speed without rewriting physics.

The other constant is the people. He cannot improve on Steve Jobs: A players want to work with A players. Plenty of candidates sign up for BUD/S, the Navy SEAL screening course, just to see whether they can keep up. That instinct, he says, is as old as Achilles, and it shows up in warfare, engineering, sports, and finance. For a stretch it was controversial to say out loud that he wanted very high IQ and did not care about background or appearance. He will not sugarcoat it. The first senior cluster, Ziegler, Wolf, Hernandez, Lehman, Smith, all came out of Harvard or MIT PhD programs, not because of a brand filter, but because Kensho had taught him that high-velocity learners beat a team a hundred times larger.

By high IQ he means something neurologically specific, not speed at a Rubik's Cube. It is François Chollet's definition: the ability to efficiently acquire new skills, to assimilate information that did not exist in your head yesterday. Grady agrees. Nadler reads Clausewitz, Machiavelli, and Sun Tzu for pleasure, and he thinks Napoleon and Alexander were not the largest men in their armies. They over-prepared, then watched none of the preparation match the battle, and rebuilt their entire framework in real time. That, in some branches of cognitive science, is high neuroplasticity, and it is what outlier success still requires in 2025.

Lightning round, question one: on what date will OpenEvidence have saved a million lives? Nadler starts to build a consulting-style tree. Grady offers his own math. Medical mistakes kill something like 300,000 to 800,000 people a year in the United States; take the low end, cut it in half to 150,000, and a fully ramped product gets to a million in about six and a half years, plus a couple of years of remaining growth. He lands on November 4, 2034. Nadler says he will use the question in interviews, then adds that the 2030s tally misses what is already happening: the MS that did not get worse, and a Rhode Island doctor who used OpenEvidence as a curbside consult, recognized a pulmonary embolism, rushed the patient back to the emergency room, and saved a life.

Are broad foundation models commoditizing? The frontier, Nadler says, does not get commoditized, because it is the frontier. The cost of the original ChatGPT wow factor will converge toward zero, as Kurzweil would predict. Intellectually, the foundation-model layer remains phenomenal. From a business perspective, the great companies will be at the application layer. On AGI, he thinks we have already arrived and keep moving the goalposts: high-school expertise in many fields, then college, then PhD-level work from coding to medicine. What people actually mean is consciousness, the movie scene where the system becomes aware. He is not sure that ever happens. Consciousness may not be an emergent property of a dense enough neural network. That is a philosophical question.

The one piece of content he wants AI builders to consume is Ted Chiang's early-1990s novella *Understand*, from the author of the story that became *Arrival*. Without spoilers, it is his touchstone for what nonlinear acceleration of intelligence feels like; if they insist on nonfiction, go read the Chinchilla paper. The most optimistic near-term picture is personalized medicine, which has sat on the horizon like quantum computing and fusion. Matching psoriasis-plus-MS to the safer biologic is already a sketch of that future. In ten years, whether through OpenEvidence or a constellation of similar systems, a patient's exact fact pattern gets matched to everything known in medicine, and the care plan is hyper-tailored. Push that far enough, he says, and Theseus's ship starts replacing planks of human biology, until 120 or 130 is no longer the ceiling and a parent might watch a child turn 100.