Deep Thought

1. The Hugeness of AI

For one of the course's final classes, Thiel brings in three founders working on the frontier question of computer science: Scott Brown of Vicarious, Eric Jonas of Prior Knowledge, and Bob McGrew of Palantir. The topic is artificial intelligence — in Thiel's view the most important and least discussed frontier of all.

We instinctively place intelligence on a human scale running from a mouse to a moron to Einstein. But that whole scale is a tiny dot. Evolution explored only one corner of the design space; engineered minds need not resemble anything nature built, just as a supersonic jet is a bird with titanium wings that no bird could evolve into. A truly superhuman AI might be as unfathomable to us as general relativity is to a mouse — which pushes the discussion into almost theological territory: would such an AI be all-powerful, and would it care about us at all?

2. The Strangeness of AI

The old Turing-test question was whether machines can think; the newer expectation is that they should also relate to us emotionally. The practical question underneath is displacement: the plow, the printing press, and the cotton gin all displaced workers yet made society richer. Does AI follow that pattern, or break it?

• Luddite paradigm: machines destroy livelihoods, so smash them before they destroy you — the textile workers wrecking cotton mills.
• Ricardian paradigm: after economist David Ricardo — technology is a trading partner; displaced Detroit workers retrain, the production frontier expands, prices fall, and everyone captures gains from trade.

The Ricardian frame holds so long as AI is only somewhat better than people: comparative advantage creates a division of labor and everyone wins. But if AI becomes vastly superior at everything, the trade frame simply stops applying. Most technologies improve smoothly with no cliff; AI may be the exception — humans stay in control right up until the moment the system goes superhuman, and then lose control entirely.

Humans don't trade with monkeys or mice.

3. The Opportunity: AI vs. Biotech 2.0

Timing is brutally hard to judge in advance: supersonic passenger jets flopped in the '70s, handheld devices in the '90s, and even Siri was arguably still too early in 2012. At a Santa Clara event called 5 Top VCs, 10 Tech Trends, the audience agreed 100% that biology was becoming an information science, voted 92% against electric cars, and split 50-50 on Moore's Law accelerating — yet most assumed AI was much further away than biotech 2.0. Unanimity should make you suspicious; consensus is where returns go to die.

Both fields may harbor hidden limits. Biotech's dream of indefinite lifespan could collide with a built-in trade-off: unbounded cell division via telomerase looks a lot like cancer, so curing aging might feed the other great killer. AI's leading candidate limit is code complexity — past some threshold, as with decades of Windows, no single person understands the system, debugging becomes near impossible, and code added to improve things makes them worse. The open question for both is whether exponential hopes eventually flatten into an asymptotic plateau.

A telling anecdote: PayPal was the first company to offer cryonics as an employee benefit — $50k for neuro, $120k for full body — and enrollment collapsed because a dot-matrix printer couldn't print the policies. Thiel's wry moral: maybe the way to make biotech work is to push harder on AI.

4. Three Ways to Tackle AI

Each guest answers the same question — how do you attack AI as a company right now? — differently. Notably, none of them requires waiting for the endgame: each path monetizes intermediate milestones along the way.

1. 1

   Vicarious — brain principles

   Extract the neocortex's computational principles (hierarchy, sparse representation) rather than simulate neurons. Start with human-level vision, then image search, robotics, and diagnostics — with generally intelligent machines as the explicit end goal. Unrestricted object recognition alone would be tremendously valuable, so each milestone funds the next.

2. 2

   Prior Knowledge — Bayesian data

   Skip the brain entirely. Use Bayesian probabilistic models — math deliberately unlike everyday human reasoning — to find patterns and causal structure in data. The five-year goal: machines that find things in data that humans can't, compounding like Linux, where apps build on a core no one has to touch.

3. 3

   Palantir — augmentation

   Don't chase strong AI at all. Pair human conceptual judgment with machine-scale data processing — the PayPal anti-fraud lesson: humans can't scan millions of transactions, computers can't adapt to shifting adversaries, but the combination can. Squarely the Ricardian gains-from-trade play; it still took three years to land a paying customer.

5. Why Now, and How to Defend the Win

Why now? Data has outrun human analysts, and AWS turned server farms into a credit-card purchase, so the need and the compute finally coincide. Brown adds a striking projection: within 14 years, the world's fastest supercomputer will perform more operations per second than there are neurons in the brains of all living people — so the real race is figuring out what algorithms to run on it. Meanwhile academia rewards marginal papers and big companies shun decade-long projects, leaving few teams even attempting a Manhattan Project for strong AI.

Why copy the brain at all? Brown's answer is the airplane analogy: the Wright brothers didn't need detailed bird physiology — they needed the principles of lift. Likewise Vicarious hunts for the cortex's governing principles. Ferret experiments hint the bet is sound: rewire optic nerves into the auditory cortex and the ferrets learn to see, suggesting one common cortical algorithm underlies vision, hearing, and perhaps language.

You can't succeed by making a thing that has feathers and poops.

• Process as moat: like the Wrights' kite-to-glider-to-flyer discipline, rivals can copy your artifact but not the experimental process that produces the next one.
• Network effects: become the AWS of image recognition — every new user makes the system better and the feedback loop more entrenched.
• Escape velocity: keep out-innovating so the lead compounds; while rivals copy V1, you've applied the tech to hearing and language and shipped an improved V1 with more data behind it.
• Talent as filter: recruit by asking candidates what they care about — people serious about intelligent machines select themselves in.

6. Danger, Baggage, and Timing

On existential risk, Jonas jokes he worries more about getting cash-flow positive than about Skynet — though he plans to name his kid John Connor. The serious position: intelligence is orthogonal to volition. An oracle that reasons about facts is a neutral tool, and fearing it conflates having intelligence with having a will. Still, Brown says the work deserves the reverence you'd bring to building bombs or super-viruses, and McGrew notes computers can threaten civil liberties well short of strong AI — which is why Palantir works with privacy lawyers and civil-liberties advocates from the start.

AI's real handicap is its baggage: the War on Cancer spent 40 years to end up arguably further from victory; an infamous early MIT summer project expected to crack AI in months; the 1980s insisted AI was just around the corner. The rebuttal is twofold. First, humanity itself is an existence proof that general intelligence is physically possible — unlike faster-than-light travel, there's no theoretical barrier. Second, the smartest people in the field once declared heavier-than-air flight impossible, right up until the Wrights flew.

Timing errors kill companies, not ideas. Project Xanadu tried to network the world's computers from 1963 until it ran out of money in 1992; Netscape arrived the very next year and opened the Internet era. Thiel closes with Columbus, who talked his mutinous crew into just three more days at sea — and landed on a continent he wasn't looking for.

Which pretty much makes North America the biggest pivot ever.

Then vs. now (2026)

Self-check quiz

Pick an answer to reveal the explanation.