Visits with your doctor are episodic and months apart because human attention is scarce. Guidelines are coarse because the evidence you need to add a branch in a guideline is expensive to collect. Disease categories are broad because many distinctions are too difficult or expensive to measure. Clinical trials are rare because recruitment, experimentation, and analysis cost millions.

But those costs are changing. AI is making many forms of work that were previously manual (and therefore scarce) cheap and abundant.

As those changes ripple through a healthcare system built around fundamentally different constraints, the second-order effects may be stranger than the first-order gains.

1) Total healthcare spending will rise dramatically.

Healthcare will become dramatically cheaper per unit while consuming a larger share of the economy.

We already spend an extraordinary share of GDP on healthcare. In 2024, US health spending was at 18% of our GDP. CMS is projecting an increase to 20.6% by 2034, and I think even that will prove conservative.

William Stanley Jevons died in 1882 (drowned while bathing :/), but his paradox that efficiency gains can increase total consumption becomes more popular by the day.

For the full Jevons Paradox to hold:

• a technological change increases efficiency

• consumer prices drop because of those efficiency gains

• an elastic demand curve leads to increased consumption that exceeds the price drops (e.g. consumers buy 3x more at 2x lower prices)

The efficiency gains are the easy part. And while healthcare is far from perfectively competitive, it’s hard to imagine a world where all efficiency gains flow to margins instead of consumers.

For the third point, for spending on a fixed service to actually rise as its price falls, demand needs to be elastic in the technical sense: an elasticity below -1.

There’s a fantastic paper that measures demand elasticity in healthcare by looking at how spending varies across types of services with different deductibles. We can expect the total consumption of the most elastic services to increase considerably (e.g. imaging and specialist visits).

But even at an elasticity of -0.44, a 10x decrease in price would produce around 3x increase in utilization. This is a huge increase, but spending on this curve still would fall substantially.

But that calculation also assumes healthcare remains the same product; that we’ll just move to a different point on the same demand curve.

Falling costs will expand the set of things that medicine can do, creating new sources of demand. Existing patients will consume more, but new patients will also enter the system, new indications will become economical to treat, standards of care will expand, and entire categories of healthcare that exist at the margins today will become normal.

In 2001, it cost $95 million dollars to sequence a single human’s genome. Today you can pay a few hundred dollars to have yours sequenced. Sequencing didn’t just become a cheaper version of the same product. Falling costs have made it economically usable and widespread tool for oncology, rare diseases, reproductive medicine.

2) There will be more diseases and more people will have one

What we label a disease is partly determined by what we can detect and what we can act on.

As measurement and analysis become cheaper, we’ll uncover and label diseases at earlier stages, when they’re milder, represent lower elevations of risk, and are less symptomatic.

Bone fragility existed as a concept in medicine before we could easily test it. But DXA scans gave medicine a way to measure bone density as a continuous variable in asymptomatic people. The WHO defines osteoporosis as bone mineral density 2.5 standard deviations below a reference value of a young adult.

Bone density and fracture risk are continuous. There isn't a natural cliff at which osteoporosis just appears. Medicine draws a clinically useful threshold across that continuous distribution. As measurement improves and intervention gets cheaper, we’ll draw many more of those boundaries.

Broad disease categories will also start breaking apart into lots of distinct diseases. There are huge classes of idiopathic diseases that that may really be collections of different mechanisms producing similar symptoms, but these are mechanisms we currently can’t afford, or simply don’t know how to, distinguish.

Broadly this is a good thing. Narrower disease subtypes make narrower interventions more viable. Catching diseases earlier means we can intervene earlier and avoid more acute presentations. More detection will keep people healthier, but will also create more anxiety around incidental findings. Culturally and psychologically, we’ll have to come to terms with knowing a lot more about our bodies, including abnormalities and risks that may never cause us harm.

3) Doctors will experiment on you more.

Treatment and research have stayed mostly separate, blending only in exceptionally high cost and acuity settings like oncology where trial enrollment is a more frequent part of treatment.

But as recruitment, monitoring, adherence support, and data analysis all become cheap and care becomes more coordinated, we will start collecting clinical evidence in many more care settings.

We use clinical trials to try to convince ourselves that treatments are safe and effective on average. But most pre-market trials don’t collect enough evidence to tell us exactly who responds, at what dose, and under what conditions.

Today, so much uncertainty gets resolved at the physician level. The specific drug to start you on is based on that provider’s preferences, experiences with other patients, and trial and error on a per-patient basis.

Part of the value of an experienced physician is that they’ve effectively run thousands of informal experiments over the course of a career. But the issue is that most of what they learn stays trapped in that physician’s head rather than bubbling back up into the shared evidence base that informs other physician’s treatment decisions.

Even ordinary treatment choices for low severity conditions could be randomized when several options are reasonable. More trial-level evidence will be easier to collect even when not incentivized for a new drug. This will make post-market evidence collection around comparing generic options, dosing, and subgroup analysis much more feasible.

During covid, almost 10% of all hospitalized patients in the UK were enrolled in a comparison of dexamethasone, an old, inexpensive drug, in patients requiring respiratory support. They were able to rapidly establish a mortality benefit for an old, inexpensive drug, that was able to rapidly increase the quality of care for a new disease.

Instead of having a parallel universe of clinical trials, experimentation will be a layer on top of care delivery.

Subscribe now

4) Existing drugs will become more valuable

Most pharmaceutical evidence collection today is focused on novel therapeutics, where the cost of research is incentivized by the profits of being the IP owner of a new, useful molecule. New drugs have owners who can monetize new evidence while older drugs largely don't. Medicine has much weaker incentives to optimize the enormous pharmacopeia of off-patent drugs.

Basic questions about dosing, combinations, heterogeneous response, sequencing, and alternative indications remain surprisingly open. If evidence collection becomes much cheaper, this backlog of questions becomes easier to answer.

The ADAPTABLE trial randomized 15,000 patients across the US to 81 mg or 325 mg of aspirin. It’s an almost comically boring question about a drug that has been around forever. There’s weak incentives for anyone to run this type of trial. But for drugs that millions of people take every day, its exactly the kind of question we should answer more often. The trial found that 325 mg didn’t improve cardiovascular outcomes over 81 mg, suggesting that millions of patients didn’t need the higher dose.

Similarly, ALLHAT was an enormous NIH trial that randomized tens of thousands of patients to compare an old, inexpensive diuretic with newer blood-pressure drugs. The newer drugs generally weren’t better, helping establish cheap thiazide-type diuretics as a first-line option.

Those sorts of trials would rarely be conducted by pharmaceutical companies incentivized to discover new drugs, but can become much more common-place as we make it easier and cheaper to collect trial-level data.

5) Fee-for-service will get squeezed to death by cash-pay and value-based care.

Paying for medical output rather than medical outcomes has always created perverse incentives. This has always been the case, but we’ve clung to FFS in the US because:

• medical output has remained at least directionally correlated with medical value

• VBC has been expensive and operationally cumbersome

• most medical care has been too expensive or unpredictable for patients to purchase directly

At its core, FFS depends on medical output being a reasonable proxy for medical value. AI will sever that relationship for large parts of medicine.

As the marginal cost approaches zero, providers will be able to produce mountains of billable output. You will be able to conduct visits, message patients, conduct reviews, order tests, and document cases for almost free, billing for everything. This weakens the already imperfect correlation between output and outcomes.

Payers will feel this first, and initially be forced to fight an AI war between payer denial agents and provider’s RCM agents.

Just as spam proliferated when email became virtually free, cheaper bureaucracy may create more bureaucracy, not less. Payers will respond by tightening utilization management, changing reimbursement, and repricing coverage.

For cheap, shoppable services that sit below the deductible, cash pay will increasingly route around insurance.

And for expensive, longitudinal, risky care, VBC will start to look even better. One hospital study of Medicare value-based reporting programs estimated 108,000 personnel hours and $5 million annually devoted to quality reporting alone.

So at the same time, the economics of VBC improve. The expensive work required to bear risk (monitoring, reporting, coordination, outreach, and early intervention) also gets dramatically cheaper.

Scarcity has shaped what we measure, what we call a disease, what we study, how often we see patients, and how we pay for care. As that scarcity recedes, we shouldn’t assume those choices remain correct.

We have an opportunity not just to automate a health system built for expensive intelligence, but to design the health system we would have built if intelligence had always been cheap.

This week Midjourney announced a machine that scans your entire body in about a minute. You step into a shallow pool and a platform lowers you through a ring of submerged ultrasound sensors, and a minute later there is a map of everything inside you, with no radiation and no needle involved. The pitch is early detection: finding a cancer while it is still small enough to cut out.

Midjourney is best known for making AI images from text prompts. The scanner is its first piece of hardware, and the company says it eventually wants 50,000 of them doing a billion scans a month.

Whole body scans aren’t new. Prenuvo has been selling whole body MRIs to healthy people for five years now, and a handful of competitors have followed. What kept them rare was money. A whole body MRI runs $600 to $2,000 out of pocket and insurance won’t cover it for someone with no symptoms, so the scans went mostly to people who could afford an expensive maybe. Midjourney is promising the same thing at roughly the price of a massage, and whether their particular machine works or not, cheap whole body imaging is clearly coming.

Every single person has something weird in their body. A whole body MRI of a healthy person turns up something abnormal in 80 to 97 percent of people, and almost all of it is harmless. One review of more than 6,000 screening scans found 91 percent of the findings were benign and about one percent were cancer. These machines reliably find things, and what they find is overwhelmingly noise.

Many doctors object for a statistical reason: in someone with no symptoms, cancer is rare, so most flagged findings are false alarms. Each one can lead to another scan, a specialist visit, sometimes a biopsy, and weeks or months of waiting.

What Midjourney built

The Midjourney scanner is an ultrasound device. It uses Butterfly Network’s ultrasound-on-chip, the same component in their handheld probes, arranged into a ring you pass through underwater so it images you from every angle. Midjourney licensed the technology in late 2025 . The company markets the result as “Ultrasonic CT,” which is a bit confusing, since CT means X-ray computed tomography and there are no X-rays involved, and it compares the images to MRI, which uses magnetic fields rather than sound.

By the company’s own account it’s not a diagnostic device yet. It has no FDA clearance, it currently outputs only body-composition maps of muscle and fat, and the founder, David Holz, said it doesn’t yet use any AI. The diagnostic version that would rival an MRI is a regulatory process the company says it intends to start later. The first location is a spa in San Francisco penciled in for the end of 2027, and the stated goal is 50,000 machines doing a billion scans a month, paired with a claim that enough early imaging could prevent 30 percent of deaths and half of all healthcare spending. There is no evidence I could find for that last figure.

Maybe Midjourney is the company that drives whole body imaging costs down an order of magnitude, maybe it’s a competitor, but cheap and frequent scanning is coming quickly either way.

What whole body imaging finds

Prenuvo’s argument is that the cancers that a whole body scan turns up are disproportionately the ones nothing else would have caught in time, the cancers with no routine screening test. Their Polaris study followed about a thousand mostly asymptomatic adults and found biopsy-confirmed cancer in 2.2 percent of them, most in people who felt fine, a lot of them kidney, bladder, and ovarian cancers that tend to surface late. Among the people whose scans came back clean, 99.8 percent were still cancer-free a year later.

Those numbers come from the company that sells the scan and are drawn from people healthy and motivated enough to spend a few thousand dollars checking on themselves, so they are worth treating with some caution. There is no randomized trial showing that whole body MRI makes average-risk people live longer, and no mortality data at all. Prenuvo has a 100,000-person study running to look for some, and the results are years out.

False positives and overdiagnosis

A scan finding can be bad in two different ways:

The first is an ordinary false positive. The scan flags a spot, the spot looks like it could be something, you do more tests, and it turns out to be nothing. The information was wrong, and the cost was the workup it took to find that out and the anxiety in the meantime.

The second is overdiagnosis, which is stranger. The scan finds a real cancer and the biopsy confirms it, so the test was not a false positive. But sometimes that cancer was never going to hurt you. It would have sat there doing nothing for the rest of your life while you died of something else decades later. Because the scan found it, it gets treated, and you absorb the surgery, the radiation, and the years of follow-up scans.

Thyroid cancer is the textbook example. As neck imaging became common, thyroid cancer diagnoses climbed steeply across dozens of countries while the death rate stayed flat. A 2024 analysis put the share of thyroid cancers diagnosed worldwide that were overdiagnoses at roughly three in four.

The hard thing about overdiagnosis is that accuracy does not solve it. The test wasn’t inaccurate and the cancer was really there. The issue comes from the health system you’re in treating every cancer it finds as urgent, including the fraction that would have done nothing if left alone.

The exception

There is one group of people for whom whole body MRI is not controversial at all. People born with a broken copy of the TP53 gene have Li-Fraumeni syndrome, which gives them an extraordinarily high lifetime risk of cancer, often several cancers, often young. For them an annual whole body MRI is recommended by the NCCN as standard care. A baseline scan finds a new cancer in about 7 percent of TP53 carriers, almost always early enough to treat, and surveillance meaningfully improves how long they live.

It produces the same flood of incidental findings, and it false-alarms about 40 percent of the time even in this group, and it is still clearly worth doing for them. The difference is the prior: in a TP53 carrier, a suspicious finding is much more likely to be real cancer.

Priors

This tradeoff is older than whole body imaging. It shows up in every screening test, and it turns on the prior: the chance the person had the disease before you ran the test at all.

Take a test that performs relatively well, one that catches a real cancer 90 percent of the time and falsely flags a healthy person only 9 percent of the time. Run it on 10,000 people in a population where one in a hundred actually has the cancer. A hundred people have it and the test catches 90 of them. Among the 9,900 who are healthy, a 9 percent false-alarm rate produces about 890 positives. So out of roughly 980 people who get a positive result, only 90 have cancer. The test worked exactly as designed, and a positive result still means the person is probably fine, because there were so few real cancers in the population to begin with.

Being one of those 890 sucks. It usually starts with a call saying the scan found something, followed by weeks or months of waiting through a follow-up scan, a specialist visit, and often a biopsy, with the possibility of cancer hanging over all of it. In a national survey of physicians, more than 99 percent said they had seen an incidental finding set off this kind of cascade, and most reported it had harmed patients financially, psychologically, or physically. Each workup also consumes a specialist slot and a pathologist’s time that someone with an actual problem then waits longer for. The harm can be physical, not only financial and emotional. A biopsy can bleed or get infected, and a suspicious spot on a kidney frequently goes to surgery, where 15 to 30 percent of the kidneys removed for a presumed tumor turn out to be benign. A number of those 890 healthy people end up losing part of an organ over something that posed no real threat.

Subscribe now

This is how testing for anything rare behaves, and in people with no symptoms the thing you are testing for is usually rare. The clearest demonstration is the mammogram. When researchers give doctors the standard numbers and ask what a positive result means, most of them get it badly wrong and guess far too high. Give them the identical information as counts of actual people instead of percentages, and most of them get it right.

A symptom changes the prior. A scan ordered because something already hurts is checking a real suspicion, so a positive there carries weight. The same scan run on someone who feels fine is searching a population that is almost entirely healthy, where most of what turns up is the ordinary background of an aging body. That is why a doctor images you the moment you have a worrying symptom and resists when you just want to look, and why whole body MRI makes sense for a TP53 carrier and not for most people.

A low prior doesn’t automatically rule a test out. Every newborn in the country is screened for PKU, a metabolic disorder almost none of them have, because the confirmatory test is cheap, the treatment is a change in diet, and the thing it prevents is permanent brain damage. The calculation is always the false-positive rate times the cost of chasing a false alarm, set against how much earlier detection actually helps. For newborn PKU screening it is an easy call. For a whole body scan in a healthy adult, given where the accuracy and the follow-up systems are, it mostly isn’t, at least not yet.

The handoff

The handoff is everything that happens between the scan and the patient: the grading, the wording of the report, the decision about what comes next. The same probably-benign spot can move through handoff and end up as a note to recheck in a year or as a biopsy referral, depending on how it's graded and described.

Radiologists already use structured reporting systems that sort findings into tiers, so that something needing real attention and something worth a glance next year aren’t delivered with the same weight. A report can give the patient the actual probability that a finding is cancer instead of the word “abnormal,” though that is rarely done. And a finding can be watched over time rather than acted on right away, which matters because a spot that has been stable across three years of scans is a very different thing from a spot seen once, and the second and third scans are what let you tell them apart.

Over-reporting in wellness scans like Prenuvo is partly an economic artifact. When a scan costs a couple thousand dollars and you get one every few years, the company has every reason to report a full inventory of what it saw, because a long itemized list is how a single expensive scan justifies itself. Prenuvo lists the cyst on your kidney and the spot on your liver because that report is what you paid for. That incentive depends on scans being rare and expensive. Once they are cheap and frequent, the reason to surface every little thing weakens, and the natural thing to report is what changed since last time.

Guidelines for what to do with test results lag behind access to the test. The screening arrives first, and the rules for handling what it finds catch up later, usually after the harms pile up enough to force the issue. That catch-up is driven by the load those cases put on the health system: enough bad outcomes accumulate that the standards get rewritten. We have been through this cycle before, most clearly with PSA.

PSA detects a lot of cancer, much of it the slow kind that would never have killed the man who had it. The enthusiasm in the early 2000s for testing every man over fifty led to a wave of biopsies and surgeries, and to many men left incontinent or impotent after treatment for cancers that posed little real risk. In 2012 the US Preventive Services Task Force recommended against PSA screening outright.

But then in 2018 the Task Force moved back toward screening for men aged 55 to 69, as a shared decision rather than a blanket recommendation either way. What changed the recommendation later was not a new test. It was active surveillance, the practice of monitoring low-risk prostate cancer rather than operating on it right away, which went from unusual to standard over the following years. As the typical consequence of a positive result stopped being an unnecessary surgery, the balance of benefit and harm shifted.

In the six years after the 2012 recommendation to stop screening, the share of men first diagnosed with metastatic cancer rose.

That whole correction depended on the cases landing inside the system, where the cost of getting it wrong eventually forced a fix. Cash-pay wellness scans sit outside that. They generate the findings and pass the cascade to a system they don’t answer to or pay for. The scans will get cheap before the rules catch up, and for a while the gap is going to land on people one kidney spot at a time.

What you call it

Say a man has the lowest grade of prostate cancer, the kind that essentially never spreads or kills, and the guidelines say to watch it rather than treat it. You can explain all of that to him, and he is still going to wake up most mornings knowing he has cancer.

The obvious question is whether you can just not tell him. But withholding a real diagnosis is a narrow and unfashionable idea in western medicine, and the direction of modern ethics is toward telling patients the truth and letting them decide what to do with it.

(I’ve always found it fascinating that in traditional Chinese medicine, patients are often not informed of a cancer diagnosis, although this practice has become less common recently.)

So the fight has moved somewhere stranger: whether the harmless thing should be called cancer in the first place. This has happened before. In 2016 a common low-risk thyroid tumor was reclassified and the word carcinoma was dropped from its name.

There is now a serious, unresolved push to do the same for the lowest grade of prostate cancer, renaming it something like “acinar neoplasm of low malignant potential,” so a person with a lesion that was never going to hurt them isn’t handed a cancer diagnosis and the years of fear that come with it. Pathologists are split, and one survey of them found about 82 percent against it. The thyroid version already went through.

Sharper signal

The scans will also get better, probably faster than the current debate assumes. False positives and overdiagnoses respond to better sensors, better software reading the images, and a growing record of what a given kind of spot turns into over years. The useful role for AI here is less about catching more and more about being confident enough to call a harmless finding harmless, so that fewer of them ever reach the patient as a scare. Repeat scanning does some of this on its own. Once there is a baseline, the question becomes whether a spot has changed rather than whether it looks dangerous, and that is both an easier question and a less frightening one to be asked. Collecting large longitudinal datasets now will let us build models that are better at categorizing risk.

There is a parallel in skin cancer screening. A dermatologist examining your whole skin for melanoma is doing roughly what a body scanner does, aimed at the one organ you can see, and it comes with the same overdiagnosis problem. US guidelines give routine skin checks for people without symptoms an insufficient-evidence rating, the same grade whole body MRI would get now. It is also one of the places AI has made the most progress at distinguishing a dangerous lesion from a harmless mole. There is little reason to expect the inside of the body to behave differently from the surface.

As the false-positive rate drops and the handling of each finding improves, the prior you need to justify scanning someone comes down with it. The line between the TP53 carrier who clearly should be scanned and the healthy adult who probably shouldn’t moves over time, toward more people reasonably being scanned and being hurt less often when they are.

More scanning is coming, and that’s good. Catching a cancer early in someone who feels fine is worth a lot, and these machines will do it for the kinds of cancer we usually find too late. After that, someone still has to tell the findings that matter from the ones that don't, and keep a harmless spot from becoming a diagnosis someone carries for years. But it’s a much better problem to have than not being able to see inside people at all.

“We literally sank to the floor while still on the phone with the geneticist,” his father, Kristofer DeVault, told me. “Walking into our little boy’s room felt like visiting a memorial. The life we thought we’d have with him felt like it had died on that call. It was an unspeakable kind of sad.”

About 400 people on the planet have been diagnosed with CTD. The actual number is almost certainly higher, since most cases go undiagnosed. The condition causes intellectual disability, autism-spectrum features, seizures, and developmental delays. Treatment, when there is a treatment, has to happen during the window of childhood brain plasticity. That window is closing on Brody now.

A drug that might help him exists. It is in the pipeline of a biotech that has indicated it would consider providing the drug pre-approval. There is a state law in Montana, signed last May, designed to allow this kind of access. None of it has reached Brody.

Twenty-one drugs

In 2018, congress passed the federal Right to Try (RTT) Act. The law was meant to give terminally ill patients access to drugs that had cleared Phase 1 trials but not yet been approved. In the seven years since, only twenty-one drugs have been supplied through it.

For comparison, the FDA’s older Expanded Access program, which the 2018 law was supposed to improve on, approves roughly 99% of requests at a four-day median turnaround. Between 2010 and 2020 the FDA received 16,380 expanded access requests and granted 16,258 of them.

The RTT bottleneck on access is not the FDA. Federal Right to Try requires the company that owns the drug to provide it at cost. But small biotechs have margins that cannot absorb the legal, regulatory, and adverse-event monitoring overhead that participation requires. One bad outcome can jeopardize the company’s reputation with the investors funding its multibillion-dollar approval pipeline. For the operators weighing the request for early access, the math almost never makes sense.

This is not a contested view on either side of the RTT aisle. Arthur Caplan, the NYU bioethicist who is among the law’s most prominent critics, has spent the years since 2018 pointing out that “the FDA is not the obstacle to compassionate use. The proponents of the Right-to-Try laws demonize the FDA, but it is almost never the problem. The main challenge is the drug sponsors and companies.” Stephen Martin, the advocate currently lobbying Congress to expand Right to Try, describes the same dynamic in different words.

Kristofer DeVault reached the same conclusion on his own.

“I already knew the federal Right to Try law was a nightmare to navigate,” Kristofer said. “It gives a company developing a drug almost no reason to participate.”

He spent the early months of 2026 working every angle, including investigating whether he could manufacture a treatment for Brody himself.

This was not rhetorical. By the time we spoke he had had formulation candidates made, located labs willing to manufacture, and collected quotes. He hasn’t gone ahead. He hasn’t ruled it out either.

Subscribe now

Montana

Senate Bill 535 was signed by Montana Governor Greg Gianforte on May 12, 2025. It establishes a state licensure framework for what it calls Experimental Treatment Centers. Centers hold a license issued by the Montana Department of Public Health and Human Services, pay a $10,000 application fee and a $5,000 annual renewal, follow informed consent and adverse event reporting standards, and only offer treatments that have completed Phase 1 trials.

The key change is that the law lets biotechs charge a profit on those treatments. Two percent of the net annual profits of an Experimental Treatment Center go into a state Insurance Premium Support Account that subsidizes ACA marketplace premiums for low- and middle-income Montanans.

Stephen Martin’s full-time job, on behalf of an investor called Infinita VC, is to call the CEOs of biotech companies and ask them to ship their experimental drugs into Montana.

He has been doing it for about a year. Stephen says about 70% say “yes, if.” The “if” is the FDA committing not to punish them for participating.

New Hampshire passed a similar bill, HB 701, in August. Eighteen states have passed a second variant of Right to Try that covers personalized treatments based on a patient’s specific genetic profile. The Goldwater Institute, which has been writing the model statutes for these laws for over a decade, calls them “Right to Try 2.0” laws. They are designed for n-of-1 therapies that were impossible to imagine when the original federal law was drafted. In February 2025, a team at Children’s Hospital of Philadelphia produced a custom base-editing CRISPR therapy for a baby named KJ Muldoon in about six months. The state laws are now ahead of the federal regulatory framework.

Brody’s case qualifies under Montana SB 535. The treatment Kristofer has identified would be a candidate. The company that makes it, he said, is “genuinely interested in participating, if the FDA risk could be taken off the table.”

Discretion

The FDA risk is that the agency retains authority over the company’s Investigational New Drug application, and could in principle withdraw the IND or otherwise penalize the company for participating in the Montana program. No one I spoke to thinks the FDA actually would do this. But “in principle, the FDA could” is enough to keep most biotech boards from approving participation.

What advocates are asking the FDA for is a piece of guidance that promises the agency will not punish companies that participate in state-licensed Right to Try programs meeting defined safety and reporting standards. “If I were on the regulatory council of a company weighing whether to participate in Montana’s Right to Try framework,” Kristofer said, “I’d want one thing in writing: a statement that the FDA will not punish companies for participating. That alone would mitigate the regulatory risk.”

This kind of guidance is called “enforcement discretion.” It is how the FDA regulates compounding pharmacies. It is how the FDA has regulated laboratory-developed tests since 1976. It is how the FDA tolerates the vast majority of off-label prescribing. The agency has used this tool hundreds of times. It has not used it for state Right to Try laws.

Kristofer made two trips to Washington this year with Stephen Martin to talk to senators and house members. He found most offices receptive.

“Most policymakers and their staff cared about the issue and wanted to help,” he said. “There was one office that opposed the idea, though. They kept pointing to the fact that federal Right to Try, expanded access, and compassionate use pathways already exist. They asked, ‘What if the big drug companies sue the FDA for allowing this to undercut their profits?’”

But the pathways they were pointing to are the ones that have provided twenty-one drugs in seven years. The lawsuit concern is not one legal analysts of state Right to Try laws have flagged. Drug companies don’t typically sue the FDA over enforcement discretion decisions.

Kris thinks the ask is more tractable than people assume. “What we’re asking is for the FDA to let states exercise their own expanded Right to Try laws without federal interference.”

Doubt

There are reasons to be cautious about Montana’s pilot. Critics like Caplan and NYU’s Alison Bateman-House have spent years pointing out that Phase 1 trials demonstrate safety, not efficacy. About 90% of drugs that enter Phase 1 ultimately fail in Phase 2 or 3. A “Phase 1 complete” drug is not a drug that has been shown to work.

Allowing biotechs to charge a profit on a Phase 1 drug creates an incentive structure that could in theory distort behavior in bad ways. Companies could choose to keep selling under Right to Try indefinitely rather than do the more expensive Phase 2 and 3 trials. Bad products could slip through. Desperate families could be exploited.

I think those concerns are valid. We do not actually know how Montana’s incentive structure will play out. But what we have now is not working, and families are suffering for lack of access.

Federal Right to Try was itself a policy experiment. We ran it for seven years. It provided twenty-one drugs. We learned that voluntary participation with at-cost pricing does not move biotech behavior. Montana is the next experiment, this time with profit on the incentive side. The only way to find out whether it works is to operate it.

This pattern of state experimentation in healthcare incentives has proven effective before. In 2019, Louisiana cut a Netflix-style subscription deal with Gilead for Hep C drugs. Washington did something similar. Several states have piloted outcomes-based contracts and installment payments for gene therapies like Zolgensma and Spinraza. CMS is currently running a Cell and Gene Therapy Access Model for sickle cell. In each case, the federal government let states test new payment structures and we learned from the results.

The drug approval system is built on the premise that we cannot predict efficacy or safety from theory, which is why we run Phase 1, 2, and 3 trials. Incentive design deserves the same humility.

Drug development costs have doubled every nine years for seven decades. Kris told me this is called “Eroom’s Law”. It’s the fucked up inverse of Moore’s Law. We have spent that time mostly trying to solve the underlying problem with technology. State-level experiments on incentive structures are one of the few levers that has worked, and one we have barely used.

Today

Brody is on a drug called Ravicti, originally approved for urea cycle disorders. Some researchers think it may help chaperone creatine across the blood-brain barrier and into cells; recent Cornell work suggests it also helps control seizures in CTD patients. Kris is not sure how much it is helping Brody. His insurance covers it. Without insurance, the list price would be approximately $180,000 a month.

“That price is driven by the staggering cost of drug development,” Kristofer told me, “and in this drug’s case, by cost inflation from acquisition.”

But the treatment that might actually help Brody, the one in the biotech’s pipeline, is not commercially available anywhere. Kristofer has been in direct contact with the medicinal chemists developing it and has read the data closely. He is one of a small number of CTD parents with the resources, network, and time to push for it personally.

“If I had to scale Everest every month to get my son his treatment, I would,” he said. “Enthusiastically.”

There are about 10,000 known rare diseases. Around 5% have any approved treatment at all. Most of the families navigating the other 9,500 don’t have someone who can spend two days a week in Washington.

A few things I’m pretty sure of after writing this:

• The existing pathways for pre-approval drug access don’t work well enough. The economic incentives are misaligned: companies have to provide drugs at cost, absorb the legal and reputational risk, and get almost nothing back. Twenty-one drugs in seven years is what those incentives produce.

• States like Montana are trying to fix that. SB 535 lets biotechs charge a profit, changing incentives for access.

• I don’t know if it will work. Nobody does. But a scoped state pilot is a promising environment to test it. The blast radius is small if it goes wrong, and the upside is access for families like Brody’s if it goes right.

The next FDA Commissioner can decide whether the experiment runs. The ask is one piece of discretionary guidance. If you have a line into the agency or administration, please consider discussing this with them.

If you’d like to get involved, the Senate will vote on the next FDA Commissioner soon. Your Senators’ DC offices take calls (numbers here). A version of what you could say:

“Hi, I’m a constituent. There are families like Kristofer DeVault’s whose three-year-old has a rare disease that could be treated under state Right to Try laws, but companies won’t participate without FDA enforcement discretion. Would the Senator consider making a public statement on this before voting on the next FDA Commissioner?”

This is the second part of my posts about sailing across the pacific ocean. If you haven’t read Part 1, start there.

Twelve hours into our crossing of the Pacific Ocean, a storm blew in and ripped our front sail to shreds.

The forecast had called for 15 knots of wind, and we ended up getting 30. We tried to wind in the headsail but it wouldn’t come in, jamming as the boat heeled over in six foot waves. I finally got up to the sail and wound it up by hand, tying it shut. But it slowly unfurled again, catching the wind. By morning the sail was destroyed beyond repair.

We were about fifty miles offshore, facing the prospect of crossing 3,000 miles of ocean without our front sail.

We talked about turning back. It was clearly the correct move. We were only a day’s sail back to land. I asked John, our experienced crew member, what he thought. His response was basically: didn’t you boys want an adventure? It matched our energy. We were too committed to turn around. So we kept pushing.

Preparations

We’d left San Jose del Cabo on April 10th to cross the first chunk of the Pacific to the Marquesas Islands in French Polynesia. Around 3,000 miles of open ocean. Most sailboats average somewhere between 100 and 200 miles per day, so we were looking at roughly three weeks.

We’d spent the weeks before making massive supply runs. We stashed enough peanut butter in every crevice of the boat that all four of us could survive on peanut butter alone for a month. Around a hundred eggs, twenty avocados, kilos of flour and rice, a bunch of meat. We heard liquor was worth double in the islands so we loaded up on bottles. Using the fridge and freezer at the same time drew too much power, so we just packed the fridge and hoped for the best. Every preparation I made or missed now could have real consequences for the next month, and I could feel that weight in everything we loaded onto the boat.

Water was top of mind. Fresh water is hard to get in the middle of the ocean. We had four tanks that held just enough for about four weeks, with almost no margin. We filled six jerry cans and lined the deck with them, plus thirty plastic jugs stashed everywhere else. A lot of cruisers have watermakers: reverse osmosis machines that turn seawater into fresh water. We couldn’t afford the power or the cost. Instead we bought an emergency inflatable water distiller designed for life rafts as a last resort.

One of the last things we were waiting for was a replacement headsail, the sail at the front of the boat. I’d been talking to a sailmaker in Mexico and asked him to repair our existing one and also make us a new one as a backup. He showed up a couple days before departure with just the repair. So we were crossing with a sail that was a decade or two old, and no backup. At some point you just have to go.

I should say something about John, because he ended up being one of the defining parts of the crossing. I was a 25 year old with almost no sailing experience trying to evaluate the competence of a 69 year old man who claimed decades of it. We’d brought him aboard because we wanted someone with real ocean experience. His resume was hard to argue with: diesel mechanic in the Navy, meteorologist for the Coast Guard, and several ocean crossings. How do you fact-check any of that from a marina in Mexico? His daily input to the actual sailing consisted of maybe twice a day telling us to tighten some line, and that was about it. He’d sit around for hours. When we asked him direct questions (the whole reason he was on the boat) he had this talent for responses that contained no information.

Life at sea

After the first couple rough days, things calmed down. The seas smoothed out and everyone’s bodies started adjusting. Appetites came back. Vito cooked the first of many lasagnas.

The four of us split the night into watches. Someone was always awake, keeping an eye on the sails and the horizon. My shifts were usually 5am to 9am and 9pm to midnight. You sit in the cockpit with the stars and the sound of the hull cutting through water. It’s a lot of time alone with your thoughts, which is sometimes welcome and sometimes not. We’d sometimes set timers and sleep for 15 minutes at a time, popping up like meerkats to check then going back. I spent a lot of my watch shifts writing letters to friends. I’d posted on Instagram asking for addresses and gotten a few dozen. I started with the easier ones (people I didn’t know as well) and was building toward the ones I wanted to put more thought into and was a little nervous to write. Four in the morning in the middle of the Pacific is a strange place to figure out what you want to say to the people in your life, but you have the time.

Vito did a lot of the cooking. On calm days we’d stop the boat and swim in the open ocean. The water was so deep it was almost black looking straight down. It’s a weird feeling, floating in the middle of the Pacific, knowing there’s a couple miles of water beneath you and nothing else around.

Early in the crossing, the wind had this frustrating pattern of dying every night. It would drop to four or five knots, sometimes nothing. As we went over each wave with no momentum, the mainsail would backwind and the boom would swing back and forth, shaking the entire boat. It meant eight to twelve hours of every day where we were barely making progress, just bobbing around.

One afternoon during one of these dead calms, Vito and I were looking at some clouds on the horizon trying to figure out if they were storm clouds. Vito asked John what he thought. John responded: “I know they’re over there and we’re over here.”

As we sailed further south we hit the tradewinds, and everything changed. The wind blew steady day and night. We’d go just as fast at three in the morning as at noon without touching anything. Those were the good stretches.

We caught our first fish about twelve days in. A smaller tuna that was the perfect size for the four of us. I was so fucking stoked about it. I made ceviche, sashimi, nigiri, and spicy tuna rolls right there on the deck. A few days later, after a string of losses, we landed a mahi big enough to keep.

Everything breaks

The spinnaker ripped when a rope came loose during a night watch. Vito and I spent two full days hand-stitching it back together on the deck. It held for a while. Then it ripped again. We taped it. The tape held.

Then came the leak. The bilge pump was kicking on at least every hour, which meant seawater was getting into the boat somewhere. It took two full days of checking every through-hull fitting and crawling around with a flashlight to find it.

It turned out it was the drain in the anchor locker, a tube that passes through the hull at the front of the boat. A mussel had grown in the gap between the pipe and the hull, widening it. I corked it and smeared putty around it, which barely made a difference. In bigger seas it was constantly streaming water. Not enough to sink us, but enough that a pump failure would become an emergency quickly.

Then the tube in the sink water filter exploded. Gallons of fresh water into the bilge before I figured out what happened. For the rest of the crossing we drank straight out of the plastic jugs.

Each failure on its own was manageable. But they stacked up. Power was always precarious. Water had less margin than I wanted. The sails were patched together. The bilge needed constant watching. Every day you were just wondering what would break next and if you’d be able to fix it.

It was in these moments of repairs and exhaustion that self-doubt crept in the easiest.

When you’re catching yellowfin tuna and your boat is planing across the water on a sunny day, you’re suddenly able to figure everything out by sheer force of will. Healthy risk tolerance is for people less effective than you.

When your boat is slowly taking on water from a hole you can’t find, the toilet hasn’t worked for two days, and you’re a thousand miles from the closest person outside your boat, you’re a recklessly naive 25-year-old who thought it was worth risking his life and the lives of his friends and strangers to cross the Pacific without any experience.

There are different ways to frame it in your own head, and different ways other people will frame it for you. I’ve come to believe that both frames are true at the same time. The kid who had no business being out there and the one who made it coexist in the same person on the same boat. It’s not about intentions. If your boat sinks in the middle of the Pacific and a container ship has to pick you up, it doesn’t matter why you started. Both versions of you are real and valid until the outcome decides which one survives. The cat is dead and alive in the box. So you cork the leak, stitch the sail, and keep going, because that’s the only thing that has any influence over which version of you is left standing when the box finally opens.

The equator

As we approached the equator, the weather changed. We started getting squalls. Fast moving rain storms would blow through with a burst of wind and then pass. For a stretch we were getting two during the day and six at night. The constant cycle of reefing and unreefing, adjusting sails in the dark, checking the horizon for the next cloud. It ground everyone down.

Subscribe now

I was sleeping when one of these storms blew in. We didn’t get the sail down fast enough and the boat got blown over maybe 45 degrees. Everything slid off the counters and crashed on the floor. Vito jumped up to help and shouted that the boat was sinking. It turned out that a big jar of honey had just broken on the floor and Vito, half asleep, had just felt the honey on his feet and thought it was water. We spent the next hour cleaning honey off of everything but it beat getting into the life raft.

We crossed the equator two weeks into the sail. There's a sailing tradition Matthew’s grandpa told him about where first-time equator crossers pay tribute to Neptune: you're supposed to be initiated from a "pollywog" into a "shellback." Our version involved tin foil tridents, a shot of tequila and hot sauce, and firing the flare gun off the stern. Not exactly the Royal Navy ceremony, but we hoped Poseidon would be pleased enough.

From there, we pushed south. The doldrums, where the northern and southern trade winds meet and cancel each other out, were exactly as advertised. Barely moving most of the day, then a squall would appear out of nowhere, blow you over, and vanish thirty minutes later. We had the strongest winds of the whole trip here. We also burned most of our diesel motoring through the calms.

The last stretch

The final week was the hardest, not because anything went dramatically wrong but because of the sheer accumulated time. By three weeks in, the excitement had worn off and the destination still wasn’t close enough to feel real. You just exist in this uncomfortable middle. The same routine, the same view, the same people. We started watching more movies.

By the last week, John and I didn’t really get along. I couldn’t hide my frustration and he could tell. The retort to all of it, and the reason I didn’t fully regret having him aboard, was that John was always in genuinely good spirits. He had a kind and open minded energy about him. He never complained, never caused drama. He was just sort of there.

I think a bit the unpleasantness of the crossing was the weirdness around John, and the most of it was the sheer length of it. Three weeks would have been challenging. A month was draining in a way that started to feel existential. But that’s sort of the point of doing a crossing.

But finally we approached the islands. In the distance the sharp peaks of the Marquesas appeared.

Land

We arrived in Nuku Hiva on May 12th after thirty-two days at sea. Pulling into the bay and seeing green volcanic mountains rising out of the water felt unreal. You could smell the land before you could see it — dirt and grass cutting through the salt air. When I stepped onto the dock my legs didn’t know what to do with ground that wasn’t moving. I just stood there for a minute.

After a month with the same three people in the same forty feet of space, everything on land felt almost overwhelming. Time to yourself. New faces. Conversations with people who hadn’t heard your stories. A cold beer and a pizza.

John stepped off the boat, and we parted ways amicably. I think the trip would have been less interesting without him.

One of the locals told us that the Marquesas were actually among the first islands the Polynesians settled, thousands of years ago. They came from the west and landed here before branching out to Hawaii and the rest of the Pacific. You could hear it in the language, which had an Indonesian lilt to it that sounded nothing like the French I’d expected.

The main town was a few roads along the beach — a couple stores, a school, a restaurant where we ate poisson cru three days in a row. Most mornings you’d see guys heading out in small boats to set tuna hooks on floaters, then coming back in the afternoon with fish bigger than anything we’d caught in a month of trying.

A girl named Emma flew in the day we arrived to join the crew. Matthew had met her on mushrooms back in Mexico, and we’d invited her onto the boat without me ever having met her. She turned out to be great and ended up staying for most of the trip. Good luck, considering the gamble. We also met a French guy named Gauthier who’d been hitchhiking between boats. Tall, big beard, looked like some kind of Greek statue that had learned to surf. He’d been bouncing around the Marquesas for a while and we invited him to sail with us to the Tuamotus.

Matthew flew home to Austin for a few weeks to check in on his business. He needed a break, and honestly so did I. The crossing had been hard on the friendship in ways we didn’t really talk about. A year ten feet apart is a lot. We spent the rest of our time fixing everything on the boat that had broken during the crossing, which was most things.

One of the best days on the whole trip happened in the Marquesas. A woman named Maeva and two of her friends invited a group of us up to their hunting cabin in the hills. Wood and corrugated metal, tucked into the bush above the coast. They cooked us a meal that I still think about. Raw tuna in coconut milk and lime juice. Breadfruit tossed whole into the fire for almost an hour. You flake off the charred skin and the inside is like mashed potatoes. Pork, lamb, rice, coconut bread. Our contribution was a tray of brownies baked on the boat. Emma couldn’t stop laughing. Gauthier and I hiked up one of the creeks afterward to find a waterfall with a pool to swim in. It felt like exactly the kind of thing I’d imagined when Matthew and I first talked about buying a sailboat.

It’s the whole point of your stomach. Your GI tract is a 30-foot disassembly line for proteins. Acid denatures them, pepsin cleaves them, trypsin finishes the job. The mucus layer on your intestines blocks absorption as well.

Anything that looks like a chain of amino acids gets shredded into fragments before it can reach your bloodstream. This is what’s supposed to happen. It’s how you digest food. It’s also the reason that for over a century, every attempt to deliver a therapeutic peptide by mouth has run into the same wall.

The first attempt at oral insulin was in 1922, one year after the protein’s discovery. Researchers gave patients the highest doses they could tolerate and got nothing close to the effect of a simple injection. That was a hundred years ago. Since then, thirteen different companies have tried to crack oral insulin. Nine decades of work, and not a single commercial product exists. Peptide drugs on average have oral bioavailability below 1-2%. For comparison, metformin, one of the most common diabetes drugs in the world, has oral bioavailability of 40-60%.

So when Novo Nordisk announced they’d figured out how to deliver semaglutide in a pill, it was a genuinely remarkable scientific achievement. They didn’t fully solve the oral peptide problem, but they just barely squeezed past it using a molecule called SNAC.

Subscribe now

How SNAC Works

SNAC, salcaprozate sodium, was developed by a company called Emisphere Technologies over the course of decades. Emisphere started screening modified amino acid compounds for permeation-enhancing effects in the 1990s, eventually landing on SNAC as their lead candidate. Their first commercial product using SNAC was an oral vitamin B12, approved as a medical food in 2015.

Novo Nordisk partnered with Emisphere in 2007 and licensed the technology for their GLP-1 program. In 2020, Novo acquired Emisphere outright for $1.8 billion, assigned about a hundred researchers in Copenhagen to improving SNAC, and got to work expanding it across their pipeline.

SNAC actually does three things simultaneously:

1. It buffers the local pH in your stomach, which suppresses pepsin activation and protects semaglutide from degradation.

2. It changes the polarity of the solution around the dissolving tablet, which prevents semaglutide molecules from clumping into oligomers, keeping them in their active monomeric form.

3. It incorporates itself into the lipid membranes of your gastric cells, temporarily fluidizing them so that semaglutide molecules can slip through via transcellular transport.

Fifteen years of development, over 9,500 patients across ten phase 3 trials, and the result is 0.8% bioavailability.

99.2% of what you swallow gets destroyed. That 0.8% is the entire product.

SNAC is not a general-purpose permeation enhancer. It’s remarkably specific. Novo tried it with liraglutide, a closely related GLP-1 analog, and it failed because liraglutide forms oligomers that SNAC can’t break apart the same way. Even within the same drug class, against a molecule that differs by a handful of amino acids, the technology didn’t transfer. It’s not a generic “absorption booster” you can swap in and out.

SNAC Alternatives

SNAC is also not the only permeation enhancer anyone’s ever tried. There’s sodium caprate, known as C10. Novo Nordisk actually used C10 to develop an oral long-acting insulin called IO338, which got to phase 2 trials and achieved an estimated oral bioavailability of 1-2%, higher than semaglutide. But it took roughly 60 times the subcutaneous dose to get equivalent blood sugar effects, and the cost of that much insulin made the product commercially impractical.

Then there’s sodium caprylate, C8, used in a technology called TPE, Transient Permeation Enhancer, developed by a company called Chiasma. TPE is the basis for Mycapssa, an oral formulation of octreotide approved by the FDA in June 2020 for acromegaly. It’s the only other FDA-approved oral peptide drug that relies on a permeation enhancer. Its oral bioavailability is about 0.7%. The 20mg oral dose replicates the pharmacokinetic profile of a 0.1mg subcutaneous injection.

So to summarize the state of the art in oral peptide delivery: there are exactly two FDA-approved products that use permeation enhancers to get peptides into your bloodstream through your GI tract. Both achieve sub-1% bioavailability. Both required over a decade of development, thousands of clinical trial participants, and hundreds of millions of dollars. Both use specific, well-characterized chemical enhancers with published mechanisms of action and extensive safety data. And scientists in the field describe the need for better permeation enhancers as one of the largest unmet needs in pharmaceutical development.

What is Hims selling?

Hims doesn’t have SNAC. They don’t have C10. They don’t have TPE. When Reuters asked what absorption technology they use in their $49 oral semaglutide pill, the company said “liposomal technology that is intended to support absorption” and declined to elaborate. No published pharmacokinetic data. No bioavailability studies. No human clinical trials. Novo’s CEO Mike Doustdar said you’re flushing $49 down the toilet. For sure biased. But he’s also describing what happens to every unprotected peptide anyone has ever swallowed.

One rat study exists showing that sodium glycocholate liposomes can improve semaglutide absorption via intestinal bile acid transporters. Rats. Not humans. And “improve” is doing some heavy lifting when you’re starting from a baseline of near zero.

I keep thinking about the person who signs up for Hims oral semaglutide. Someone who couldn’t afford Wegovy at $1,350 a month, watched the price drop to $149 and still couldn’t swing it, sees $49 and finally feels like the system is working for them. They take their pill every morning. They believe their obesity, a genuine chronic disease, is being treated.

Or maybe they’re paying for a mint flavored placebo while a serious condition goes unmanaged for months.

After the Shortage

Hims started during a legitimate drug shortage, filling a real gap, using a compounding exemption Congress created for exactly this situation. Patients got access to injectable semaglutide they couldn’t otherwise get. Revenue went through the roof. Stock hit all-time highs.

The shortage ended in February 2025. The FDA gave compounders until May to wind down. Hims kept going, rebranded “compounding” as “personalization,” built a million square feet of manufacturing facilities, and launched an oral pill with unproven absorption tech while the FDA’s warning letter about their deceptive marketing was still warm.

We decided as a society that "it might work" isn't good enough for something you put in your body every day to treat a chronic disease. We built an entire regulatory apparatus, the FDA, clinical trials, bioavailability studies, post-market surveillance, specifically because "trust us" from a company with a financial incentive to sell you something is not a substitute for evidence.

The system is slow and expensive and sometimes maddeningly bureaucratic and certainly delays life saving innovation. There are times for regulatory discretion.

And honestly, some of the most important medical advances in the last decade have happened in the gray space of regulation. Compounding pharmacies during the semaglutide shortage were operating in that gray space, and patients genuinely benefited. Right to try laws exist because sometimes the regulatory timeline and the patient’s timeline don’t match. The tension is valid and I don’t think the people who push on regulatory boundaries are always wrong.

But there’s a difference between operating in a gray area because the rules haven’t caught up to a genuine medical need, and operating in a gray area because it’s profitable and you’re hoping enforcement is slow.

The FDA just referred Hims to the DOJ. Hims pulled their oral semaglutide. Their stock is down 60% from highs. And somewhere a person is taking their $49 pill every morning, trusting that it works, with no published evidence that it does.

In 2023, I sailed across the Pacific Ocean, from San Francisco to Sydney, Australia, with my best friend Matthew. We had basically never sailed before we left and figured it all out along the way. We met a ton of fantastic people on our trip, and we would not have made it without them.

The whole journey took a year, and I journaled almost daily. These series of blog posts are based on the journal entries, memories, and videos of the trip. I’m planning on splitting up the year of sailing into a few posts.

Part 1 of this series goes from us having the idea to buying the boat and sailing down to Mexico.

Matthew and I met when we were 18 in our freshman dorms at UT Austin. We were fast friends. We connected over our love of the outdoors and spent a lot of our time going on roadtrips and backpacking.

Fast forward 6 years and we were sitting in my apartment in San Francisco during covid. I was working as a software engineer at Apple, and Matthew had started a moving company in Austin that he’d grown into a successful business. We were both watching YouTube, stumbling on the sailboat cruising lifestyle channels. Offhandedly, Matthew said we should buy a boat together and sail. I agreed, and after a couple more conversations figuring out how much money we might need and what the timeline would look like, we just committed and never backed out. We were 24, had some savings, no kids or girlfriends, and no good reason not to. If we were ever going to do something this dumb, now was the time.

We had years to prepare and we spent most of them fantasizing. Before we started boat shopping, the trip felt (and probably was) very much hypothetical. I spent most of my time reading books. Sailors have lots of time on their hands, and so sailing books are very detailed. I remember reading “The Voyager’s Handbook” several times. It covered what felt like everything: how to read storm clouds, diesel engine maintenance, recipes for meals aboard, legal considerations. I read several books on how to survive heavy storms at sea. Lots of them had conflicting advice of how to handle bad weather. When should you try to outrun it? When do you just close all the hatches and hide inside the boat? I had nothing to base my opinions on, but felt like I’d be able to figure it out when the time came.

We found a Groupon that covered two weekends of sailing courses for around $400 in the Texas town of Kemah, sailing around a lake and briefly out into the bay of Galveston. We talked to our classmates about our plan to buy a boat and sail across an ocean, which was taken about as seriously as you should take what a possibly stoned 24 year old says to you in the parking lot of a marina in Kemah. We probably spent about 8 hours sailing over the course.

Boat Shopping

By early 2022, we decided to take it seriously and start looking for a boat. We were incredibly flexible largely because we knew very little. The used sailboat market for boats that can handle larger oceans turns out to be smaller than you’d expect. There are hotspots of sailing in the US, where most of the boat market is concentrated. Florida might have the most, but many of the boats there are absolutely sun beaten and run through. The northeast has their share of boats around Virginia up through Rhode Island. More racing style boats. California has a mixed bag, and the boat market trails all the way up the west coast, with the boats getting hardier as you get closer to Alaska.

I learned a bit about boat design. You’ve got small sailboats for bopping around a bay, racing boats that are fast but spartan, comfortable cruisers for weekend trips, and then (probably the smallest category) boats actually meant to cross oceans. There are formulas for figuring out if a boat is capable of handling offshore conditions, “blue water,” but it’s mostly a spectrum with tons of grey area and varied opinions.

Many boats sell through brokers, and while it’s common to just have a seller’s broker and not a buyer’s broker, I decided we had no clue what we were doing and enlisted a buyer’s agent, a friendly man named John Kuony who had significant experience ocean cruising and brokering boats. We told him our budget and experience, and although he seemed a little skeptical, he was much more open to our plans than most people we met.

We probably ended up seeing 10 boats in person, with most of them in disrepair. The boats that are built to cross oceans are understandably more expensive, and the market much less liquid. The ones that were true blue water boats within our budget were old—’80s or ‘70s—and needed a lot of work. Teak decks would need to be ripped up, masts would have to be replaced. We didn’t have the time for a project boat like that. So we ended up looking for some boats that were in better condition and still within our budget, sacrificing some of the hard “made for crossing oceans” requirements.

Our original plan was to buy in the early summer, fix it up, and leave in the fall—you want to get out of California before winter storm season, make it to Mexico by December, and keep moving south from there. The easiest way to survive a tropical storm is to not be in one.

SV Zeppelin

But by the time we found the boat we’d actually end up buying, it was already late summer. We found “Malu Lani” (which we’d later rename to Zeppelin), a 40.5 foot Hunter Legend that an ER nurse was living on in San Francisco. It was definitely more of a weekend cruiser—a boat comfortable to live on and shallow enough to fit into the small channels of the SF Bay. But it was in relatively good condition, we were growing impatient, and John figured it would probably be able to cruise in open ocean. I found one example of another version of the boat crossing the Pacific online, so it seemed possible. But it was certainly not what the boat was intended for, which, in combination with our lack of experience, gave us some doubts.

We bought it anyways. I started living on it in Alameda, still working in Cupertino and doing repairs at night after work. Almost immediately someone in the marina said “I hope you’re not taking that thing out under the bridge” (i.e., into the ocean). I told him I certainly was. I still think about that guy a lot. Sailing attracts people with strongly held opinions based on personal experience who will tell you anything. Tons of noise. Not just in sailing, most people don’t know what they’re talking about. And if you’re trying to learn from others, you have to figure out how to find the people who have a clue.

We took our new boat to a boat yard in the Bay Area to get some more serious work done—replacing all the rigging that held the mast up, sealing holes, fiberglass work. A big crane took the mast off the boat. When we told the boat yard people we were planning on heading to South America, they also just told us we wouldn’t make it, without clarifying why.

The work there dragged on through Thanksgiving, and eventually to December. By the time we had the boat back, it was almost January. We were way behind schedule. I quit my job, we bought a satellite phone and a life raft, and we left the marina, motoring out of the bay underneath the Golden Gate Bridge, our first time ever in the open ocean.

Under the Bridge

People had warned us that the waves would grow as soon as we made it out of the bay, but our first experience in the Pacific was mild and choppy. Not really knowing how to sail, we just motored down the coast to Half Moon Bay, found a dock, and spent the night. Half out of ignorance and half out of laziness, we never raised the sails once during our first couple days. I think we were worried that actually sailing would force us to confront just how little we knew. We made it most of the way down California on the back of our 50hp diesel engine.

The coast of California is dotted with enough towns and protected bays that you can make day trips almost the whole way—leave in the morning, pull into a slip by evening. That winter, California was getting hit by atmospheric rivers, so we spent a lot of time in marinas finding new leaks in the boat as it rained. But there were gaps in the weather, and we’d make dashes down the coast between storms.

The one exception is Big Sur—about 300 miles of coastline with no protection. You just have to send it in one go.

The day we sailed into Big Sur started lovely. We put the sail up for the first time in the Pacific. You could see the green cliffs in the distance as we sailed a few miles offshore. But as the evening came, a storm swept in. The winds picked up, and we still had our full sail up. When you have a storm blowing in, you’re supposed to pull the sail down—called reefing—before the stronger wind blows you over. But I had only reefed our sails once before, in the bay. So in the dark, with the winds howling, we struggled to get the sails down. We turned on the deck lights, which made the boat feel like a stage bobbing in blackness. I clipped into a harness and went forward to try to unjam the reefing mechanism. The whole boat was heaving up and down six feet at a time. Eventually I just pulled the sails down completely with my hands, lashed it to the boom, and we motored through the rest of the storm, bobbing off every wave until morning.

Southern California

We made it from there to Santa Barbara, Los Angeles, and then San Diego, largely without incident. As we pulled into San Diego, we couldn’t find a marina with a vacant guest slip. We were forced to anchor for the first time, in a bay that was undersized for our boat. We had literally never anchored before. We set the anchor, it felt alright, then we went to sleep. I had an app on our iPad that would sound an alarm if our GPS showed a bigger change in distance than the amount of anchor chain we let out, which would mean our anchor was dragging, not holding the boat in place.

That night, for what would become the first of many restless anchor nights (we really only got good at anchoring toward the end of our trip), the anchor alarm went off as winds picked up. We pulled up the anchor, drove to a different spot, and tried to reset it. You drop the anchor and the only feedback you have is 100 feet of rope and chain. It feels like it’s snug. You reverse on it to see if it’s holding, but it’s hard to tell if you’re staying still when the water is choppy and everything is moving. You’re so tired you convince yourself it’s fine. Thirty minutes later, alarm. Then we pulled our anchor up and tried hooking to a mooring ball that wasn’t being used. Made for boats smaller than ours in winds lighter, the mooring ball started to pull. This was all in a bay where we only had a few boat lengths before we’d wash ashore. We ended up staying up the entire night, repeatedly trying to anchor and different mooring balls, with the wind howling and us never finding a safe place until the sun came up. Exhausted, I begged the marina next to us for a guest slip. They said they had none. But we had no other place to take the boat and the wind was only supposed to get heavier, so we just pulled into their marina and tied off to the end of a dock without permission. It was the safest feeling in the world and I was so happy to be off anchor that I was ready to pay whatever fine they wanted us to pay, but we ended up staying in that marina for a couple days without being confronted. We had to hop the gate every time we wanted to leave.

We ended up stuck in San Diego for weeks waiting for our boat registration paperwork to clear from the Coast Guard. We moved to the guest slip operated by the city—maybe 10 slips, right next to one of the nicest marinas in San Diego with its mega yachts. Our marina was mostly people living on their boats full time, pulling into a guest slip whenever they could. A lot of boats that weren’t really seaworthy.

Living on a boat attracts alternative people. You see two archetypes everywhere: there’s Frank and Susan, the retired couple who sold their roofing business in Utah and now spend half their year cruising on a catamaran. And then there’s Eric—a deadbeat father in his 40s-60s with a whole ass family he doesn’t talk to back in Washington state, a healthy drinking problem, and a boat small enough to singlehand four beers deep. Uses sailing as distraction. Talks your ear off about the best way to insulate your fridge. In the San Diego marina, we were phalanxed by Erics. Hippy Jesus Erics, crackhead Erics. Sailing hobo purgatory.

We went to the San Diego Zoo one day and saw this guy at the gorilla enclosure standing by the glass, showing one of the gorillas something on his phone. The gorilla was glued to the screen. We went up and talked to him and he said he comes every single day after work to show the gorilla TikToks of other gorillas. Just videos of gorillas eating bananas and stuff. He said the gorilla wasn’t interested in any non-gorilla content. The gorilla was just sitting there like a person mindlessly scrolling. Interesting and also a little hellish—this guy spreading the TikTok gospel to other apes.

One close call: we pull the boat out to dump the sewage tanks. I untie from the dock and walk around to meet Matthew at the pump station. He lets the wind push the boat out of the slip, but when he throttles up, nothing happens. The boat’s stuck in neutral. He revs the engine as we start drifting toward a $20M yacht in the marina next door. Everything is happening extremely slowly, but over the next 60 seconds we’re on a collision course. Matthew—who was somewhere between the #6 and #2 fastest swimmer in Texas high schools—ties a rope to a cleat, jumps in, swims to a dock, and pulls the boat just hard enough to catch it about ten feet from the mega yacht. I swear to god Matthew dragged the 20,000lb boat against the wind just by swimming. By now all the employees at the nice marina have run up and started shouting. They had, very understandably, profiled us as crackheads with an out of control boat. We got towed back to our slip. When we opened the engine compartment, we found the prop shaft coupling had fallen off—the plate connecting the engine to the propeller. Our engine vibrated at idle, and after enough hours, the bolts had just shaken loose. We were lucky it happened there and not somewhere worse.

Subscribe now

Mexico: First International Waters

We eventually got our paperwork and pulled out of San Diego, making a one day trip to Ensenada—our first time dealing with customs.

Every country is different with sailboat inspections. Sometimes they search every cupboard of your boat, sometimes they just make you pinky promise you’re not carrying kilos of cocaine. It’s always the agent’s first time doing this, so they’re just as confused as you are. You could never be sure when you were bribing and when you were paying a fee.

Ensenada was a shitty touristy port full of cruise ships and pharmacies with giant Viagra mascots. We got scammed by a shipyard we hired to replace some cracked windows. But our friend Eddy joined us there for our first real passage: Ensenada to Cabo, down almost all of Baja and a few hundred miles offshore.

This was our first real offshore passage—multiple nights, too far to motor the whole way. It was time to actually sail. We got lucky with generally good weather.

There’s an art to balancing how much sail you have up for the conditions—too much and the wind pushes the boat over harder than it needs to. We hadn’t figured this out yet. There’s a lot of stuff like that in sailing, things you don’t know you don’t know until you’re out there learning the hard way. We spent most of the trip down Baja overpowered, the whole boat leaning over, shaking and dipping when gusts blew through.

You get used to having a third point of contact as you walk around. The stove gimbals on a hinge so it stays level, which means from your perspective as the chef, the stove is swinging wildly back and forth while you try to cook.

The Baja peninsula is barren. There aren’t many stops where you can do more than anchor by sand dunes, so we made just one stop down the entire length: Bahia Tortugas, this tiny Mexican town with dusty streets and locals who stared at Matthew as he walked around shirtless. As we pulled into the bay, a guy approached on a boat offering to sell us diesel straight from his hull. We negotiated, bought some, and left the same day.

Bored offshore, we invented new activities. We had a surfboard on board, and Matthew wondered if he could drag behind the boat on a rope like a wakeboard. We stopped the boat a hundred miles offshore—which feels like the middle of the ocean—and Matthew jumped into the water with the surfboard. After a couple tries, he was standing up behind the boat, riding back and forth. Eddy coined this “chumming,” because it was like we were chumming for sharks with a naked Matthew on the rope.

Eddy also blessed us with a culinary invention he called the calorie stack: about 20 crepes, layered with bacon, ham, the most revolting plastic cheese he could find, and syrup between each layer. The end result, an abomination, was sliced like a cake.

As we passed Cabo San Lucas, a whale breached about 50 feet from the boat—something that would become a common occurrence. We stopped in San Jose del Cabo for a week before heading to La Paz.

In San Jose, we met Michelle. Her husband had recently passed, and she wanted to scatter his ashes at sea. We took her out for it, and she ended up joining us for the whole sail to La Paz. We listened to “Chicken Fried” by the Zac Brown Band probably a dozen times. She told us about her husband and their life together. We ended up taking one of his bottles of tequila with us across the whole Pacific.

La Paz: Pacific Crossing Preparation

There are a few hubs of sailboat cruising, and La Paz, Mexico is one of them. It’s well sheltered and sits on the Sea of Cortez, the most biodiverse sea in the world. Many Americans sail down to La Paz for the winter months. And a few use it as a jumping off point for crossing the Pacific.

Unlike Ensenada, La Paz felt like an actual town where people lived. There are boat yards to do repairs, sailmakers to patch sails, shops that sell reverse osmosis water makers and solar panels. It was a lovely place to spend a few months and chip away at the never-ending list of issues and upgrades on the boat.

We met a ton of characters. One was Tim, an Australian guy with a wild story. He’d been a software engineer in Sydney, got stage 4 pancreatic cancer, went to live in the bush and fast, and his cancer went into remission. He ended up bouncing around the world, landing in cattle country in interior Mexico. When covid hit and Australia closed its borders, he decided he’d sail home. He bought a tiny Polynesian-style trimaran in Puerto Vallarta and was in La Paz preparing to cross the Pacific. He had a dog named Bella who had followed him onto the boat and refused to leave.

Tim started his crossing around the same time we did. A few weeks later, a storm knocked out his electronics. He drifted for three months, surviving on raw fish, until a Mexican tuna boat found him 1,200 miles from land. He and Bella both survived.

Our friend Kirema came to visit and stay on the boat. We explored the mangrove forests in La Paz—taking the dinghy into the mouth of a mangrove river until it got too shallow, then wading through the maze. One of the cooler places I’ve ever been.

We spent a lot of time in La Ventana, a kitesurfing town about an hour from La Paz. Staying at a hostel there, we met a Belgian-Italian guy named Vito. He had torn his ACL and was limping around, but he was an experienced kiteboarder and ended up giving Matthew and me free lessons. We became friends quickly.

We still hadn’t agreed on what to do next. We were way behind schedule to make it to South America as originally planned and still have time to cross the Pacific after. With storm seasons as a forcing function, your window for leaving across the Pacific is just a few months in the spring. We had just heard that a whale had sunk a boat making the same crossing. Everyone survived, getting rescued by another sailboat that happened to be in the area. But still, that can happen even to sailors who know what they’re doing.

I desperately wanted to cross. I felt like it would show us places we’d never be able to see any other way—the adventure we’d been looking for. It would also be easier to sell the boat in Australia than somewhere like Ecuador, and I wasn’t sure we’d want to sail back to the US once we got there.

But Matthew was understandably hesitant. We really had no business attempting a Pacific crossing. We had only sailed maybe 400 miles at this point, and the first leg of the crossing was 3,000 miles. We argued a lot—me stubbornly, eventually progressing to begging. It was the only thing that made sense in my head. Other people in the hostel would join the arguments and I’d kick them out when they sided with Matthew. Vito pushed gently for us to cross. I offered a compromise: I would find someone more experienced to sail the first leg with us.

Eventually, under this condition, Matthew agreed to cross the Pacific. He made it very clear he was only on board because he knew how much it meant to me.

By this point, Matthew and I needed a bit of a break from each other. On a boat, you’re living 10 feet away from the other person constantly, trading sleeping shifts. You’re just stuck together, more than you ever are with a roommate or partner you live with. So we found it was good to take little trips apart when possible. We took turns—one person staying on the boat preparing for the crossing, the other backpacking around Mexico.

Vito came to visit us on the boat, cooked dinner, and told us he wanted to join us for the Pacific crossing. He had never sailed before and we had only met a few weeks ago, but it felt right. We told him we’d love for him to come.

That still left finding the experienced person to satisfy my deal with Matthew. I posted on sailing forums and DM’d sailors I followed. I laid out our plans and inexperience, and offered to pay for plane tickets for anyone foolish enough to join us. I got three responses. One was a 40-year-old man who kept asking if there would be girls on the boat. Immediately rejected. Then there was a 79-year-old and a 69-year-old, both of whom I’d met on a forum. They both seemed like experienced sailors. But the 69-year-old—we’ll call him John—had the most impressive resume: diesel mechanic in the Navy, meteorologist for the Coast Guard, computer science professor, several ocean crossings under his belt.

I talked to him on the phone for a couple hours, then we all agreed. We’d spend the rest of the month provisioning the boat and getting everything ready, then John would fly down to join us and we’d leave April 10 for the Marquesas Islands, the first leg across the Pacific.

All babies are floppy. They’re born with muscles not even strong enough to keep their heads up. A three-day -old deer can hop across traffic on I35. Humans take almost a year to be able to crawl.

But still, this guy’s pretty floppy. Out-of-distribution floppy.

He was progressing normally at the start, but at 3 months, he’s not lifting his head up to look at you. So you take him to the pediatrician.

The pediatrician refers you to a pediatric neurologist. 2 month wait. Baby gets floppier. The neurologist sees your baby and orders a genetic test. 2 week wait. Results come back.

You find out that your child is missing the gene that codes for the Survival Motor Neuron (SMN) protein. This protein is expressed in all cells in your body, but (for a reason we don’t fully understand) the neurons in your spinal cord that control voluntary movement die off in the absence of SMN and don’t come back. By 9 months old, most babies with the most severe form of this condition, Spinal Muscular Atrophy, have lost 95% of their spinal motor neurons.

Ten years ago, diagnosis of Type 1 SMA in a baby would come with a discussion of hospice care. There is no cure; take your baby home and love them. Most don’t make it past a year.

About 1 in 40 people carry the recessive version of this missing gene. If a 1-in-40 has a baby with another 1-in-40, 1 in 4 of their babies will have SMA. About 1 in 10,000 babies are born with this condition.

The Cure

In 2016, the first treatment was approved by the FDA. Spinraza is an antisense oligonucleotide therapy. It’s injected into the spine and helps the body produce more of the missing protein.

Spinraza doesn’t give you a copy of the missing gene, it just helps your existing genes express the functional proteins better. It requires lifelong treatment, with lumbar punctures every 4 months for the rest of your life. It didn’t work for everyone, but many babies on it who would’ve previously died started sitting up, crawling, and even walking.

In 2018, Novartis acquired a US gene therapy startup AveXis, in late stage trials for a true gene therapy for SMA. They paid $8.7 billion. And just one year later, their therapy Zolgensma was approved by the FDA.

Zolgensma uses a viral capsid to deliver a functional copy of the missing gene directly to the motor neurons. You only need it once. Your body gets to keep the new genes forever. It’s not a complete cure, but it has completely changed outcomes for babies with this condition. Some babies treated pre-symptomatically are essentially normal.

SMA quickly became the poster child for gene therapy. It’s a well-understood, single-gene disorder. In five years, it went from a death sentence to a treatable condition. The results are dramatic and very visible, with babies sitting and talking. SMA is now screened for at birth, catching cases when treatment is still most effective.

The Math

Zolgensma is also the most expensive drug in existence: $2.1 million dollars for a single treatment.

You hear a price tag like this, and you picture Martin Shkreli gouging desperate parents. But Shkreli raised prices on existing drugs with captive markets. Gene therapy companies are struggling to break even on drugs that didn’t exist five years ago. The economics are genuinely different.

This shit is truly just very expensive, and the $2.1 million price tag isn’t even expensive enough to justify development.

Let’s do some napkin math.

Novartis has to recoup the $8.7 billion acquisition of AveXis. From 2018–2024, Zolgensma has generated roughly $7 billion in revenue. But that’s not profit. For each dose of Zolgensma, Novartis probably pays something like:

• $200k-500k for manufacturing the gene therapy

• $125-150k to license the AAV9 Vector technology to deliver the therapy

• $200k-300k for marketing, distribution, regulatory overhead, and ongoing trials

Standard pharma margin is 70-80%, but gene therapies are much more expensive to produce, pushing margin closer to 40-60%. A generous profit estimate puts Novartis’s profits from Zolgensma around $4B to date. The annual return on investment is something like 5%.

Five percent ROI. Barely beating inflation, and an atrocious allocation of capital by biotech standards. A normal biotech investment expects 12-20% returns. Novartis would have made more money buying index funds.

Novartis didn’t do the deal only for financial reasons. They wanted to establish gene therapy pricing precedent and build manufacturing capabilities they could reuse. There were strategic reasons to push this drug to market.

And sure, Novartis probably overpaid for AveXis. ProPublica found that much of the drug’s early development costs were born by small charities and donations from parents. But everyone involved in developing something like this wants to be compensated for the risk they take. Investors who funded AveXis made billions when Novartis acquired them. Was that fair? Maybe. It’s hard to attract capital to these risks with lower returns.

The Economics of a One-Time Dose

The high price is mostly because development costs must be recovered from a tiny number of one-time treatments, not spread across decades of refills.

Even if manufacturing costs dropped 90%, Zolgensma would still cost over $1 million. The cost isn’t in the molecules; it’s in the years of development spread over a few hundred patients.

And more and more new therapies are one-time. We’ve spent a century working on treatments for chronic diseases. Most of the unpicked fruit on the pharma tree is in rare diseases, and many rare disease treatments will be one-time applications.

This isn’t just gene therapies. CAR-T Cell Therapies and Hep C antivirals also fall into this bucket of expensive, mostly one-time, curative therapies.

The number of FDA approvals for gene therapy has doubled every year since Zolgensma was approved in 2019. There are roughly 4,000 gene therapies in the development pipeline.

We’re about to have 50-100 of these multi-million-dollar cures. We have the technology to cure many of these diseases, but we haven’t figured out how to pay for them.

Despite the increasing gene therapy approvals, pharma companies are pulling back, announcing plans to reduce investments in rare diseases. Zolgensma sales are lower than predicted. Bluebird Bio’s Zynteglo, a gene therapy for a rare blood disorder, was withdrawn from European markets entirely after price negotiations failed. German authorities wanted to pay less than half the list price. An FDA-approved cure, unavailable because we couldn’t agree on payment. Bluebird has since exited Europe completely.

Uncertainty about future prices is a massive burden on development. If pharma companies know roughly what they’ll be paid for a successful therapy, they can develop more confidently and access cheaper capital.

Ideally, society would have a clear view of how much a given therapy is worth to us. Pharma could look at those prices, estimate development cost, and produce the therapies that are profitable to make. Instead, the back-and-forth negotiation in many health systems muddies the signal.

Healthcare policy is about incentive alignment, and aligning incentives for curative therapies is turning out to be one of the hardest problems in medicine.

Subscribe now

What We’re Trying

Every health system in the world is experimenting with ways to make this work. Most of the ideas fall into the same bucket: try to spread a huge, one-time cost over more people and more time, and try to capture the externalities of a cure (avoided hospitalizations, productivity, caregiving, etc.) that don’t show up on a single payer’s balance sheet.

Here’s what we’re trying:

Government Backed Risk Pools

Just a few $2 million claims can blow up a state Medicaid budget. Risk pools spread that cost across larger populations.

The federal government’s most significant intervention launched January 2025, with CMS negotiating outcomes-based agreements with manufacturers on behalf of state Medicaid programs. 33 states plus DC and Puerto Rico have joined, covering approximately 84% of Medicaid beneficiaries with sickle cell disease.

The model covers two gene therapies: Vertex’s Casgevy ($2.2 million) and Bluebird Bio’s Lyfgenia ($3.1 million). If the therapy doesn’t work, manufacturers refund the payers.

Subscription Portfolio Access

What if states just paid a flat annual fee for unlimited access to a drug? Louisiana tried it.

Louisiana launched its subscription model in July 2019 with Asegua Therapeutics (a Gilead subsidiary), paying $35-58 million annually for unlimited access to a Hepatitis C cure for Medicaid and corrections populations. Before the model, less than 3% of Louisiana’s 39,000+ HCV-positive Medicaid patients received treatment at $35 million total cost.

By June 2022, over 11,100 Louisianans had accessed treatment. Per-patient costs fell from approximately $35,000 to potentially less than $10,000, with projections as low as $5,645 if all targets are met.

This worked in part because Hep C is common and pills are cheap(er) to make. Gene therapies are neither.

Bounties & Prizes

What if the government just offered a prize for curing a disease, then let anyone manufacture the cure? It’s worked for vaccines.

The Pneumococcal Vaccine AMC committed $1.5 billion from the Gates Foundation and several governments to encourage distribution of the vaccine globally. Manufacturers received up to $7 per dose initially, subsidized to a “tail price” of $3.50/dose after AMC funds depleted, with 10-year supply commitments at the tail price. It’s estimated to have saved hundreds of thousands of lives so far.

But the vaccines already existed. The prize accelerated distribution, not discovery.

Bernie Sanders has repeatedly introduced the Medical Innovation Prize Fund Act, proposing ~0.55% of US GDP annually to reward successful drug development while denying exclusive manufacturing rights. No version has made it out of committee.

Unlike software or vaccines, gene therapies have substantial per-unit manufacturing costs, making “winner-take-all” prizes less efficient. Clinical uncertainty, small patient populations, and US payer fragmentation all add friction.

Outcome-Based / Installment Payments

Outcome-based agreements (OBAs) structure payments contingent on clinical outcomes, with manufacturers providing rebates or refunds if therapies fail to deliver specified benefits over monitoring periods typically spanning 2-5 years.

Novartis offers a 5-year installment option for Zolgensma at $425,000 annually. Several states negotiated Medicaid contracts with rebates payable if defined clinical health outcomes are not met over 5 years. Specific milestones include motor function measures like the ability to sit without support.

But most contract terms are confidential, hindering learning across payers and limiting transparency for patients and policymakers. Each payer has to reinvent the wheel, and manufacturers can price-discriminate silently.

Reinsurance Markets

Reinsurance and stop-loss products protect payers against catastrophic individual claims, spreading the risk of $2-4 million gene therapy costs across larger populations or transferring risk to specialized insurers.

Optum and Aetna have both brought reinsurance products to market like this, providing some security to self-insured employers.

A typical structure for a $2 million gene therapy: the employer pays the first ~$50K, captive insurance covers $50K–$1M, catastrophic reinsurance covers the rest.

But stop-loss carriers often exclude conditions known to be eligible for gene therapy (like sickle cell disease and hemophilia), creating adverse selection. Employers with known eligible employees struggle to get coverage, shifting costs back onto them.

AI Driven Development

AI platforms have dramatically accelerated early drug discovery. Insilico Medicine, for example, moved from target identification to preclinical candidate in 18 months at a fraction of typical costs. Exscientia reports big reductions in the number of compounds synthesized and tested. Companies like Pathos are using models they’ve trained to make more intelligent decisions about therapies to acquire in early stages.

No AI-discovered drugs have received FDA approval yet. AI accelerates preclinical work but doesn’t necessarily shorten clinical trial duration. A modest reduction in clinical failure rates would save far more than a big reduction in preclinical cost.

Conclusion

Even if money weren’t an issue, we’d still face a prioritization problem. There are thousands of rare diseases. Each gene therapy takes years of specialized expertise to develop, requires manufacturing capacity we don’t have enough of, and needs clinical trials with patients who are hard to find. We can’t cure everything at once. Society has to decide which diseases to tackle first, and in what order.

That’s an economic problem, not a scientific one. And it requires economic mechanisms to solve.

The financial systems we built for medicine in the 20th century were designed for a different kind of drug: chronic treatments for common conditions, sold to millions of patients over decades. Those systems (insurance pools sized for predictable annual costs, R&D models built on blockbuster revenue, pricing based on cost-per-pill) don’t work when the goal is a single curative treatment for a few hundred patients.

We need new mechanisms. Ways to signal to developers which cures society values most. Ways to spread one-time costs across time and populations. Ways to reward successful cures without requiring monopoly pricing. The experiments happening now are early attempts at building that infrastructure. None of them fully work yet. But they’re what iteration looks like.

And there’s reason for optimism. As AI accelerates drug discovery and manufacturing processes mature, the cost of bringing each therapy to market will fall. The denominator in that brutal equation (development cost divided by treatments) won’t change. But the numerator will. Diseases that are economically impossible to cure today might become feasible in ten years, then routine in twenty.

The scope of what we can cure as a society is going to expand. The question is whether we’ll have built the financial systems to actually deliver those cures to patients, or whether they’ll sit on shelves, technically approved, waiting for someone to figure out how to pay.

Before urbanization, all doctors were generalists—delivering a baby in the morning and treating pneumonia in the afternoon. The idea that you could choose to treat only one specific type of medical issue was impractical at the time.

With cities came the ability to specialize. Obstetricians no longer performed amputations. Doctors could focus on one aspect of human health, seeing far more cases in their specialty than any generalist could. This concentrated expertise allows specialists to follow patient outcomes more closely, notice subtle patterns, understand how patient differences affect outcomes, and innovate based on larger sample sizes.

Specialization in medicine drove fantastic outcomes. Infant mortality plummeted over 90% from 100 per 1,000 births in 1900 to under 7 by 2000, with the steepest declines following the establishment of pediatrics as a specialty. As anesthesia became its own specialty, instead of a responsibility of general surgeons or doctors, mortality rates under anesthesia dropped 100-fold.

Experience Driving Outcomes

Specialists contribute in two ways: they innovate because they can dedicate their full attention to solving specific problems, and they excel through repetition—performing the same procedures until they achieve mastery.

Complex procedures show 2-3 times higher mortality when performed by low-volume surgeons versus high-volume specialists. Specialization is a proxy for volume; volume drives pattern recognition and better outcomes. Surgeon mortality rates are inversely related to case volume across all procedures—the adjusted odds ratio for operative death ranges from 1.24 for lung resection to 3.61 for pancreatic resection when comparing low-volume to high-volume surgeons.

5-year survival rates for cancer patients at specialty cancer centers are 53% versus 44% at community hospitals.

The doctors at Sloan Kettering are likely more experienced, but critically they're also more specialized. The oncologist on your tumor board has seen perhaps 100x more patients with your specific cancer type than a rural general oncologist. That disparity means 100x more opportunities to refine their clinical decision-making through real-world feedback.

Diminishing Returns

But evidence suggests that we may have reached diminishing returns on human specialization. The U.S. has a 2:1 specialist-to-generalist ratio—one of the highest in the world—yet demonstrates poorer health outcomes than countries with stronger primary care systems.

We’re hitting a wall in improving our quality of care.

A large contributing factor is care fragmentation. Patients shuttling between multiple uncoordinated clinicians, often repeating tests and receiving conflicting treatment plans. A Danish study of 4.7 million patients found that highly fragmented care—seeing multiple specialists without coordination—increased mortality risk by 2.59 times, even after adjusting for disease severity. In the U.S., 30% of Medicare beneficiaries see five or more physicians annually, creating communication breakdowns and treatment conflicts.

Why Specialization Keeps Growing Despite the Evidence

Despite evidence that fragmentation worsens outcomes, hospitals keep adding specialists. Largely because of financial incentives.

The way we’ve set up our healthcare system encourages care to take place in hospitals, by specialists. And hospitals have gotten way more consolidated in the last 30 years, mostly through regulatory capture.

Notably, the Affordable Care Act sharply limited Medicare participation and expansion of physician-owned hospitals. These physician-owned facilities had consistently outperformed traditional hospitals on quality, efficiency, and cost.

Between 2007 and 2013, the number of physicians employed by hospitals nearly doubled from 96,000 to 182,000. When hospitals acquire physician practices, they can immediately increase billing through site based billing. Medicare pays significantly more for services performed in a hospital setting. And specialists bring in almost 3 times more revenue per clinician than PCPs. Hospital margins have increased significantly through consolidation and specialization, which has been a major driver for total healthcare spending.

The same specialization that improved medicine now drives revenue through fragmentation. Hospitals will find it challenging to reduce fragmented care—the financial incentives all point the other way.

Specialist Access

Hyper-specialization creates cascading access problems. Patients can't find the appropriate specialist for their condition. A patient with joint pain must navigate between rheumatologists, orthopedic surgeons, sports medicine physicians, and pain management specialists—multiple consultations before finding care.

Wait times compound with specialization. Pediatric rheumatologists average 3-6 month wait times; some subspecialties lack practitioners in every state.

The training pipeline cannot scale to this. Subspecialty fellowships now extend training to 7-10 years post-medical school. The narrower the specialty, the fewer doctors we produce.

Shifting Expertise to Systems

To maintain the quality that specialization brought medicine, specialization must shift from practitioners to systems.

AI tools already augment generalist capabilities. Google's diabetic retinopathy algorithm enables primary care clinics to screen for vision loss with specialist-level accuracy. Point-of-care ultrasound with AI allows emergency physicians to obtain cardiac scans previously requiring sonographers. ECG interpretation algorithms identify atrial fibrillation with cardiologist-level accuracy, enabling immediate treatment decisions in primary care.

Clinical decision support systems transform pattern recognition. Mental health platforms analyze speech patterns and behavioral data to flag depression severity, providing psychiatrist-level assessments to counselors and general practitioners. These tools don't replace clinical judgment—they augment it with pattern recognition derived from populations no single specialist could observe.

Many of these tools are FDA-cleared for specific uses—typically detection or triage—not full diagnosis. Clinicians remain responsible for medical judgment and treatment decisions.

Rather than choosing between generalists and specialists, the approach involves providing generalists with access to specialist-level pattern recognition through systems that have learned from more cases than any human specialist could see in a career. Doctors can focus on building relationships that drive adherence, navigating complex decisions with patients, and coordinating care.

Specialization saved medicine by concentrating expertise. Now we can embed that expertise in systems accessible to both doctors and patients. Your primary care physician can access specialist-level pattern recognition for complex conditions. You can get dermatologist-quality skin assessments from your phone. We need specialist knowledge everywhere, available to everyone who needs it.

At the time, scurvy was devastating naval forces. Naval historian Nicholas Rodger wrote "It has been seriously suggested that a million British seamen died of it in the eighteenth century – a figure which implies that everybody who served in the Navy died of scurvy approximately twice." Even though Lind ran out of citrus after six days, the sailors eating lemons recovered quickly enough to care for their companions still suffering through sulfuric acid treatments.

Today, we run 74,000 clinical trials simultaneously. Finding the right patients is considerably harder than rounding up sailors with scurvy. And we're a bit pickier about sample sizes.

The Trial Matching Landscape

The business of finding patients for clinical trials is massive. Pharmaceutical companies spend billions annually on clinical trials, with a significant portion going to Contract Research Organizations (CROs) and trial operators who handle patient recruitment. These operators can charge tens of thousands of dollars per enrolled patient, making successful trial matching incredibly valuable.

But recruitment challenges are widespread: approximately 80% of clinical trials are delayed or closed due to recruitment problems, with delays costing sponsors millions. Even when trials do launch, 37% of sites under-enroll and 11% fail to enroll a single patient.

For patients, particularly those with cancer, clinical trials can offer access to potentially more effective treatments. Since the trial sponsor covers costs, trials provide a route to cutting-edge therapeutics that might otherwise be inaccessible, especially where public health systems don't cover newer treatments.

Trial matching grows more complex as medicine advances:

Precision Medicine: As treatments become more targeted to specific genomics, eligibility criteria become increasingly specific. This shrinks the potential participant pool dramatically, forcing operators to search harder for qualified patients.

Geographic Complexity: Modern trials recruit internationally, with decentralized setups where no single site hosts all participants. Operators coordinate across multiple locations while maintaining consistent protocols.

Diagnostic Barriers: Many eligibility criteria require biomarker or genetic tests that aren't part of standard workups. Trial operators often need to fund additional testing just to determine if someone qualifies.

The value and complexity of this problem should make it attractive to startups. But despite the massive opportunity, it's become a graveyard for new companies due to complex enterprise sales cycles, siloed patient data, and complicated stakeholder relationships (read: healthcare). LLMs have started tackling pieces of the problem (like parsing eligibility criteria), but we're still far from a complete solution.

A Digital Registry

ClinicalTrials.gov was created in 2000 as a repository for clinical trials, both in the U.S. and internationally. Later in 2007, it became mandatory for any clinical trial operating in the U.S. to register on this site, dramatically expanding its scope. Each listing typically includes:

• Title and study description

• Trial type and phase

• Locations (sites) administering the trial

• Eligibility criteria, specifying who can and can’t participate

The registry now contains over 500,000 trials and sees millions of visitors each month — it's become the global standard for trial registration and transparency. Researchers, clinicians, and the public can freely access detailed information about almost any intervention being tested.

There's no shortage of AI trial matching software, but almost every solution I found lives behind enterprise walls, only accessible through specific health systems. This makes business sense - these tools need to integrate with medical records and often require complex infrastructure. But it also means we can't compare how well different approaches work, and leaves patients and doctors outside those systems without access to these tools. It's frustrating — there's been massive effort to make trial data public through ClinicalTrials.gov, but the tools to make that data truly accessible aren't available to most people who need them.

I wanted to explore how modern AI search and reasoning techniques could help make this public data more accessible. The first challenge was improving on ClinicalTrials.gov's search functionality.

Semantic Search

ClinicalTrials.gov has basic keyword search, but that's not enough - when you search for "metastatic melanoma," you should find relevant trials even if they use different technical terms like "stage IV melanoma" or describe the condition in other ways. Embeddings make this semantic search possible by mapping text into a space where similar meanings sit close together, regardless of the exact words used.

Mapping Searches to Trial Structures

When we embed text into a high-dimensional vector space, semantic similarity translates into geometric proximity. But there's a nuance: even a clinically detailed query like "seeking Phase 2 immunotherapy trials for metastatic melanoma" doesn't naturally map to how trials are structured in the database. The search “looks” different from the documents being searched.

Clinical trials follow strict documentation patterns. A single concept might appear differently across a trial's listing:

• The title follows a rigid academic format ("A Phase 2 Study of Anti-PD1 Therapy in BRAF-mutated Melanoma")

• The summary uses standardized medical terminology and trial conventions

• The detailed description contains background, methodology, and rationale

To bridge this gap, I use hypothetical documents - imagined versions of how the user's query could appear if it were actually a clinical trial. When someone searches, the system generates multiple variations that mirror real trial documentation: a title version, a summary version, a detailed description version, and an eligibility criteria version.

By embedding these hypothetical documents alongside our actual trial database, we can project the user's query into the same spaces where trial information naturally lives.

The final search uses a weighted combination of these embeddings. This helps ensure we catch relevant trials regardless of where and how they encode the matching criteria - whether it's in the technical language of the title, the structured format of the eligibility criteria, or the detailed methodology sections.

Visualizing Trial Similarity

The embeddings we generate live in hundreds of dimensions - far more than we can visualize directly. To help validate how our search system is working, I use UMAP to project these high-dimensional vectors down to 3D space. While this reduction loses some nuance, it gives us a way to sanity check how different trials cluster together.

You can explore these visualizations in the tool by toggling "Show 3D trial view". The visualization shows your search query (projected through the hypothetical documents) as a point in space, surrounded by the nearest clinical trials. Different types of queries create distinctly different patterns:

A focused query like "breast cancer" creates a tight cluster - you'll see trials grouped closely together, with related terms like "breast carcinoma" appearing nearby. This clustering suggests our hypothetical documents are successfully replicating the standardized language used in oncology trials.

More complex queries reveal interesting patterns. Search for "gunshot triage and high blood pressure" and you'll see the query sitting between two distinct clusters: trauma-oriented trials on one side and cardiovascular trials on the other. The most relevant trials appear at the intersection, bridging these different medical domains.

While these visualizations are mostly a curiosity (the actual search uses the full high-dimensional embeddings), they help validate that our embedding approach makes intuitive sense. When we see trials clustering naturally by disease area and methodology, it suggests our hypothetical documents are successfully projecting queries into the right regions of the embedding space.

From Searching to Matching

Semantic search gets us part of the way there, but trials have complex eligibility requirements that go way beyond what you can capture in a simple search. Even if we find trials that are conceptually related, we need to try to filter out ones that are clearly not a match. That's why I built an agent system on top of the semantic search. Here's how it works:

1. Parse the User Query

The first challenge is making sure we understand the query properly. When someone enters something like "metastatic melanoma with brain metastases," the agent tries to extract and structure the key medical details. This helps ensure we're starting with high-quality inputs before we even begin searching.

2. Run Semantic Search

The agent uses our embedding system to find potentially relevant trials. This gets us a good starting set, but semantic similarity alone can be misleading - a trial for "BRAF-positive melanoma" will appear very similar to one for "BRAF-negative melanoma" in embedding space, even though they're looking for entirely different patients. This is why we need the next steps.

3. Identify Missing Details

Clinical trials often need specific information that most people wouldn't think to include in a search. The agent looks at the eligibility criteria of the most relevant trials and generates targeted questions: "Has the patient had prior immunotherapy?" or "Do you know their BRAF mutation status?" For each question, it analyzes how different answers would affect which trials remain viable matches.

4. Filter Trials by Eligibility

The system shows users both the most relevant trials and the follow-up questions. As users answer these questions, they can see in real-time which trials would be ruled out by their responses. We also show "near miss" trials and explain why they might not be a match, which helps users understand their options better.

Try It

trialsearcher.com is now live and pulls updated data from ClinicalTrials.gov daily. Go try it for free. There’s a small cost to each search, and the cheap Postgres instance I’m using for vector search means performance can be slow. But I’d love to keep it free, so please give it a try and let me know if it’s helpful.

• Is it better or worse than your usual way of finding trials?

• Any features you’d like to see added?

• Interested in a higher-volume usage or integrating it with your own system? Let me know—we can explore hosting options that won’t slow down the public tool.

Final Thoughts

Vitamin C wasn't isolated until 1928, almost 200 years after Lind's scurvy trial. But scurvy was cured long before anyone understood what caused it. It's tempting to think we've progressed beyond this empirical approach, but sometimes throwing things at the wall remains one of the fastest ways to advance therapeutics.

I've had the privilege of talking to some brilliant folks in trials, semantic search, and precision medicine. Their insights helped shape this approach and kept expectations realistic. While LLMs can help with parts of trial matching (like interpreting eligibility text), the full problem remains far from solved and goes much deeper than trial eligibility matching.

Disclaimer: This tool is a proof of concept and not meant to replace professional medical advice. Please consult healthcare professionals before making decisions about clinical trial participation.

Two people with identical cancers - same type, same stage, same tumor size, same location - walk into different hospitals. This cancer has been studied extensively through clinical trials, case studies, and meta-analyses. There are experts worldwide who have dedicated decades to studying and treating this type of cancer. If you gathered all these experts in one room, they would agree on an evidence-based course of treatment (or further tests) for either patient.

In an ideal world, both patients would receive this optimal treatment at any hospital they visit. But the reality of medicine is complex - doctors work under time pressure, have different training backgrounds, and may not have immediate access to the latest research. This is where clinical guidelines come in. Guidelines distill the vast body of medical evidence into structured recommendations, helping standardize care across different hospitals and clinicians. While individual responses to treatment vary, following evidence-based guidelines gives patients the best chance for positive outcomes.

The NCCN and Modern Guidelines

Every year, top oncologists who specialize in a particular type of cancer lay out and enumerate the different types and situations they might see in care.

• What if there's early stage breast cancer, with a previous partial mastectomy, and the patient is currently pregnant?

• What if there's invasive cancer in the breast ducts, and the tumor is between 0.6 and 1cm long?

For each situation, the panel reviews existing clinical evidence and recommends an action. That could be ordering certain tests, prescribing a course of treatment, or scheduling future check-ins to see how the situation is changing. The panel meets regularly to review new evidence and update the guidelines. There are several groups around the world who all draft different guidelines. Some courses of treatment aren't available in some countries, so guidelines are regionally applied.

The National Comprehensive Cancer Network (NCCN) is an alliance of 32 leading cancer centers in the United States. Each year, they bring together panels of multidisciplinary experts for each cancer type. These panels review the latest research, debate the evidence, and work to reach consensus on the best approaches to screening, diagnosis, treatment, and supportive care. The resulting guidelines—“The NCCN Clinical Practice Guidelines in Oncology”—are the most comprehensive and widely-used standards for cancer care globally.

The Reality of Implementation

This is a page from the NCCN breast cancer guidelines:

Each of those hyperlinks (like BINV-6) is its own page with its own flow chart, pointing to more pages. As a doctor trying to interpret guidelines for a patient, your job is to apply the patient's situation to the flow chart, continuing until you hit some information that you don't know (like the HER2 status), in which case you order tests to find it out. Or you hit a recommendation for a course of treatment.

It's not perfectly standardized. Sometimes there's multiple options for treatment, each with similar levels of support. Sometimes there's a patient situation that is relevant to treatment but isn't enumerated in the guidelines. Doctors still have to make judgment calls, even when they work with guidelines. But the space of possible judgment calls is smaller. And the chance that a patient receives a treatment even though another treatment has stronger clinical support is lower.

Centers of excellence like Mayo Clinic and MD Anderson achieve better cancer outcomes for several reasons: They're well funded, their oncologists are highly specialized, and they have broader access to potentially life-saving clinical trials. But one other important factor is their systematic approach to care, including robust processes for implementing clinical guidelines. Many of these institutions also play a crucial role in developing the NCCN guidelines, contributing their expertise and research.

At any hospital, care can deviate. Criteria can be missed. Maybe a postmenopausal patient is given a treatment that is better supported for premenopausal patients because it was missed in her charts. Doctors see hundreds of patients, and even with access to guidelines, it’s inevitable that something will be overlooked.

There's so many different ways to improve cancer outcomes. You can discover new drugs. You can improve surgical techniques. But an immediate, universal, and attainable approach is to reduce the gaps in evidence-based care.

The PDF Problem

Part of the issue with guidelines is their format. The guidelines represent hundreds of thousands of hours of expert oncologists’ time. But the end result is a dense and hard to navigate PDF.

Imagine an oncologist faced with a rare subtype of breast cancer they see perhaps once every few years. They need to first locate the correct PDF version, then identify the right section based on the cancer's characteristics, then follow multiple hyperlinks across different pages while keeping track of various patient factors - all while managing their heavy patient load. The guidelines are also constantly being revised, with new versions releasing as separate PDFs. It's easy for doctors to miss updates or continue referencing outdated versions that have gotten buried in their files among other clinical documents.

At their core, guidelines are decision trees. The teams that have drafted them have spent an considerable effort wading through evidence and corralling it into a decision tree representation. But the guideline documents that describe these trees are difficult to follow.

With properly structured data, machines should be able to interpret the guidelines. Charting systems could automatically suggesting diagnostic tests for a patient. Alarm bells and "Are you sure?" modals could pop up when a course of treatment diverges from the guidelines. And when a doctor needs to review the guidelines, there should be a much faster and more natural way than finding PDFs.

The organizations drafting guidelines should release them in structured, machine-interpretable formats in addition to the downloadable PDFs. Ideally, all guideline organizations would agree on the same format, and all patient data systems would be able to interpret that format. But that sort of data approach isn't their specialty.

Giving Guidelines Structure

Recently, I had the chance to work with some talented oncologists as a software engineer. It gave me a window into this world of cancer care and guidelines that I never expected to see. The problem of these guidelines being trapped in PDFs stuck with me, even after I moved on to other projects. I couldn't stop thinking about these crucial decision trees sitting in these dense documents.

So I built a small proof-of-concept tool to see if I could extract the guidelines into something more structured that machines could understand. If we could get the guidelines into a proper format, we could build better interfaces for doctors to navigate them, and maybe even get computers to help follow them.

Schema Design

I started by defining a schema that could represent most of the information in the NCCN guideline. The guidelines look like trees of information. Because some references can loop around, there can be cycles. So the whole set of guidelines for a type of cancer breaks down into a few disjointed directed graphs.

Each point in the graph is represented as a node, whether it is a condition, qualifying characteristic, treatment option, or something else. The nodes also store the associated references and footnotes along with them.

Each arrow in the flow chart or reference from one page to another is represented as a relation between nodes.

Data Extraction

To construct the full guideline tree, I perform a first pass over the document with an LLM to create references for each page in the guideline. These page references are the parent nodes for the top nodes on each page. When another node points to a certain page, it points to that page's node.

Then, I use an LLM to read through the guidelines in small chunks of pages, and extract the guidelines in a JSON representation of the schema, passing in the page references:

I ran this extraction on the NCCN Breast Cancer Guidelines and saved the nodes and relations to a database. I ended up with 271 nodes in total.

Visualization

The next step was building a graph viewer. I used the React Flow library, and built an interface that displayed all the root nodes (the pages that didn’t have incoming references). Clicking on a node reveals its children, and you can follow a train of nodes down the line.

Agent Implementation

I wanted to understand how this structured representation would make it easier to go from a patient case to guideline recommendations. I built an agent that could take a patient history and traverse the graph. First, it finds the top-level node that most closely matches the patient's situation. Then, it looks at the children node and chooses the one that matches the patient most closely. It repeats this until it arrives at a leaf node, or it does not have enough information to continue.

You can try the tool for yourself here.

Next Steps and Limitations

The data in the demo has not been reviewed and contains errors and omissions. Please don't use it for any clinical decisions.

I spent a few hours working on the extraction process (and even less on the agent), and you could achieve more accurate results with more effort. I sit around 70–80% accuracy right now. LLM extraction is an easy way to prototype this, and with human review it would be reliable, but ideally these guidelines would be drafted in a structured way from the beginning.

Other people are working on this. I purposefully didn't look around much before starting this, because that usually kills my motivation to build anything. If you're working on this, or work with guidelines, or are building a product you think this could be useful for, please reach out. I'd love an excuse to spend more time on this and collaborate.

The data is still only semi-structured. Information like causal relationships and how nodes should be evaluated still requires natural language understanding of each node. Future work could define an even more structured schema that would be easier to evaluate.