We discuss these and other timely topics with John. At an early age, he fell in love with science fiction and built vacuum-tube amplifiers in his parents' basement, before studying physics and materials science at Stanford. There he became a research scientist until pivoting to investing, first at StartX and later Spark Capital. Born out of firsthand experience, he co-founded Standard Metrics alongside the 8VC Build team to create a better solution for portfolio management software.

We begin our conversation with John’s path from academia to investor and founder. Next, we explore how Standard Metrics centralizes data, improves portfolio intelligence, and powers smarter investment decisions. Then we dive into the new possibilities with AI, from technical diligence copilots and investment stress-testing to new workflows and internal AI analysts. In an era of new agentic tools and shifting competitive advantages, we discuss new moats for existing software companies built around network effects, data, and more. The world of finance is changing quickly; John offers a unique perspective on the AI wave and how top investors are leveraging new workflows to get ahead.

Subscribe now

00:00 Episode intro

02:10 Academia to investor and founder

08:00 The pain point that led to building Standard Metrics

12:10 How AI is transforming venture capital

14:20 AI as a technical copilot

17:10 AI Analyst vs human in the loop

19:20 Parsing pitch decks and new AI tools

23:10 How AI search is changing marketing

28:30 New capabilities and workflows

33:47 How quickly is everything changing?

Raised in a small town in North Carolina, Will developed an early love for building, starting with electronic kits and later unique apps. After studying at Penn and Wharton, a trip home revealed the looming crisis: millions of local businesses that lack succession plans. This sparked the creation of American Operator, which pairs retiring owners with high-agency operators, along with the capital and know-how to build a thriving business.

We begin with Will’s entrepreneurial journey and the coming “Silver Tsunami.” Will explains why only 48% of small business owners have succession plans and what this means for the next generation of owner-operators. He highlights stories like Greenway Painting in Jackson, Wyoming, and how American Operator is helping create millionaires on Main Street while keeping local business local. Next, we explore the historical and cultural importance of small business in America, dating back to John Hancock’s import-export firm and Ben Franklin’s printing shop. Looking ahead, we examine AI’s impact on small businesses. Learn why Will believes SMBs are a great AI hedge, while also having unique upside as new AI tools make small cohorts of people dramatically more productive. Finally, Will explains the operate-to-own model and how American Operator aligns incentives so that original owners become valuable advisors, new operators can earn their way into majority ownership, and American Operator remains a committed partner over the long run.

Subscribe now

00:00 Episode intro

01:25 Will’s entrepreneurial journey

04:15 The small business succession crisis

06:40 American Operator’s approach vs. private equity

11:55 Does AI mean boom or bust for small business?

15:55 Forget law school? Buy a small business instead?

21:20 AI use cases for small business

23:00 Why many business owners shut down instead of sell

26:00 How the operate-to-own model works

33:50 Optimism for Main Street America

Subscribe now

00:00 Episode Intro

01:50 What sets Joby aircraft apart

04:50 Key partnerships with Toyota & NVIDIA

06:50 Closing in on FAA Certification

08:00 Air Taxis coming to 12 states this year!

13:35 Autonomous flight is the future

20:00 Hydrogen aircraft for the Pentagon

24:10 Why AI is a game changer

28:15 Expansion in Ohio and Global Competition with China

34:40 Addressing critics / future vision

35:50 Making air travel ubiquitous

We discuss these and other timely topics with Scott and Russell. Scott was a three-time gold medalist at the International Olympiad in Informatics and world champion at age 17. After high school, we hired him at Addepar, where he became a top software engineer. Russell began his career as a machine learning engineer on Tesla’s Autopilot team before selling his video data company, Helia, to Scale AI. In 2023, Scott and Russell co-founded Cognition, and a year later, they shocked the technology world with the release of Devin, the first AI software agent.

We begin our conversation by discussing the incredible collection of young talent at Cognition, and why the next generation has new advantages in the AI era. Next, we catch up on Cognition’s explosive growth: Devin usage in the first few months of 2026 already surpassed all of 2025. Scott reveals that Cognition engineers no longer write code and explains how they’re able to test and ship new products faster than ever before. Then, we dive into the new era of software abundance and what it means if everyone has access to high-quality engineering, from modernizing large legacy enterprises to supercharging small businesses. We also discuss Cognition’s recent foray into government services and its work to modernize complex outdated systems. Finally, we explore the talent flywheel that has drawn so many former founders to Cognition, and why Scott and Russell believe we’re moving from Minecraft “survival mode” to “creative mode” — where the only limit to building is imagination itself.

Subscribe now

00:00 Episode intro

01:35 Why technical talent & execution matters in AI

06:10 Do young people have an edge in the AI era?

08:26 Cognition’s rapid growth

11:55 The new era of software abundance

14:30 Cognition engineers don’t type code anymore

19:20 “Never sleep while Devin is idling”

21:25 The case for AI disinflation

23:50 How Devin generates 12X productivity gains

28:25 Cognition for government / taking on complex, broken systems

36:40 The AI race / competition with Anthropic

39:00 Forward deployed engineers?

43:40 How fast are LLMs improving?

47:10 AI-led small business explosion

We recorded this conversation the day after Jake appeared with President Trump at a Kentucky rally. He explains why he’s speaking out on politics, even if it costs him financially, and why authenticity is key to cultural influence. Next, we explore the origins of Anti Fund, and how Jake and Geoff had the conviction to back Anduril before defense tech was in vogue. They also invested in OpenAI and were instrumental in rolling out Sora; learn about the strategy behind the launch and the challenges around AI slop and identity authentication. Additionally, Jake is the co-founder of Betr, and we discuss its recent partnership with Polymarket, what’s next for prediction markets, and how to build guardrails for customers. Finally, we explore Jake’s success in pioneering the creator economy and building consumer brands. In the 1960s and 70s, most kids aspired to become astronauts. Today, they want to be content creators, in part because of Jake’s influence. We conclude with Jake’s advice for young people in the AI era, and how we can persuade the next generation to love their country and embrace a builder mindset.

Subscribe now

00:00 Episode intro

02:20 On stage with President Trump

06:00 Attention is the most valuable currency

08:00 The origins of Anti Fund and backing OpenAI, Anduril, Ramp, and others

09:10 Betr + Polymarket and the future of prediction markets

14:58 Launching Sora with OpenAI

16:14 Are you worried about AI slop?

19:10 How AI is changing culture; revival of in-person events

21:20 Who should become a creator?

24:30 Fighters as cultural icons

31:30 How to win the culture

We discuss these timely topics and more with Administrator Loeffler. Growing up on a farm in rural Illinois, Kelly raised cattle and was exposed to the challenges of small business early on. She became the first in her family to graduate college and built a distinguished career in financial services at Intercontinental Exchange. In 2018, she became the founding CEO and first employee of the fintech startup, Bakkt, which brought institutional infrastructure to crypto. Kelly was appointed to serve in the U.S. Senate in 2020 and later confirmed as Administrator of the Small Business Administration in February 2025.

We begin our conversation with Kelly's entrepreneurial journey, before diving into her mission at the SBA. Her leadership comes at a pivotal juncture: a $10 trillion generational wealth transfer amid a technological revolution. As Baby Boomers retire and new AI tools emerge, Kelly believes young talent should think twice about traditional pathways and consider small business ownership instead. She lays out the administration's agenda, including expanding SBA loans for reindustrialization and bolstering the manufacturing jobs of the future. Next, we dive into her other core responsibility: exposing fraud. Learn how the SBA uncovered over $400 million of COVID-era fraud in Minnesota, $9 billion in California, and they're just getting started. She's also investigating 8(a) minority contracting — a program for disadvantaged communities hijacked by the Biden administration for race-based contracting and grift. The SBA is bringing in top talent and technology, including Palantir, to expose the fraud, revamp the programs, and protect taxpayers.

Subscribe now

00:00 Episode intro

01:40 Farming to finance

04:46 Building Bakkt & institutional infrastructure for crypto

09:00 AI & the future of small business

13:58 How to supercharge manufacturing in America

16:15 Exposing billions in fraud

18:50 The scandal of minority contracting

20:25 Using Palantir to investigate fraud

27:15 The NGO grift scheme

29:50 Optimism for the future

Subscribe now

00:00 Episode intro

01:30 Top Gun as a service

07:40 Founding a billion-dollar payments company

09:30 Pioneering private spaceflight

12:00 What happened to NASA?

14:40 Getting back to the Moon by 2028

20:47 Jared announces new plans for Artemis

25:15 Future of the space economy

27:50 Nuclear power in space

29:00 Sci-fi futures becoming reality

Subscribe now

00:00 Episode intro

01:28 JD's entrepreneurial journey

05:18 Wins and challenges at Opendoor / meme stock moment

09:55 WithCoverage and upending insurance brokerage with AI

13:11 Mimetic frameworks for entrepreneurs

15:45 How AI is changing the way companies are built

20:21 The impact of AI on existing SaaS

26:50 How to stop AI spam and imposters

31:27 How should young people prepare for the AI era?

34:20 The new invention era / why AI levels the playing field

The most basic instinct of an American entrepreneur is that when something is broken, rather than complaining, you need to build something new. And there is a lot that is broken, like our media. In early 2024, I started talking with a younger colleague, Max Meyer, about some ideas to fight back…

Out of those discussions, Arena Magazine was born: a bold voice for American technologists, capitalists, and patriots in the form of a high-end print magazine. Yesterday, Arena announced its seventh issue. It now has thousands of subscribers and is sold in Barnes & Noble throughout the US. Max told me that the latest print run of magazines used over 10,000 pounds of paper.

I’m very proud to be a supporter of Arena, and I encourage you to be, as well. You can subscribe for your home or your office.

At a time when the legacy media is full of negativity and ugliness, Arena is daring to be both optimistic and beautiful at the same time. Technology and entrepreneurship are so important to the success of the USA in the 21st century, and our mainstream media is totally poisoned against them.

If you’re anything like me, you remember the great magazines we had a generation ago in the technology space. Scientific American, Popular Mechanics, and even Wired used to be awesome. They used to celebrate Americans who were trying to do hard things. To say that they have gone in another direction would be a serious understatement.

Restoring that legacy — bold optimism, celebrating the builders — is the mission of Arena. It’s why I and so many other leaders in the innovation world are proud supporters. We need great media institutions again.

If you want to check out Arena, start with some of the recent articles. Here are two that I enjoyed recently, including from my friend Judge Glock on the wild history of Wall Street in the 1980s, or another about the last 250 years of American technological history. Arena has a new free Substack newsletter (link below). And if you want to get the print magazines, subscribe here.

Arena Magazine on Substack

The Great Eighties Wall Street Bonanza

“It was a vast space, perhaps sixty by eighty feet, but with the same eight-foot ceiling bearing down on your head. It was an oppressive space with a ferocious glare, writhing silhouettes, and the roar…The writhing silhouettes were the arms and torsos of young men, few of them older than forty. They had their suit jackets off. They were moving about in …

Read more

2 months ago

The progress has been inspiring over the last two years, and as more people wake up to the need for optimistic, beautiful media, Arena is going to continue to grow and inspire patriotic Americans.

Joe

Subscribe to Arena Magazine

Alex Kolicich is a Partner at 8VC, where he specializes in IT and Bio-IT investments. Prior, he was a Partner at Formation 8 and worked alongside Peter Thiel at Mithril Capital Management. Alex was also an engineer and early-product advisor at Clarium, Palantir, and Google. Jack Moshkovich began his career at 8VC and helps lead our investment practice alongside Alex out of San Francisco, where he focuses on enterprise software.

We begin with Silicon Valley’s most recent technological wave: smart enterprise. In 2013, we published our smart enterprise thesis, predicting the B2B software explosion that drove tremendous value creation over the past decade. Now, we’re entering the next landmark shift: the AI wave. Learn why AI is fundamentally disruptive (you can find our thesis here), and the enormous leap from data retrieval to automated decision making and from linear to non-linear problem solving. Next, we dive into the six layers of the AI investment stack, and why we are focused on level five: the application layer. We dissect the tremendous opportunities for AI-native upstarts, as well as the challenges — and upside — for smart enterprise pioneers that can reorient around AI. Finally, we discuss some of the key questions in the years ahead: When and where will we see massive productivity gains in the economy? Will LLMs become commodities, or is each frontier model sufficiently differentiated? And are we entering a bubble, or is the AI wave just getting started?

Subscribe now

00:00 Episode intro

01:40 Smart enterprise thesis

04:00 Did enterprise software dramatically boost productivity?

08:30 Understanding the AI wave

11:45 Are LLMs becoming commodities?

17:10 OpenAI vs Anthropic vs Meta vs Grok

20:30 Frontier models vs enterprise incumbents

25:20 What did Palantir get right?

32:30 AI case studies/new possibilities

43:00 The most surprising aspect of the AI wave

46:10 Are we entering a bubble? Or just getting started?

52:40 Biggest concern with AI

58:00 Biggest reason for optimism

Kevin began his career in the U.S. Air Force, defending the U.S. military from some of the earliest cyber attacks. He went on to found and lead Mandiant, scaling it into an industry leader which Google acquired in 2022 for over $5 billion. In 2013, he famously exposed China’s massive state-sponsored cyber espionage inside America. Kevin has also served on the National Security Telecommunications Advisory Committee (NSTAC) and the CISA Cybersecurity Advisory Committee. He recently launched Armadin to address new AI possibilities in the cyber domain.

To begin, Kevin takes us behind the scenes of the early cyber skirmishes in the 1990s, and what it was like confronting some of the first cyber incursions from Russia and China, as well as the rise of criminal cyber enterprises. In the early 2000s, Kevin noticed glaring gaps in conventional antivirus approaches and launched Mandiant. Learn how he quickly became the first person that CEOs and governments would call if they’d been hacked. After scaling Mandiant and exiting to Google, Kevin is now back in the arena building cutting-edge solutions to new AI cyber threats. He explains the dramatic increase in the productivity and sophistication of AI agents, and why it’s upending how we build cyber defenses. Find out what he’s most worried about in 2026, and what steps governments and CEOs need to take to secure their networks and infrastructure.

Subscribe now

0:00 Episode intro

01:30 Kevin's journey into cyber; early attacks from Russia & China

05:00 The evolution of cyber espionage and crime

08:00 Founding Mandiant; why Fortune 500 CEOs call Kevin

13:30 What's it like getting hacked by Russia?

21:30 AI hackers are here; Kevin's warning

27:00 Protecting critical infrastructure; harnessing AI for cyber defense

31:30 Engineers vs AI agents

36:40 Optimism for the future

This week, we’re honored to feature a conversation with Paul Atkins, Chairman of the Securities and Exchange Commission. He also served as a commissioner at the SEC from 2002 to 2008, and worked on the staff of two SEC chairmen from 1990 to 1994. Prior to returning to the SEC in 2025, Chairman Atkins was the founder and chief executive of Patomak Global Partners, a financial services firm he founded in 2009.

We begin with Chairman Atkins’ visit to Texas, the significance of Texas’ new business courts, and the importance of innovation in our financial markets. Whereas his predecessor, Gary Gensler, took an adversarial, regulation-by-enforcement approach, he’s embracing new technologies and regulatory clarity. In this vein, the Chairman lays out his strategy for making IPOs great again. Learn why the number of public companies has dropped by nearly half since 2007, and how Chairman Atkins is reforming the rules around litigation and corporate governance to deter activists from harassing America’s builders. We also discuss President Trump’s vision for making America the crypto capital of the world, and what it looks like to provide clear guidelines that promote innovation and protect investors. Finally, we explore the steps he’s taking to enable retail investors to access private markets and other alternative assets.

Subscribe now

0:00 Episode intro

1:35 Texas' new business courts / evolution of U.S. markets

7:20 The SEC's new approach to crypto

16:00 Making IPOs and public companies great again

25:00 Weaponization of shareholder proposals and ESG

32:10 National security concerns / China's bad actions

36:30 New rules for retail investors & private companies

44:35 Optimism for America's future

As co-founder of Airbnb, Joe turned a living-room experiment into a multi-billion dollar global brand. Then, he launched Samara, helping homeowners reimagine their yards with state-of-the-art Accessory Dwelling Units. He also sits on the boards of Airbnb, Airbnb.org, and Tesla. Last February, he joined Elon Musk and the DOGE team in D.C. to overhaul retirement for federal workers, before setting his sights on modernizing the nation’s 27,000 .gov websites.

We begin with Joe’s journey to D.C., and how a former Democrat came to support President Trump and take on problems that have plagued government for decades. His first challenge: federal retirement and “the mine” — a storage facility inside a Pennsylvania mountain that houses all federal employees’ retirement paperwork. Learn how Joe worked alongside career civil servants to transform a broken, six-month slog into a digital-only process that takes mere days. Now, he’s set his sights even higher: overhauling the government’s digital portals and delivering an Airbnb-like experience, starting with the Trump Gold Card, TrumpRX.gov, and, more recently, RealFood.gov and TrumpAccounts.gov. We go inside the Oval Office and Joe reveals his conversations with Pres. Trump on the importance of aesthetics, functionality, and design. Drawing inspiration from President Nixon’s beautification initiative, he worked with Pres. Trump to launch the government’s first digital makeover. It’s inspiring to see Joe’s leadership, and if you, or someone you know, wants to join his elite team of designers and engineers, there are a few spots remaining.

Subscribe now

00:00 Episode intro

01:15 Lifelong Democrat to Trump White House

07:50 Text from Elon / Joining DOGE

10:20 “The Mine” & fixing federal retirement

13:50 “What if we did it in a week?”

21:45 What happened to DOGE?

23:30 Becoming America’s Chief Design Officer

26:45 Inside the Oval Office with Pres. Trump

31:00 Fixing 27,000 .gov websites

34:00 Making America beautiful & functional again

We’re honored to go inside the Pentagon for a special episode with Emil Michael, Chief Technology Officer and Under Secretary of War for Research & Engineering. He oversees the Defense Advanced Research Projects Agency (DARPA), the Office of Strategic Capital, and other critical organizations. Prior, he helped scale multiple businesses, including D-Wave Systems, Tellme Networks, and Uber Technologies. Earlier in his career, he also served as a White House Fellow and later on the Defense Business Board.

We begin with Emil’s decision to join the Trump administration, and why he believes this is a generational opportunity to transform the culture and pace of innovation inside the Pentagon. Learn how he rolled out GenAI.mil (Google’s Gemini) to the entire force — roughly three million people — in a matter of months. Next, we dive into his decision to pare down the all-important Critical Technologies List, which drives the Pentagon’s R&D spending, from 14 areas to six. Then, we discuss the Department’s shift to a “commercial first” approach and what that means for the defense industrial base. Finally, we cover his priorities for the Office of Strategic Capital, plus how he’s thinking about talent and encouraging our best and brightest to step up and serve the country.

[NOTE: This episode was recorded prior to the United States capture and extradition of Nicolas Maduro]

Subscribe now

00:00 Episode intro

01:40 Time to disrupt the Pentagon

04:20 What did you learn from your first threat briefing?

06:30 Can you bring Uber speed to Dept of War?

08:45 Redefining the Critical Technologies List

13:30 Hypersonics and Golden Dome

18:00 Commercial first approach and changing the culture

23:00 How to recruit the best talent into government

28:45 Optimism for the future

Subscribe now

00:00 Episode intro

01:34 MIT engineer turned governor

06:35 Fighting special interests

12:50 Lessons from DOGE

16:37 Forcing a Balanced Budget Amendment

26:00 AI and welders vs lawyers

29:00 Making a Balanced Budget Amendment a reality

We’re joined by Scott for a timely conversation about Tech Force, his role as Director of the Office of Personnel Management, and key civil service reforms. Previously, Scott built one of the leading venture capital firms as Managing Partner of Andreessen Horowitz. He also penned “Secrets of Sand Hill Road,” a definitive guide for raising capital and navigating the startup world.

We begin our conversation with the importance of Tech Force — a two-year program designed to surge top programmers and designers into Washington, D.C. and create an enduring talent pipeline between the public and private sectors. Next, we explore Scott’s career in Silicon Valley, how he helped build a16z from the ground up, and the lessons he’s bringing to the White House. Then, we dive into his ambitious OPM agenda, starting with addressing rampant performance inflation, improving incentives for top performers, and removing artificial barriers, like years of service or college degrees, so we can get the best talent into the right jobs. Finally, we revisit the Pendleton Act of 1883 and how to bring merit back into civil service. Learn how Scott overturned a consent decree that will allow him to bring aptitude tests back into federal hiring — this could be one of the most consequential achievements of the Trump entire presidency!

Subscribe now

Philip is the Founder of the Common Good, best-selling author, and one of the leaders in government and legal reform. His book “The Death of Common Sense” became a powerful force for bipartisan reforms in the 1990s between President Bill Clinton and Congressional Republicans. I’ve followed his work for many years, and his newest book, “Saving Can-Do,” offers important frameworks for injecting accountability and human judgement back into governance.

We begin with the genesis of the red tape state, and why Philip believes the collective guilt of the 1960s led to well-intentioned but ill-designed policies that broke the government. He explains how law began to supplant human judgment, politicians stopped making hard decisions, and governance was outsourced to an instruction manual run by the professional class. Case in point: there are now 150 million words of binding federal rules, most added post-1970. The U.S Constitution, by contrast, is 7,500 words. Next, we dive into his new book, beginning with what makes the American “can-do” attitude unique. Then, he offers solutions for reinstating human authority into our institutions. And finally, a new framework for enabling America to build again. Philip makes the compelling case that what our republic needs most of all is a return to accountability over compliance, amateurs over professionals, and liberty over safetyism.

Subscribe now

Today we released the most important white paper we’ll publish at 8VC this year, authored by my colleagues Sebastian and Finn.

There are few topics I’m more obsessed with than how we can make healthcare better and more affordable for all. (Building inexpensive tunnels with Elon’s Boring Company to get rid of all traffic around cities and create economic access might be the only other one; but if I’m being honest, fixing healthcare is more important for preserving our civilization)

When it comes to AI, it’s fun to talk about everybody directing their own movies and symphonies, cool new defense applications, advances in science and cures for disease, or even personalized education or coaching that makes us all better. But there’s an elephant in the room.

Our country has made over **$100 trillion** in entitlement promises to provide healthcare to a massive aging population, and when you do the math, we simply can’t afford it — not even if we take away all the money from all the billionaires (ahem). The truth is that our government is going to go bankrupt as these costs continue to compound, if we don’t “bend the curve”. And the only way to do that is by getting productivity up. Way up.

AI is the miracle that solves this existential problem for our civilization. It also makes healthcare better and way more convenient, for everyone. How? What do we have to change? What concepts do we need to approach and discuss and debate the path ahead?

ASSISTIVE, SUPERVISED AUTONOMOUS, AUTONOMOUS...

I don’t want to steal Sebastian and Finn’s thunder, but this is a topic that will require strong leadership against a bevy of special interests who don’t want innovation or competition to disrupt their gravy train. The concepts below teach us what this will look like. It’s critical that we succeed!

The whitepaper is a little too long to show up in full in your inbox, so you can read the first half below - for the full thing, read it on the web version.

- Joe

A Vision for Healthcare AI in America

By Sebastian Caliri and Finn Kennedy

In a few short years, AI has revolutionized software engineering, writing, and entertainment. Waymo’s autonomous fleet now captures over 27% of rideshare trips in San Francisco, and the number of self-driving Teslas on American streets is multiplying daily. Every aspect of life, even the DMV, seems to be changing rapidly with the introduction of AI.

But not the doctor’s office. The only thing changing there is your copayment, and not in the right direction. Is the technology ready? It has now been more than three years since humanity first turned sand and $0.005 cents worth of electrons into medical expertise capable of passing USMLE Step 1 – so the answer is yes. What gives?

Why should I care about healthcare AI?

“Why should I even care?” you reply. “I don’t want big tech oligarchs looking at my medical records anyway.”

First, the proliferation of excellent open source models means that AI can transform care without Sam Altman getting anywhere near your data. If that wasn’t enough, the robust patient protections embedded in HIPAA have existed since 1996 and are not going anywhere.

With that out of the way, maybe you should not care about healthcare AI if you are among the 1% wealthiest Americans that can afford concierge medicine. Things work fairly well for you already. But everyone else, especially the working class, has much to gain from a future where entrepreneurs, doctors, and technologists join forces to bring AI to our healthcare system.

1. Health insurance is making wage growth impossible

Do you know the owner of a small business? Call them up and ask how much more they are paying for health insurance in 2026 compared to 2025. They may not be able to accurately describe how much worse the benefits got as they became on average 7% more expensive, but they definitely know the 7% (or, anecdotally, 16%) is forcing tough budgetary decisions.

Consider that every employee at a firm earns a different wage, but that health insurance premiums are the same if you make $40,000 or $400,000. Consider also that at 7% growth, 2026 is a bad but not particularly anomalous yearfor insurance prices. Over time health insurance has become a bigger and bigger component of working class Americans’ pay; what was once 12% is now closer to 30% of total compensation.

If the price of care drives premium growth, is there any technology you can think of that could impact those prices and therefore premiums? Are there any low cost substitutes to visiting the doctor you could imagine?

Marc Andreessen famously said “software is eating the world,” but for a lot of Americans the thing that is eating the world outside of working two jobs is healthcare costs. Healthcare AI is one of the few tools we have to funnel dollars from bloated insurance companies back to the pockets of blue collar workers.

2. It is a nuisance to seek and receive care

There have probably been times in your life when you’ve gone through great pains to schedule a doctor’s appointment. When the day arrives, you leave work, drive across the city to the office, sit in the waiting room for 20 minutes, only to answer four or five questions and hear “sounds good, keep on taking the medication”. You turn around to leave: “sir, with the deductible your payment will be $225. And before you head out, let me verify your insurance information…” Couldn’t this all have been a short phone call? And why did so little cost so much?

Routine tasks like post-hospitalization follow ups, medication adjustments, and chronic disease management are an excellent match for the capabilities AI has today. If we choose to make it possible, a quick chat with medical AI could replace low-complexity consultations and return missed hours of work (and income) to you. Instead of spending time on the phone with the UCSF scheduling office listening to hold music, you could be on the phone with your friends and loved ones.

This is to say nothing of ballooning wait times for a doctor’s appointment in the first place: 31 days in 2025, up from 21 days just a decade ago. At minimum, AI shortens the waiting window to discuss a concerning bump, rash, or medication that makes your heart race. In some cases accessing AI care immediately could be the difference between life and death.

3. Liberty and dignity are at stake

There are many doctors who take great care to explain their thinking to patients and involve them meaningfully in decision making. The chances you have met one such doctor on a day they aren’t overscheduled is similar to the odds of winning the lottery. The inevitable result is an unequal dynamic where patients become accustomed to nonparticipation; medicine conspires to make patients spectators the second they set foot in a hospital ward.

Physicians have real expertise that non-physicians do not, but is it possible to imagine average Americans gaining an understanding of their health, the diseases they suffer, and the treatment options in front of them? Without a lot more physicians with a lot more time to spend educating and explaining, no. But with the broad diffusion of AI, yes.

One of the great possibilities of healthcare AI is to let ordinary Americans ask questions before an appointment starts and after it ends, receive answers in language they understand, and put them back in the driver’s seat in matters of their health.

4. Doctors deserve better

When only 37% of American physicians would recommend their own profession to their children, it is fair to say that something is rotten in the state of medicine. Is it prestige or pay? The former is in the eye of the beholder, but the latter can be assessed empirically. We know doctors are taking on perhaps $250,000 in debt to begin earning a salary in their thirties, but just how bad is it out there? If we think about medical education as investment and salaries as returns, some quick math suggests internal medicine is not much better than putting money in the stock market instead. Sadly, the stock market gets closer to winning every year.

This financial argument comes on top of well-known reports of physician burnout, fatigue with compliance, and frustrations with documentation requirements. Medicine takes some of the very best people in society, saddles them with years of debt, only to pay them salaries that don’t keep up with inflation while using software that barely works.

We need to find a way to pay physicians more. Paying physicians more requires growing productivity. The non-insulting way of doing this is not cramming more RVUs into doctors’ schedules but rather giving them an AI iron man suit to do more with less. New AI tools like scribes are already showing promise reducing the drudgery of button clicks and making the day-to-day grind a little less painful.

5. Millennials and Gen Z should have a shot at a good life

What do you call a system where early participants get far more out than they ever put in, where today’s payouts come straight from the pockets of new entrants, and where everything collapses the moment the base of contributors stops growing? A Ponzi scheme, right? Actually we just described Medicare. Another and somehow more charitable way to describe it would be a deeply unfair intergenerational transfer program from Gen Z and Millennials to Baby Boomers.

Baby Boomers enjoyed a system where there were four workers for every retiree to support; today that figure is well on its way to two workers per retiree. Those two workers are supporting Medicare not just with their taxes but also with the massive, growing federal debt that they will eventually need to pay down once the Boomers are long gone.

Healthcare AI is currently illegal

Whether the benefits of healthcare AI resonate with you or not, we will now return to the motivating question: why hasn’t AI done much for medicine yet?

The reason is that healthcare AI is currently banned de facto or de jure in the United States. By and large, healthcare AI can’t receive payments from insurance which makes it a losing economic prospect. Even if payments existed, all fifty states have medical practice acts that would likely prohibit AI from prescribing or diagnosing: offending H100s would be ripped out of us-east-1 and locked up in Manassas Regional Adult Detention Center for the illegal practice of medicine. At the federal level the Social Security Act puts further limitations on any healthcare practitioner not licensed by individual states. And so on: the restrictions are manyfold and located across a web of federal regulations, state law, and federal statute.

The silver lining is that we have chosen to keep AI out of healthcare but could build it if we wanted to. What would it look like to choose differently?

A vision for healthcare AI in America

“Healthcare AI” today can be more of a Rorschach test than a helpful and specific term. Everyone has opinions on it but the semantics are all over the place. Therefore we will start by defining terms.

Healthcare AI refers to AI that supports care delivery (in e.g. hospitals, clinics, or at home) or is a substitute for those services. AI used for payer applications, pharmaceuticals, and biomedical research is outside our scope – but if it relates to doctoring it’s healthcare AI our book. Inspired by the six levels of self-driving that provide a common taxonomy for regulators and industry in autonomous vehicles, healthcare AI can similarly be stratified into four distinct levels:

Level 0: Administrative
AI that supports healthcare providers in back office or administrative tasks.
(E.g. Scheduling voice agents, AI scribes)

Level 1: Assistive
AI that assists clinicians but does not diagnose, treat, triage, or prescribe medications to patients.
(E.g. AI coaches, advocates, and navigators)

Level 2: Supervised Autonomous
AI that diagnoses, treats, triages, and/or prescribes medications to patients, with all or a subset of decisions monitored by a supervising clinician.
(E.g. AI medication management for chronic disease with physician oversight)

Level 3: Autonomous
AI that autonomously diagnoses, treats, triages, and/or prescribes medications to patients.
(E.g. Fully-autonomous AI emergency triage line)

Steps towards a golden age of American healthcare

Currently, solutions are being developed and deployed at level 0. In a few short years of existence, AI scribes like Commure, revenue cycle solutions like Candid, and patient scheduling agents like Hello Patient are already capturing over $1B in annual revenue. These administrative solutions are creating real value for healthcare providers. But they are not going to steer the Titanic that is American healthcare away from the iceberg.

Levels 1, 2, and 3 are the steps forward that can transform patient experience, costs, and ultimately health outcomes; these are also the types of healthcare AI that are variously banned today. It is important that each level is not only legally permissible but also investable to build and proliferate new AI systems.

We envision a future where we move from level 1 to level 3 through a series of policy changes that build upon each other and gradually widen the aperture for innovation in our healthcare system. Such an approach is more practical than a revolution, and also allows us to learn from each advance forward and accelerate or apply the brakes as we collectively see fit. Here is how it could work.

Level 1: Assistive

What is it?

Imagine being able to ask your doctor a question at any time of day via text, phone call, or video, and immediately getting a response. “Am I supposed to feel a bit nauseous with this new medication?” Or, “if I have to avoid potassium in my diet, remind me if I can still have a banana in the morning? What are some alternative ideas?” Or even, “watch me make this movement – should I keep wearing the knee brace or is it time to take it off? And can I do the hike with my friends on Sunday?”

Imagine no more: level 1 healthcare AI, assistive technology, could give you a personalized response to each of these based on your medical history, biometric data, and your doctor’s preferences. Level 1 cannot write a prescription or make a referral – those types of scenarios would still need escalation – but it could certainly make your life easier with advice, coaching, and education that wraps around the care you already receive.

How do we get there?

Level 1 healthcare AI exists today albeit under limited circumstances. Digital health companies like Omada, Superpower, and Citizen deploy assistive agents already as health coaches and advocates that help inform patients and supplement clinicians.

The main blocker for assistive AI care models is that no broadly billable CPT code matches the services they render. The companies mentioned above rely on alternatives like bespoke case rates, PMPM fees, or they just skip insurance reimbursement all together in favor of cash payments. Investors theoretically could fund more new businesses in the category but the uncertainty of ever being able to get paid by insurance is a massive impediment to capital efficiency and therefore investment activity.

There are also a few state issues to overcome as overzealous state legislatures have scrambled to “do something” about AI over the past year. And the HTI-1 Final Rule, published in 2024 under the Biden administration, could subject some of the assistive AI described here to questionable certification requirements as a de facto condition for market entry.

1.A. Implement pay-for-value reimbursement for assistive AI care models [CMMI]

Care models using assistive AI cannot currently receive insurance reimbursement because those services do not match any AMA-defined CPT codes. In fact, assistive AI could actually reduce provider revenues if applied towards services that bill time-based codes. Assistive AI needs private sector investment, and that investment will not be forthcoming if the only possible business model is charity.

Our previous writing on the topic of AI reimbursement describes a CMMI model for making value- or outcome-based payments to AI-enabled services that could result in durable codes, case rates, or monthly fees, readily adoptable by private payers. The more transparent the rate determination methodology, the stronger and clearer the signal to capital allocators and the more that resources can flow to the space.

Solving reimbursement takes level 1 healthcare AI from a paradigm of building technology followed by a multiyear lobbying effort to be paid by insurance companies to a world where valuable innovation can be very predictably rewarded.

1.B. Pre-empt heterogenous state restrictions and disclosure requirements [Federal Law]

Rather than supporting healthcare AI, many state governments have begun adding various restrictions to hobble it. Some examples of recent state laws include:

• California: mandates that every message that touches “patient clinical information” and was drafted by generative AI must carry a conspicuous AI notice at the top of emails, continuously in chat, and with clear instructions for how to reach a human clinician.

• Nevada: prohibits anything resembling AI mental-health care. If your product had a CBT coach, it is now disabled and if your marketing copy for your AI system implies “therapy” or “counseling” you face penalties.

• Texas: recently introduced two new laws that require patient disclosure when AI is used in healthcare, as well as a data localization rule that forbids physically offshoring medical records.

These heterogeneous restrictions imposed by individual states turn one nation-wide market into fifty distinct markets each with unique requirements for products to be introduced. This increases the cost of innovation for healthcare AI not just at level 1 but across all levels. If you need supporting evidence just compare the venture capital flowing into Medicare and employer solutions to the money going towards Medicaid innovation.

In July 2025, President Trump called for “one common-sense federal standard [for AI] that supersedes all states”. In the context of healthcare AI, federal law should clarify common disclosure and data storage requirements and prohibit wholesale bans on particular uses of AI to protect professional cartels. Other regulatory areas important to individual states (e.g. patient age verification) should not be dismissed but still ought to be preferentially addressed at the federal level.

1.C. Refocus HTI-1 certification on outcomes, not inputs [ASTP/ONC]

When you use ChatGPT, how often do you find yourself asking, “that was a great and useful response, but I wonder if the training dataset used Common Crawl from April 2025 or November 2024?”. These questions you might ask if you are gathering competitive intelligence on OpenAI’s state of the art models but not what users care about.

Under the Biden administration, ASTP/ONC promulgated the HTI-1 rule that would require some assistive AI systems to document their models’ training data provenance, update policies, evaluation methods and meet a long list of other compliance measures for official certification — in turn a de facto requirement for market entry. Compliance here encumbers small startups in favor of big tech and the large AI labs. Depending upon the details surfaced to EHR vendors and end users, there is also risk of trade secrets being exposed to competitors.

HTI-1’s predictive DSI requirements should be narrowed to disclosures on what providers and patients actually care about and what regulators should care about: how well they work in the real world, not how they were built.

Moreover, there should not be one single authority on level 1 system evaluation. HTI-1 should define de minimis “table stakes” disclosures, and a number of valid competing approaches should emerge bottom up to protect against regulatory capture.

Level 2: Supervised Autonomous

What is it?

Level 2 builds on level 1 by adding the ability to legally practice medicine: diagnosing, prescribing, and referring just like a doctor. Actually, more like a nurse practitioner or physician assistant: level 2 systems would be deployed to patients and overseen by a supervising clinician. All AI activities would be auditable and all or a flagged subset of medical decisions (e.g. unusual prescription changes, a decision flagged by AI for uncertainty, or anything flagged at the patient’s own request) could be subject to supervisor review before actioning.

Imagine you are hospitalized for congestive heart failure and prescribed a whole pharmacopeia of medications right before discharge. Some of the dosing would be right and some would need adjustment. Rather than coming back every few weeks (at a time no doubt incompatible with picking your son up from soccer) for changes, suffering from discomfort in the meantime, what if you used a home blood pressure cuff, AI chat, and mail order pharmacy to deal with it all on your own time? If you ever suspected something was off, or if AI suggested an usual recommendation, your doctor would get a notification to review the prescription first.

Young parents desperate for something to make life easier would also find relief in level 2 systems. Is it a viral infection or is it probably bacterial? Or did junior just somehow get a Lego into his ear? Forget about rushing to urgent care to find out. Your pediatrician has already given you access to an AI tool that looks at images of tongues, mouths, and little ears, and automatically writes a prescription for antibiotics if needed, and counsels other remedies if not.

How do we get there?

All autonomous AI is considered Software as a Medical Device (SaMD) by the FDA. The FDA label would determine what a level 2 solution can and can’t do – prescriptions, diagnoses, and so forth. We can think about a level 2 AI’s “scope of practice” as defined by its FDA approval.

The FDA evaluation process for AI systems is immature. Reviews often take too long, standards for approval are often too high (or in some cases, inappropriately low), and that only about 2% of approved SaMD has an authorized PCCP tells us that something is amiss with how we accommodate updates and upgrades.

Much like level 1, insurance reimbursement is also a barrier. Technologists will not build it if no reimbursement will come. At this stage, services will be billed under a supervising provider’s National Provider ID (NPI) number but assigning an NPI to the AI system will allow for prescribing and ordering and set the stage for insurance billing at level 3.

Meanwhile, certain state laws will need to be changed as level 2 autonomous agents come online. Each state has a number of “practice acts” that define types of healthcare practitioners and the services they are legally allowed to provide. Unsurprisingly, AI agents are not considered legal practitioners in any state.

2.A.1. Align FDA approval benchmarks with real world standards, not hypothetical ideals [FDA]

One problem with our FDA evaluation process for autonomous AI is that the performance bar is often set unreasonably high. LumineticsCore is an illustrative case study: the FDA required the tool to catch at least 85% of diabetic retinopathy cases, but studies show board-certified ophthalmologists land closer to 77% on the high end, and possibly as low as 33%. Our current approach is like making everyone who wants a driver’s license match Lewis Hamilton’s track record at Monaco: sure, you’d have fewer crashes, but almost no one would get a license and many would not even try.

AI should be measured against care people actually receive instead of subspecialists most patients never see. Although performance would be no worse than an average doctor visit, risk would be further mitigated by default escalation of low-confidence judgements from AI to supervising clinicians.

In the case of diabetic retinopathy only about half of eligible Medicare patients get screened at all owing to the accessibility and cost of services. We are making AI outperform the best of the best when the benefit of increasing access to average care in many cases dramatically outweighs the risks.

2.A.2. Reform PCCPs to allow continuous model improvement [FDA]

PCCPs don’t work in practice. PCCPs as constructed today must be highly specific at filing, and the AI research landscape is simply progressing too quickly to anticipate what changes may come in future years. It is also challenging to predict opportunities for systems to be extended to new populations or input data (e.g., adding CT when you start on X-ray; analyzing images captured with higher fidelity from newer machines) years before it is deployed in the wild.

FDA should revisit PCCPs to allow developers to make improvements while still minimizing risks. PCCPs should pre-approve a short list of upgrades that can launch with notice and live monitoring, like retraining on new data, like-for-like model swaps, and added inputs within the same modality. Sponsors should be able to expand plan scope by notification when risk does not increase, and full resubmissions should be reserved for changes that materially change the risk profile or add indications.

2.A.3. Require standardized postapproval monitoring [FDA, ASTP/ONC]

If the FDA approval bar is inappropriately high in some cases, it may actually be too low for SaMD on the 510(k) track. In a JAMA Health Forum review of about 950 510(k)-cleared AI devices, 60 were later recalled and many within one year of launching. These recalls were concentrated in devices that were approved by showing functionalequivalence to other devices, but without submitting validation from actual use in the clinic.

Whether 510(k) approval should require some clinical validation or not is not a topic on which we will opine, but what should be uncontroversial is the need for rigorous and standardized postapproval monitoring. It should be technically possible to implement reporting through ASTP/ONC EHR certification requirements. Such a rule not only serves the purpose of catching cases where recall is appropriate as early as possible, but can also serve as a foundation for evolving the PCCP, allowing sponsors to test updates at small scale to support more frequent and less burdensome improvements over time.

2.A.4 Create an FDA Center for AI (CAI) [FDA, HHS]

The FDA’s CDRH was created in 1982 to review medical devices and all kinds of radiation-emitting devices whether medical (e.g. x-rays, CT scanners) or non-medical (e.g. airport x-rays, consumer devices like TVs and monitors). This suite of responsibilities was already motley in 1982 and adding responsibility for cutting-edge AI is simply a bridge too far.

Evaluating healthcare AI will require a great deal of creativity and talent that simply does not match up with the expertise at CDRH. FDA should create a new Center for AI whose sole responsibility is the evaluation of level 2 and 3 healthcare AI. Perhaps there would be no need to even suggest items 2.A, B and C if we had started this way.

It makes logical sense to have a team of software experts working with software innovators instead of forcing a square peg into a roundhole. This can be done with the authority already vested in the FDA commissioner and the HHS secretary.

2.B. Enact state Medical AI Practice Acts [State Law]

Building a functional level 2 system for medication titration or chronic disease management would be a miraculous achievement. But today it would be explicitly illegal in all 50 states because it contravenes state practice acts for AI to prescribe, treat, diagnose, and refer without an appropriate medical license.

The cleanest path forward would be for states to pass Medical AI Practice Acts that award a new class of medical licensure to FDA-approved healthcare AI. Approved SaMD could have a scope of practice aligned with its FDA label. An associated state Medical AI Board defined by the Practice Act can expand or contract scope as the label evolves, work with the FDA to monitor performance (in line with item 2.A.3.), and sanction AI companies that do not respond appropriately to inquiries and warning flags related to failure or degradation. States will ideally adopt uniform language for formation of Medical AI Boards to reduce the complexity of market entry for new models and create the strongest possible environment for spurring investment.

An alternative approach could be to bundle a federal AI Practice Act with the pre-emption of state disclosure laws described earlier (1.B.). A federal practice act would likely be seen as more intrusive on states’ rights than a national disclosure standard, but on the other hand would be a much faster path to making level 2 AI investable than fighting legislative battles in 50 distinct state capitals.

2.C. Implement provisional T-code payments to bridge FDA approval and reimbursement [CMS]

The FDA approval of a level 2 or 3 system may bless something as safe, but it may not provide CMS and other payers with the information required to determine a final reimbursement rate. But if we stack a costly and (possibly) unpredictable reimbursement timeline on top of a costly and (possibly) unpredictable FDA timeline, the uncertainty will make it hard to attract capital and talent for development.

After an FDA authorization, CMS should grant temporary T-codes with small provisional payments during this bridge period, billed using the NPI of the supervising clinician. This approach would be somewhat analogous to the New Technology Add-On Payment (NTAP) program for new devices. Reimbursement could be clawed back if prices were ultimately set lower than provisional rates or if a device’s approval was later revoked by the FDA.

Such a T-code system would kill two birds with one stone. It would provide cash flow for innovators, while also allowing CMS to gather evidence of value from real world deployment as an input to a reimbursement determination and code assignment.

2.D. Framework for autonomous AI coding and pricing determination [CMMI/CMS]

With a clear FDA process and provisional T codes to investigate real world performance, level 2 systems are off to the races. But how will reimbursement work in the long run and how will pricing be determined?

An approach could build upon the pathway described for assistive AI (1.A.) through (an) additional CMMI model(s) defined specifically for supervised autonomy. On pricing, others have written eloquently on autonomous AI reimbursement rates. How the details should work is a question for CMMI analysts and CMS actuaries. We only say that a clear process that results in a code, case rate, or other payment model that is easily billable by a supervising provider NPI and portable to private payers is what can facilitate development of the level 2 ecosystem.

2.E. Create NPIs for licensed AI providers [CMS]

FDA approval, state licensure, and insurance reimbursement still won’t quite unlock the potential of level 2 AI. There is one last administrative issue to resolve: transactions including e-prescriptions, referrals, and order placement require signing with a CMS-issued NPI as part of the data transmission standards defined by HIPAA.

CMS should create a new class of NPI for FDA-approved and licensed autonomous AI and require its use to be supported across all relevant senders and recipients of healthcare data: EHRs, e-prescribing, labs, payers, clearinghouses, etc. Ordering and referring rights are typically tagged to specific NPIs by payer systems and in this case could be inherited from the FDA label in the same way that state scope of practice can be.

While not used directly for billing at level 2, these AI system NPIs will unlock their usefulness beyond assistive systems and set the stage for direct billing at level 3. These AI NPIs will furthermore create a vector for assessing model behavior in the real world via inclusion in claims data.

Level 3: Autonomous

What is it?

Level 3 AI would have the ability to prescribe and diagnose, but unlike level 2 it would not need to be deployed by and directly monitored by a physician. Level 3 is in some sense a true “AI doctor”.

A world of level 3 autonomy looks like autonomous AI urgent care, available 24/7 without wait time, even in rural areas that lack other options to receive care. It looks like agents that persistently monitor biometrics across the medicine ward and catch signs of decompensation before a human possibly could, improving survival and reducing labor required to run a hospital. Perhaps there are behavioral or autoimmune conditions where AI can synthesize data and make better prescribing and dosing decisions than humans.

Routine and occasionally perfunctory work like refills for low risk drugs could be handled by AI too instead of wasting physician time. Is it really necessary to have all those doctors and nurses on the other side of the screen at Hims & Hers and Ro, clicking “yes” to prescribe your erectile dysfunction medication?

A human connection is essential to many aspects of medicine and it is unlikely that the most effective way to deploy level 3 systems will be wholesale substitution of your doctor in most cases. A better starting point to envision their role is asking “where are there places where autonomous systems can do things that humans alone cannot?” As well as, “where can human-computer symbiosis yield the strongest possible patient experience and patient outcomes?”

How do we get there?

Only a few policy changes are required to open the door to level 3 once level 2 is permissible. The biggest remaining issue is embedded in the Social Security Act, Title XVIII – the very source code for Medicare – namely, defining AI as a type of practitioner that can be eligible for reimbursement.

3.A. Expand state Medical AI Practice Acts [State Law]

Going from level 2 to level 3 requires an expanded scope of practice that can be assigned to AI, in particular changing requirements for supervision and oversight to permit full autonomy where indicated by the FDA. This would be enacted via changes to state Medical AI Practice Acts defined earlier (2.B.).

Uniformity of approach across states would once again create the best possible environment for investment – a single market is much better than fifty – but there is a case that states should be able to express different preferences. An FDA approval for level 3 autonomy could still be used in a supervised fashion (level 2) in some states without necessarily breaking the business model for innovators. Perhaps states may also want to license level 3 use for rural or underserved areas but only as a fall back when clinician supply cannot meet demand.

3.B. Amend the Social Security Act to allow Medicare payments to licensed AI [Federal Law]

Medicare was created in 1965 through amendments to the Social Security Act (SSA). Within the SSA various types of providers are enumerated as eligible for reimbursement by Medicare in section 1861. In 1965 this list included just licensed physicians; in 1977 Congress added physician assistants; in 1989 we added nurse practitioners. Today the list includes various types of social workers, therapists, and midwives. AI is not currently listed under section 1861. Even if CMS created a billing code for a level 3 autonomous service, it would be illegal for Medicare to make a payment directly to an AI company instead of a supervising physician.

Section 1861 of the SSA must be amended to allow Medicare to make payments to FDA-approved autonomous AI providers, using its own NPI, when acting within its label-defined scope of practice and licensed by the relevant state board.

Interestingly, Medicaid may not be subject to the same restrictions. Medicaid was also defined in 1965 as an amendment to the SSA and similarly enumerates what types of providers and services are eligible for payment in section 1905. The language leaves significantly more flexibility for interpretation by the HHS secretary to support payments to new types of licensed practitioners defined by individual states. It may be possible to run early pilots for level 3 autonomy in specific states to address high areas of need for Medicaid populations before seeking legislative support to serve and bill for Medicare beneficiaries.

The future of American healthcare will not be born without a fight

The one in six American adults that asks an AI chatbot medical questions each month can see a future that puts the world’s best specialists in your pocket. They can see a country that looks after working class and rural populations, dignifies Medicaid beneficiaries with access to care, takes the problem of physician burnout seriously, and gives Gen Z and Millennials hope for the future instead of saddling them with a soul-crushing amount of federal debt.

But powerful groups are working assiduously to make sure this future of abundance never comes to pass. Professional associations like the National Association of Social Workers, the cartel behind a ban of AI behavioral therapy in Illinois, will protect their monopoly at all costs. The American Medical Association similarly seeks to thwart AI systems while veiled behind a concern for safety. Big business hopes to capture healthcare AI regulation, crush any competition, and sequester all future profits for themselves.

Political factions are no better. Short-sighted socialists on the far left would rather sacrifice the future than allow private companies to profit from innovation. On the far right, a deep distrust of the technology industry has figures like Steve Bannon getting dangerously close to a call for Butlerian jihad.

We are on the side of the American people, patients, and physicians. We want safe and effective AI that makes healthcare accessible and excellent so that our country will flourish. Virtually all writing on healthcare AI is written by its critics and focused on the problems, risks, and restrictions that must be enacted. Almost none is focused on the benefits and the path to get there.

The American healthcare system causes all kinds of unnecessary suffering in obvious places like the hospital and in non-obvious places like your paycheck. AI is our best hope to alleviate that suffering. This plan has ambiguities and ideas that will be wrong when making contact with reality, but is my attempt to share a vision and a way to make that vision real.

Acknowledgements:

• Joe Lonsdale

• David Brailer

• Colin Rom

• Michael Abramoff

“Thoughts without content are empty, intuitions without concepts are blind.”

Immanuel Kant, Critique of Pure Reason

I have known Peter Thiel as a colleague, friend, and intellectual sparring partner for almost twenty-five years. Peter is not just one of the leading investors or entrepreneurs, but thinkers of our time. Over the last year, he has given a series of interviews and private lectures (some at our University of Austin) on the subject of the Biblical Antichrist, and possible existential threats to (and defenders of) our world.

If it were a random person — or even the average VC — deciding to wade into a subject like this, one could be forgiven for looking away from something that sounds wacky in our modern culture. But consider the source.

This is the person who made the first investment into Facebook, and wrote the first check from Silicon Valley to Donald Trump. In those two bets alone, Peter demonstrated an uncanny ability to focus on important ideas and people that will shape our civilization before others. Not to mention: he founded PayPal and led its team of future legends; together we co-founded Palantir; he was a very early (and key) backer of SpaceX at a time when no private space venture had ever succeeded. He predicted the hollowing-out of American universities. And he was a backer of JD Vance well before our now Vice President entered politics.

Given this record, when Peter focuses this much on something, one should keep an open mind. For me, as a student of history and philosophy like Peter, drawing on concepts and models from the texts that form the basis of our civilization is powerful way to access and think about complex topics, such as how a history-ending equivalent of the Antichrist might manifest as a one-world state of total “safety.”

Immanuel Kant, in his Critique of Pure Reason, explains the importance of unifying our intuitions with models and concepts. For Kant, we actually cannot think without them.

“Intuition and concepts … constitute the elements of all our cognition, so that neither concepts without intuition corresponding to them in some way nor intuition without concepts can yield a cognition. Thoughts without content are empty, intuitions without concepts are blind. It is, therefore, just as necessary to make the mind’s concepts sensible—that is, to add an object to them in intuition—as to make our intuitions understandable—that is, to bring them under concepts. These two powers, or capacities, cannot exchange their functions. The understanding can intuit nothing, the senses can think nothing. Only from their unification can cognition arise.”

Kant is a bit dense to some (okay, to me too usually), but think of it this way: if you want to build intuition for how to play a complex game, you have to form a hypothesis about how a good strategy might work, with a simple model in your head. Any guess is better than none. Then, play. Our minds evolved to iterate on conceptual models: you start with a model that’s wrong in various ways, and build intuition, improving it over time. Similarly, if you want to think and discuss hard concepts intelligently and come to new conclusions about not just games but business, investing, or even more abstract topics like how a threat analogous to the “Antichrist” might be generated or restrained, you have to build intuition by starting somewhere with relevant concepts.

For most people, going all the way back to some of the toughest Biblical texts like the Book of Daniel and the Book of Revelation and an idea like the Antichrist is just too difficult. And in an increasingly secular society, that makes Peter’s intellectual confidence to “go there” all the more worthy of our attention — whether you are secular or religious; a Jew, a Christian, or neither.

“The default political solution people have for all these existential risks is one-world governance,” Peter told the New York Times’ Ross Douthat, giving us a strong clue why he might have come up with these Antichrist lectures. Those risks might be AI, climate change, pandemics, nukes, or something else. In Peter’s telling, “The way the Antichrist would take over the world is you talk about Armageddon nonstop. You talk about existential risk nonstop, and this is what you need to regulate.”

In Christian thought, the only throne for the whole earth is that of Christ; and for us Jews, it is the covenants with God that were made so many generations ago. The Hebrew prophets and the authors of the New Testament—whose wisdom is the rock of our whole civilization—took the concept of Earth-consuming empires very, very seriously as violations of divine order, insofar as they try to substitute for either Christ or the Covenants. Genghis Khan, Alexander the Great, Adolf Hitler, and others wanted physical thrones over the earth. But it’s a fear that goes back much further, to the biblical texts from which the idea of an Antichrist arises.

In the Babylonian Captivity, the Jewish exile Daniel had a series of terrifying, apocalyptic dreams of four beasts that precede the end of times. These dreams are recounted in the Book of Daniel in the ketuvim (‘writings’), the final section of the Hebrew Bible. Medieval Jewish thinkers thought the four beasts might be allusions to Nebuchadnezzar, who besieged Jerusalem and under whom the Jews were captive in Babylon; Belshazzar, the last leader of that dynasty; Alexander the Great; and the Roman Empire. It is in this book that the idea of the beast as a one-world state originates. In the the final verses of Daniel 7, God speaks to Daniel about a beast (kingdom) that would “devour the whole earth” and then be conquered:

‘The fourth beast shall be
A fourth kingdom on earth,
Which shall be different from all other kingdoms,
And shall devour the whole earth,
Trample it and break it in pieces.
The ten horns are ten kings
Who shall arise from this kingdom.
And another shall rise after them;
He shall be different from the first ones,
And shall subdue three kings.
He shall speak pompous words against the Most High,
Shall persecute the saints of the Most High,
And shall intend to change times and law.
Then the saints shall be given into his hand
For a time and times and half a time.

‘But the court shall be seated,
And they shall take away his dominion,
To consume and destroy it forever.
Then the kingdom and dominion,
And the greatness of the kingdoms under the whole heaven,
Shall be given to the people, the saints of the Most High.
His kingdom is an everlasting kingdom,
And all dominions shall serve and obey Him.’

The Book of Revelation, the last book of the New Testament, had similar apocalyptic visions:

“And the whole earth was amazed and followed after the beast; they worshiped the dragon because he gave his authority to the beast; and they worshiped the beast, saying, ‘Who is like the beast, and who is able to wage war with him?’ A mouth was given to him speaking arrogant words and blasphemies, and authority to act for forty-two months was given to him. And he opened his mouth in blasphemies against God, to blaspheme His name and His tabernacle, that is, those who dwell in heaven.”

The word Antichrist (Ἀντίχριστος) means in place of, opposed to, or against the anointed one (Christ, χριστος). Katechon (κατέχον) is “that which withholds” or “the restrainer.” Both appear in the Apostle Paul’s Second Epistle to the Thessalonians. You will immediately recognize the similarity of Paul’s description of the “man of sin” to Daniel’s description of the evil king.

Now, brethren, concerning the coming of our Lord Jesus Christ and our gathering together to Him, we ask you, not to be soon shaken in mind or troubled, either by spirit or by word or by letter, as if from us, as though the day of Christ had come. Let no one deceive you by any means; for that Day will not come unless the falling away comes first, and the man of sin is revealed, the son of perdition, who opposes and exalts himself above all that is called God or that is worshiped, so that he sits as God in the temple of God, showing himself that he is God.

Do you not remember that when I was still with you I told you these things? And now you know what is restraining, that he may be revealed in his own time. For the mystery of lawlessness is already at work; onlyHe who now restrains will do so until He is taken out of the way. And then the lawless one will be revealed, whom the Lord will consume with the breath of His mouth and destroy with the brightness of His coming.

The scripture provides the basis for how people have understood concepts like the Antichrist. But one need not believe that a beast with ten horns and seven heads will physically arise from the sea, or even that the Antichrist is a single person. Christian thinkers have long debated these images. Peter leaves it up in the air what he believes. But contrary to the mockery from South Park or anyone else, Peter is not actually indulging apocalyptic fantasy, or weirdly obsessed with something irrelevant here. He is using his wisdom and wisdom from the ancients to help leaders build intuition about the biggest threats we face, and the dangers in how we confront them.

We must take freedom seriously without falling into the trap of believing we can protect freedom only by surrendering it to a global sovereign. One of Peter’s deepest and most unsettling points is that a katechontic bulwark—perhaps the one you and I still call “the last, best hope on Earth,” the United States of America—can itself become the Antichrist if it evolves into a homogenized world order. History is littered with empires that began as defenders of pluralism and ended as devourers of the earth.

The katechontic task is to confront those dangers head-on while simultaneously restraining the seductive argument that only a single planetary regime can save us. Indeed, a katechon itself — safetyism “saving” us from AI, or a one-world government “saving” us from war or climate change — can morph into an Antichrist if we are not careful.

Peter makes a lot of other points that it’s not my intention to cover here; and to be clear this is my take on his lectures, not his! It’s unclear to me if our culture is transcending or falling back past the age of the Enlightenment, as he also seems to hint, and if old superstitions, mimetic mobs, and other human instincts will play a bigger role in our society in the coming decades; but there is certainly wisdom in building intuition for what this could mean.

As a Jewish American, one of the most important aspects of our reality that I’ve come to understand today is how the reaction to Nazism and nationalism itself have shaped the last 80 years. On the one hand, it was absolutely necessary to do what the United States and our Allies did. But in reacting in horror at nationalism and embracing extreme forms of globalism after the war, some of our leaders and especially European leaders have created havoc in their countries. The reaction to their excessive globalism today is once again nationalism — some of which is clearly needed against the evils of extreme globalism, but which at the other extreme could threaten to echo the evils of the 1930s.

It would be impossible to go into detail on all these subjects, but what makes the models so useful is in testing our theories of present and future reality.

And as for Peter: giving us some of these models is the type of serious intellectual work that in previous ages would have been done behind the cloisters of universities and monasteries. And in doing that, Peter should earn immense credit. If you’re a leader interested in the future of humanity, I’d pay closer attention. Oh, and don’t immanentize the Katechon!

We hosted this conversation at the 2025 Cicero Institute Courage Awards, where we recognize bold policymakers on both sides of the aisle for standing up to special interests and fighting for liberty. Amid the regulatory battle over AI, we were honored to feature Michael, who currently serves as Director of the White House Office of Science and Technology Policy and a Senior Advisor to President Trump. In the first Trump term, he was the nation’s Chief Technology Officer and later acting Under Secretary of War for Research and Engineering. He started his career working for Peter Thiel, first at Clarium Capital and later Thiel Capital.

We began our conversation with Michael’s journey to the White House, and his role leading the White House’s science and technology policy agenda. He outlined three key areas to winning the AI race: the right regulatory frameworks, the necessary infrastructure, and international engagement. How do we not only stay ahead of China but also persuade other nations to adopt our AI stack? Michael lays out the strategy. He also makes the case for a federal AI standard and explains why a patchwork of laws stifles innovation, hamstringing the upstarts and entrenching the incumbents. Finally, he lays out what AI will mean for the American worker, and as the man in charge of overseeing AI, quantum, space, and more, he reveals what keeps him up at night. We have a generational opportunity with AI to bring down the cost of living and lift up millions of lives; we’re grateful to have leaders like Michael at the helm.

Subscribe now