Clinical Guidelines Are 11% Evidence. Curai’s Founder Still Wants GPT-4 in the Exam Room

Neal Khosla, founder and CEO of Curai Health, opens with a claim that sounds almost rude to modern medicine: for as long as people have practiced it, the job has been a short conversation with a doctor. Wealth buys more of that time — medieval kings got more attention, today’s concierge patients still buy the same scarce resource — but the lever has not changed. Hosts Nathan Labenz and Erik Torenberg had asked for a patient’s-eye view of 2025 and 2030. Khosla started further back, with the practice itself.

The computing revolution, he argues, rewrote almost every profession except this one. Drop a physician from 2023 into the same job 60 years earlier and they would function just fine. Imaging machines arrived in the 1970s; some therapeutics changed; the core work did not. Doctors remain oracles who are supposed to hold the latest biomedical knowledge, spend 15 minutes with you, and issue a recommendation. Reimagining that job around data and computing, Khosla says, is the actual starting point — not a nicer portal on top of the same visit.

His version of zero-marginal-cost expertise is specific. If there is one physician in the world who is the expert on your condition, that person should be available around the clock, even if the patient never speaks to them directly. Everyone should be able to pull out a phone and get best-in-class, personalized knowledge at roughly zero extra cost. That matters in the United States. It matters more for the eight billion people who will never walk into the Mayo Clinic. Curai had already raised more than $50 million before the current AI boom; Khosla had been an ML researcher at Google and Stanford. The thesis predates GPT-4. The model, he thinks, finally makes the supply curve movable.

Medicine, Khosla says, is not a particularly data-driven science. It is a judgment-based art. A review from roughly five or six years earlier asked how many clinical guidelines — the establishment’s official “how to treat this” — were based on grade-A clinical evidence. The answer was about 11%. The rest is, in practice, expert opinion from panels that sit down and talk it through. A further look found that doctors follow those guidelines only about half the time. Multiply the two and, as a patient, you are not getting a data-driven recommendation most of the time. “A couple of smart people sat down and talked it out” is still the operating system in 2023, he says, even though millions of people have already lived through the same conditions.

Building the infrastructure to collect longitudinal data — what happened to the patient, what was done, what resulted — and then surface it as decision support at the point of care is a 20-year project, not a three-year product. Utah is one of the few places with a glimpse of that world, because a relatively homogeneous Mormon population stayed put for generations and adopted electronic health records and genealogy early. Even that, Khosla argues, is probably insufficient. The rest of the planet does not have the glimpse. Modern medicine is a miracle, he is careful to add. It is not yet the data-driven practice he would want as a patient.

Patients today reach medical expertise in two ways: they talk to a doctor, or they research themselves on Google, Reddit, and forums. Those two paths, he thinks, need to become one. Self-research is unvetted and self-directed; the 15-minute visit never hears about the fish oil Erik takes every morning. The interface Curai wants is a text box. You say what you are wondering about — a supplement, a diet, diabetes, trying to get pregnant — and the system’s first move is not an answer. It is more questions: why this, what else have you tried, are your triglycerides actually high? Search engines never do that. Dialogue is the product.

Even “simple” patients blow past a context window after a few turns, so prior notes have to be indexed and retrieved into the window in real time. Language models also cannot, under current U.S. rules, tell a patient to take a named drug; that has to come from a doctor. Mental-health guardrails sit outside the model after a news story about a suicide that a partner linked to chatbot use. Before a final recommendation, Curai always connects the patient to a physician who can prescribe, order labs, and close the loop. The goal is not to replace the visit. It is to let tools do perhaps 80% of it so doctors can actually respond to the full range of what patients are dealing with.

Because quality and safety matter so much, Curai is building something Khosla has not seen much of elsewhere: unit tests and regression tests for clinical cases. A hypothetical patient with a bacterial infection needs an antibiotic; the model interviews, the doctor reviews a summary, they prescribe together. The company wants a large suite of such cases so it can see how the model handles a common range of scenarios as it changes. Language models are non-deterministic, and fine-tuning or human feedback can move them. You would not want antibiotics to work today and go haywire tomorrow. Labenz notes that OpenAI’s evals library is a glimmer of the same instinct; he guesses Curai is going “comfortably 10x deeper” on the cases that matter.

Labenz’s own early GPT-4 red-teaming, recreating routine visits from his life, felt almost indistinguishable from the real conversations. Khosla’s anecdote is blunter. He would not recommend it to anyone who does not understand the risks, but most of his own medical advice now starts in ChatGPT with GPT-4 — including mid-workout, when he hears something about a supplement. A physician on his team who moonlights at Stanford insists on bringing the tools into clinic and says they have made him a better doctor. Quantitatively, a January paper suggested GPT-3.5 could pass the USMLE; after GPT-4 launched, Microsoft formally reported scores above the average physician. Curai had just published a paper that, Khosla says, pushed past that state of the art. By any objective measure on current benchmarks, these systems look as capable as a physician. He still wants more rigorous proof before “egregious claims at scale.”

The exam questions themselves are a poor picture of practice. He reads one aloud: a 67-year-old man with transitional cell carcinoma of the bladder, tinnitus after neoadjuvant chemotherapy, five options ranging from inhibition of thymidine synthesis to DNA cross-linking. No doctor is handed five answers. Curai stripped the choices and made the model generate an open-ended response, then used another model to judge whether “high cholesterol” matches “cholesterol over 200.” They call the metrics GPT recall and GPT precision. Multiple choice both inflates scores, by shrinking the problem, and can deflate them if an old few-shot script forces a letter too soon. Real patients are messy. Early medical AI built huge graphs from findings to diseases; a lactose-intolerant patient may only say they feel bad after cereal, never “I drank milk.” Language models can represent those findings in abstract terms. Diagnosis also evolves — a differential on arrival, another after more questions, another a week later — and static vignettes do not.

Curai’s paper, *Dialogue-Enabled Resolving Agents*, uses two instantiations of the same model. One is the “decider,” which takes the case and offers conclusions. The other is the “researcher,” which pokes holes in every step. The pair enters a Socratic dialogue; the original working title was “student and teacher.” Across a variety of medical tasks, Khosla says, the setup reached state-of-the-art performance. He would like someone to try it outside medicine. The recurring lesson of the moment is that a clever prompt to one GPT-4 is good, and several agents interacting is better. Biomedical knowledge is already in the latent store. The question is how to elicit it. Skeptical questioning, in their results, pushes the system to a better resolution.

Classical ensembling says two copies of the same model should share failure modes, so they should not help. Khosla’s reply is that the state space is so large that you are really trying to surface the right knowledge, not average independent errors. He points to new work on chain-of-thought arguing that reasoning emerges from local clusters of variables influencing one another — “locality of experience,” as he recalls it. Mixing GPT-4 with GPT-3 and Bard, against his old machine-learning intuition, did worse than several GPT-4 instances working together. Labenz analogizes to Deep Blue pouring compute into runtime search, versus today’s models spending compute in training and then emitting one token at a time. Meta’s Cicero, which played Diplomacy, was another attempt to bring compute back to inference. Multi-agent dialogue is a medical version of that bet.

Multimodal systems are still mostly a suspicion. Khosla had watched a model ask a patient to upload a photo that morning. A rash that is red in the image and “itchy and flaky” in the patient’s words should be more informative together than either channel alone. What already surprised the team is generalization to data the model almost certainly never saw as medical time series. GPT-4, out of the box, interpreted continuous glucose monitor traces: at 1 p.m. the patient’s glucose spiked to 170, a sign of a carbohydrate-heavy lunch. Companies that treat glucose data as a moat should be nervous, he says. As more modalities go in, out-of-the-box models will do more of what used to look like secret sauce.

Khosla’s “inconvenient truth” for the industry is that OpenAI is “way better than everybody else and it’s not particularly close.” He has tried Google’s models and others; he is willing to name Google and not the rest. If the problem is pure intellectual horsepower, using anything else is a sacrifice. Content tools like Jasper, he allows, may already be good enough that cost and open-source models win. Curai’s problem is not that. Performance in medicine is critical, so they need GPT-4 plus scaffolding: safety, unit tests, prompt engineering, guardrails, longer memory, and agentic reasoning of the kind in their paper. Training a language model for a specific use case, depending on the case, ranges from “correct but insignificant” to “completely delusional.”

The CGM example is also his attack on data moats. These models are extraordinary few-shot learners. Give them 10, 100, or at worst 1,000 examples of a cognitive task and they generalize well. Owning a million data points of thing X does not stop someone from labeling a thousand examples and capturing 90 or 95% of your performance. Only if you truly need 99% does the pile become a substantial advantage, and for many products it will not. Competition then returns to UX and distribution. Domain-specific models are overblown for the same family of reasons, unless you are optimizing cost or latency. GPT-4 will probably beat your medical-only model.

A patient came back from Burning Man with hurting lungs. A system trained only on medical records and biomedical text would not know what Burning Man is, would not know the Las Vegas desert puts dust in lungs, and might miss a rare desert fungal infection. Lactose intolerance hidden in “I eat cereal” is the same class of clue. “Life is big, the world is big, the world is messy,” Khosla says. Labenz agrees that scoring well on the USMLE with a narrow model is imaginable; interacting with a real patient, where hints are sprinkled into conversation, seems to need something closer to a world model. On the product side, Curai is also exploring “goal-oriented medicine”: if the patient wants to lose weight or control diabetes, can a model, AutoGPT-style, stay in the relationship over time and work the goal instead of answering one snapshot?

Curai already has a direct-to-consumer app: match with a doctor, ongoing access, $14.99 a month at curaihealth.com. The business push is enterprise — health plans and provider systems. “The cat’s out of the bag,” Khosla says. Over the next three years, most patients will start with ChatGPT for medical advice, the way they started with Google, only more so, and more dangerously, because some will never come in at all. Insurers also care that a chatbot sending someone to neurology is expensive if the referral was not needed. The pitch to health systems is to get in front of that trend with a consumer-centric, AI-centric front door that still has humans in the loop and can actually complete the job: medication, lab tests, follow-up, not just information.

He uses Stanford Health Care as a *fake* example — they are not a partner — to describe the flow. You download the system’s app, tap “get care now,” and you get Curai’s stack first, not a raw GPT-4 chat. It triages, including emergencies, then a clinician jumps in for the last 20%, especially the active decision of what to do. Backend tools draft notes, automate post-visit check-ins, and assemble care plans. At recording time, early April 2023, only about 5% of patients were on the new models; the rest were being aged off older ones in a slow rollout. He expected 95% or more within three months. GPT out of the box is not HIPAA-compliant, so the team has to be precise about where they can use it, where other models apply, and where they build their own. Microsoft and others will eventually solve that, he thinks. For now it is an obstacle.

ChatGPT changed the sales conversation. CEOs of health systems and large plans now ask how this will change their business; a year earlier, Khosla says, 95% of the establishment insisted the models could not do the work. Thoughtful critics still raise hallucinations and the need for explainability when deploying to doctors — combining GPT-4 with other models for interpretability is a live question — but the “these models can’t” line has collapsed in three to six months. On regulation, his position is sharp. With doctor oversight, treat the system as clinical decision support; the physician stays responsible. If you want autonomous practice, treat it like a medical device. IDx-DR, a diabetic retinopathy screening system, received full-scale approval to read retinal images and drive a care plan, which he offers as a case study: show the FDA evidence on defined clinical cases, then earn autonomy, with a high bar. Releasing current models to consumers to “practice medicine” would be a mistake.

Malpractice today is judged against the standard of care: what would the average doctor have done. These models, he notes, are extremely good at the average thing, and “right now they probably already meet that bar.” Unlike autonomous driving, medicine has no “perfect driving” to compare against, only ambiguity about the body. He thinks the same human standard should apply, and that if models can replicate a median physician they can invert supply and demand. Pure AI-first apps that leave the patient to decide are not, in his view, the competitive threat. ChatGPT can say rest and ice. It cannot start a patient on rosuvastatin. Prescriptions, medication changes, lab orders, and result interpretation still require a doctor in the United States. Completing the job to be done, and staying with the patient as the problem evolves, is the advantage.

Survey data Khosla cites showed a direct correlation between optimism about AI and a country’s wealth, inverted: poorer countries were “abundantly optimistic.” India has more than a billion people and no realistic way to staff them with doctors. Practical obstacles are specific. Ask a large model what to prescribe and you get a Western, U.S.-indexed answer; India’s drug supply chain is much narrower, so the answer may be irrelevant. Health literacy and the ability to describe symptoms are further gaps. The regulatory environment in those countries will likely be more favorable, he thinks. Curai still operates mainly in the U.S., under tight physician supervision. Building here, he says, is the difference between brewing a Starbucks coffee and building the entire Starbucks: more complex, higher value, and the path to collecting your own data, human feedback, and fine-tunes.

China, only half-joking, might put a large language model in everyone’s hand within six months and leave doctors behind — “some serious damage” in the short term, strong long-term results if they do. The U.K.’s NHS is a different bottleneck: huge bureaucracy, a system he describes as crumbling, theoretically in desperate need of language models, hard to move. The U.S. advantage is many doors; you can start small, gather evidence, then face Medicare, Medicaid, and the giant payers. Britain has interesting mental-health experiments — Limbic does LLM triage in a shortage specialty — but core medicine turns into “territory fighting.” Labenz asks whether there is an obligation to deploy imperfect systems where the alternative is nearly nothing. Khosla’s answer for parts of the developing world is yes. Raj Panjabi’s Last Mile Health trains high-school graduates in Africa on the ten most common local conditions, including malaria, to expand primary care. Bloomberg has been associated with training high-school graduates to do C-sections. Where access is that poor, these systems are “an absolute must,” if they can be adapted to local biomedical norms. There is, he says, “a moral imperative to get these things to a lot of people ASAP.”

Asked whether he would take a Neuralink-style implant for thought-to-text, Khosla first wants the gadget that records ideas while falling asleep. The serious answer is that he does not know. In 10, 15, or 20 years there will be so much cognitive infrastructure that the question becomes what spiritual fulfillment means — VR, hiking, an implant — and entrepreneurs rarely have time to decide. His larger hope is that society can shift from a scarcity mindset to an abundance one as wealth and, he thinks, prosperity need redistributing; classical economics may not fully hold. The failure modes are a few people controlling the world, or the West trying to stop the technology while China and Russia build it. Bill Gates’s line about a human whose purpose is to be a hamburger chef sticks with him: work displacement is a real distribution problem, but the jobs themselves are not the meaning. The exciting 25-year world is free energy, free intelligence, people playing music and colonizing more of the universe. Getting there, he argues, requires working with existing governments rather than building in isolation — which is why he does not want AI to “run wild in medicine,” and why he still thinks collaboration could invert the supply of care.