Stanford OnlineAI in Healthcare Series: Have We Already Bent the Healthcare Cost Curve?
Stanford Online
Guests Debate a Provocative AI Paper 0:12
The episode opens with Justin welcoming Eric Larson, a longtime healthcare advisor and investor, to discuss a recent paper by Zeke Emanuel, Neil Kosa, and Vinod Khosla arguing that autonomous AI can already exceed physicians. Matt, the practicing radiologist among the three, calls the piece provocative but grounded in a real trend: in study after study, AI plus physician performs worse than AI alone. He resists the leap to letting AI treat patients autonomously, citing what he calls the doorman problem, borrowed from a Bob Wachter essay: AI can beat a physician on a narrow task like spotting pneumonia, which is only about one percent of a radiologist's job, but medicine is a huge bundle of tasks that no single benchmark captures. Watching any clinician for five minutes, he says, reveals a dozen more things AI would need to master.
Scarcity, Chess, and Dislocation 4:31
Eric argues the paper echoes Vinod Khosla's 2016 "Dr. Algorithm" essay and compares medicine's moment to Garry Kasparov's 1997 loss to Deep Blue, followed by a human-plus-machine era that ended around 2012 when machines began beating everyone outright. He suggests healthcare's six-trillion-dollar structure has always rested on scarcity of cognition, such as the twelve years needed to train a radiologist, and that this scarcity is now dissolving. Matt adds that the conversation still isn't thinking exponentially enough, comparing it to early COVID debates over docking cruise ships when spread was already exponential.
A Predictable Pattern of Resistance 10:31
Eric cites Elton Morris's 1955 essay describing how societies meet radical technology: first ignoring it, then rationally rebutting it, then mocking its champions, until a credible leader forces adaptation. He sees medicine, law, finance, and consulting all entering this cycle, with intellect itself as the prestige hierarchy now being challenged, raising unresolved economic and social questions about mass job dislocation.
The Cursor Example 16:02
Matt Ross describes how a small seven person team built Cursor and outcompeted Microsoft, even though Microsoft held OpenAI's technology, owned GitHub, and had the world's leading coding tool. The lesson he draws is that focus and lack of legacy baggage let a small insurgent beat a giant incumbent, and he asks what this means for healthcare, a field built on the assumption that knowledge and expertise are scarce. He points out that academic medicine, its teaching methods, evidence gathering, and peer review, were all designed around that scarcity, yet technology now moves faster than papers can be published or systems can be updated.
Cognitive Scarcity Becoming Abundant 19:30
Eric Topol calls this the most exponential technology in history and notes that even the most prestigious, centuries old institutions, like the AMA, the Cleveland Clinic, and the Mayo Clinic, are confronting it. He explains that hospitals exist as trillion dollar real estate complexes because specialists need to be physically gathered together, a response to medical knowledge growing too fast for any one mind, noting that the time for medical knowledge to double dropped from fifty years in 1950 to about 73 days by 2019. His question is what happens to institutions built to guard and ration expertise once large language models make that expertise nearly free and available to anyone, turning individuals into polymaths who can act as their own lawyer, doctor, or consultant. He notes the US spends 18 percent of GDP, six trillion dollars, on an inequitable health system, and asks whether existing institutions can survive in their current form or whether the transition must be handled thoughtfully and compassionately.
Can Incumbents Self-Disrupt 25:32
Justin Norden pushes on whether current institutions can adapt or whether insurgents will force the change, and Topol points to Satya Nadella losing a trillion dollars in market value around the Stargate announcement as proof that even strong incumbents can be caught flat-footed. He argues incumbency is a head start, not a guarantee, especially in a highly regulated, litigious sector where moving too fast can cost lives, and mentions Tsinghua University's new AI hospital in China as an example of building healthcare from scratch. Ross offers a middle path, comparing it to Salesforce, where an incumbent could become headless and agentic rather than being replaced outright, letting existing systems like Epic or Oracle serve AI agents instead of only human users, though he is unsure any current system of record is ready for that shift.
Current tools already outpace healthcare adoption 32:01
Leading organizations like OpenAI, Anthropic, and startups such as Cursor are already using today's AI tools orders of magnitude more effectively than the median healthcare organization, and that gap keeps growing. The point made is that healthcare leaders don't need to wait for better technology before acting, since the tools available right now are more than enough to fundamentally change the industry if applied well.
A Harvard paper on healthcare costs 33:00
Eric introduces a Harvard paper by Cutler and Clarett, still a preprint, which he says changed his thinking. He had been arguing informally that AI could cut 500 to 700 basis points off the share of US GDP spent on healthcare, a claim that made him unpopular. Healthcare is a six trillion dollar sector employing 23.8 million people, one in six working adults, and it suffers from Baumol's cost disease, meaning it must raise wages to compete for labor without gaining matching productivity, which pushes prices up.
Costs already came in far lower 35:32
CMS actuaries in 2010 projected healthcare would reach 21.2 percent of GDP and 6.3 trillion dollars by 2024. The actual figures were 18 percent and 5.3 trillion dollars, a gap of 6.7 trillion dollars saved over that 14 year span, the largest divergence since tracking began in 1960. The authors attribute about 14 percent of this to technology, with the rest from shifts in site of care, a healthier population, pre-authorization, and payer clawbacks, all before AI played a role. Eric also notes GLP1 drugs, though costly at first, are already showing a 5 percent drop in total healthcare costs among adherent employees, echoing past shifts like inpatient to outpatient care such as total hip and knee procedures, which alone saved 94 billion dollars in 2024.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
