Cliff Sosin: Why Free Intelligence Won't Change Everything for Investors

Investor Cliff Sosin argues that superintelligent AI has hard limits, that good-enough intelligence is becoming nearly free, and that the biggest winners may own data, attention and physical scale rather than models. He makes the case to Patrick O’Shaughnessy on Invest Like the Best, in a 71-minute episode released October 27, 2026. Sosin, whose concentrated portfolio is roughly 80% Carvana, offers a deliberate counterweight to the “unlimited demand for intelligence forever” view that dominates AI investing.

TLDW

Sosin says intelligence is better thought of as a ratio than an unbounded quantity: it fills the gaps between known facts, but chaotic systems like weather, markets and people cap what pure thinking can predict, so progress eventually requires slow real-world experiments. If good-enough intelligence becomes cheap and ubiquitous, investors should own its complements, such as data, attention, physical proximity, scale and even regulatory capture. He doubts the frontier labs can escape commoditization, treats Jevons paradox as one point on a U-shaped demand curve, and avoids the infrastructure buildout because it is too hard to forecast. The back half covers AI agents that make consumers ruthlessly price-sensitive, America’s capital markets, an AI adoption playbook for CEOs, income distribution, portfolio concentration and education.

Thoughts

The sharpest idea here is that AI agents could break aggregation theory. Ben Thompson’s framework explains why internet platforms captured so much value: consumers click the first result, so whoever owns the top of the funnel wins. Sosin points out that the 1990s analysts who predicted the internet would crush margins were wrong mainly because of aggregation. But an agent does not get tired after the first result. It reads everything, compares price, quality and service, and switches the moment a better offer appears. His garbage collection example makes it concrete: today nobody switches haulers to save 50 cents a month, so lazy pricing survives. With agents, the low-cost hauler can win a whole street, get cheaper still, and prices start moving weekly instead of rising 5% a year unnoticed. That is a threat to every business whose margin depends on customer inertia, from bank deposits to subscription software, and a gift to genuine low-cost producers. It is also the logic behind his Carvana position.

His answer to Jevons paradox deserves more airtime in AI debates. Jevons says cheaper resources get used more, and bulls treat it as proof that falling token prices mean exploding spend. Sosin’s correction is that every market is U-shaped: demand grows as prices fall, until further cuts no longer bring enough new usage and the market shrinks. His example is moving data. Bandwidth costs collapsed, traffic exploded, and yet Cogent, which by his account routes 30 to 40% of internet traffic through one division, is a few-billion-dollar company. He pairs this with a “diamond-shaped” view of demand: use cases widen as models improve, then narrow at the top, because few tasks really need a 300 IQ. You do not have to agree to see the value. It turns “will AI demand grow?” into the better question of where on the curve each layer of the stack sits.

The CEO playbook is the most practical segment, and it is humbling. Step one is that everyone vibe codes, starting with the board and the CEO, so the whole company learns what is now possible. Carvana already runs these hackathons and calls them pizza parties. Step two is that the CEO demands read-only AI access to every internal system. The IT head will say it is hard, and that is the point. Then come write permissions, then a process for changing systems. Four bullet points that are really years of work. Sosin’s question for any incumbent is whether it has enough time to finish that migration before a startup that began with none of the baggage reaches it. That is why he calls it the best time to start a company, “until tomorrow,” because anything easy to build is easy to copy and the agents will switch providers daily.

His take on education fills the gap left by “why learn anything if the model knows it?” Personal knowledge matters because it is how you direct a model. Sosin says he could not run a serious quantum physics conversation with an AI because it would be explaining things “in crayon,” while in his own field he can steer it well. So kids should learn foundations without models, multiplication before matrix multiplication, not to do the math by hand later but to understand what the model is doing. And for teachers worried about students cheating on Shakespeare essays, his answer is to assign something a hundred times harder, let them use every tool, then make them defend the result. Learn the basics alone, then work with AI on something ambitious: that is a better policy than banning or ignoring the tools.

Key Takeaways

Sosin thinks of intelligence as a ratio, like the roundness of a ball, not an unbounded height. It interpolates between known facts and tops out in domains that cannot be solved in closed form.
Most interesting questions involve nonlinear dynamic systems like weather or human behavior, where small errors compound, as in Edward Lorenz’s discovery of the butterfly effect. Beyond a point, you have to “touch reality” with experiments.
Good-enough intelligence is rapidly falling in price, and it will answer the vast majority of questions people need intelligence for.
When intelligence is cheap, own its complements: data, attention, emotional attachment, physical proximity, scale and regulatory capture.
Seen from space, Meta spends its roughly $250 billion of revenue on intelligence, through data centers, inference, engineers and content review. If that cost falls 100x while its data and attention remain, Meta wins on costs, creator content and ad targeting.
Smaller companies with lots of data and attention but limited engineering, like Grindr, could benefit from having the equivalent of 10,000 top engineers.
Frontier models move up the stack, and the tasks they leave behind commoditize. Sosin sees a “diamond-shaped” demand curve with few uses for super-elite intelligence.
Jevons paradox describes one spot on a U-shaped curve. Bandwidth got cheaper until the industry shrank, leaving Cogent a few-billion-dollar company.
Sosin avoids the infrastructure buildout because supply and demand cannot be forecast within orders of magnitude, and the last one to stop building will regret it. He prefers to ask what comes after the railroads.
Personal AI agents will make consumers far more price-sensitive and could break the aggregation economics that protected internet platforms.
America redeployed about 3% of GDP into a speculative technology almost overnight, which Sosin calls an underappreciated achievement of its capital markets. Higher rates partly reflect paying people to save for that future.
Cheap intelligence could flatten incomes by competing away rents and commoditizing professional work, or steepen them by multiplying returns to ambition. He calls himself a two-handed economist.
AI will take over investing at short time scales first, working out toward longer horizons where results take years to verify.
Kids should learn foundational skills without models so they can direct models later, then use AI on far more ambitious assignments.

Chapters

0:40 Intelligence Is a Ratio, Not a Tower

Sosin rejects the idea that intelligence climbs from amoeba to Einstein to God. Intelligence interpolates between scattered facts, and math lets it approach perfection, but chaotic systems do not. He retells the weather-model story behind chaos theory, where rounding the starting data turned a sunny forecast into a downpour. To simulate the universe perfectly, you need the universe.

6:27 The Complements to Intelligence, and Meta as Personalized TV

Everyone will carry a team of 200 IQ agents optimizing the details they are too lazy to handle, which is bad for businesses that profit from consumer inertia and for software that is “crystallized intelligence.” His test for any company is to assume everyone has great intelligence and ask what is left. Meta keeps its data and attention while its costs fall, and Grindr could build far more than its team size allows.

15:50 Frontier Labs, Commoditization and Jevons Paradox

O’Shaughnessy presents the bull case: the labs are slowing distillation and keeping their best models internal. Sosin says saturated tasks commoditize quickly, Chinese models keep collapsing costs, and curing cancer needs experiments, not just IQ. He answers Jevons paradox with the bandwidth industry, and says that even AI generating its own questions is bound by how fast real experiments run. He has stopped tracking which model his own agents use.

24:50 Automated Labs Still Hit Speed Limits

Closed-loop robotic labs are a great idea, he says, but if an experiment needs mice that take a year to grow, it takes a year. A model might pick experiments with a 1% success rate, ten times better than the average researcher, and it is still 1%. AI will mine forgotten connections in existing papers and speed research, but it will not suddenly reveal a world we cannot comprehend.

26:57 Why He Avoids the Supply-Side Buildout

Forecasting inference demand, cost declines, value-chain bottlenecks and competition all at once is too hard for Sosin. Consumer demand is price-sensitive: he would not pay $5 for the math explainer videos he generates for his daughter. The best entrepreneurs are building supply because a buildout this big needs strong price signals, but society is like a household saving for a house, and the last builder to stop will regret it. Even Nvidia could face AI-driven commoditization, while TSMC’s tacit process knowledge is harder to copy.

35:08 Agents, Garbage Collection and Price as a Miracle

Consumer choices sit on a spectrum from fully delegable, like deposit rates, to fully personal, like a favorite sweater. Agents will move delegable spending to the low-cost provider, and local density winners can consolidate whole streets. On investing, he expects a long human-plus-machine era like freestyle chess before models outforecast humans. He calls price “a signal wrapped in an incentive,” and doubts unlimited intelligence will replace it.

44:40 America’s Capital Markets and Higher Rates

Sosin praises the speed at which U.S. markets moved about 3% of GDP into AI, and expects fraud to surface eventually but far less than in the railroad era. His theory of American dynamism mixes self-selected risk-takers, culture and institutions such as bankruptcy law and venture capital. Higher long-term rates partly reflect a society paying people to defer consumption for the future.

51:10 The AI Adoption Playbook for CEOs

Everyone vibe codes, starting at the board, and the CEO demands read-only AI access to every system, followed by write permissions and change management. That is years of work, and incumbents must finish before startups with no baggage catch up. It is the best time ever to start a company, but easy-to-build products are easy to copy, so founders need a wedge into lasting economics.

55:05 Margins, Vertical Stacks and Income Distribution

More consumer surplus is the default outcome, unless AI lets companies run complex vertical stacks, like a pet company spanning food, insurance and vet hospitals, with Carvana as an early example. Cheap intelligence could flatten incomes by removing rents and commoditizing professional work, or steepen them by rewarding initiative. He closes the section by telling both the shortage and the overbuild stories “with equal certainty.”

1:02:58 Concentration, AI Portfolios and Education

O’Shaughnessy describes a friend whose AI-run portfolio sizing beats his own. Sosin expects AI to win at short time scales first and defends his own concentration in Carvana as an owner’s bet on a business with few bad outcomes. On education, he wants kids learning foundations without models so they can direct them, and harder assignments that use AI fully.

Notable Quotes

“Really good enough intelligence is rapidly falling in price. And it turns out really good enough intelligence will solve the vast majority of questions where people kind of need to use intelligence.”
Cliff Sosin, on where AI is heading

“If intelligence becomes cheap and ubiquitous what you want to own are the complements to intelligence.”
Cliff Sosin, on his investing framework

“I think figuring out how to cure cancer requires touching reality. And I’m not sure that more IQ gets you that.”
Cliff Sosin, on the bull case for frontier models

“Price is a miracle, right? Because it’s a signal wrapped in an incentive.”
Cliff Sosin, on whether AI could replace prices

“It’s the best time to start a company until tomorrow.”
Cliff Sosin, on founders in the agent era

“I can tell both stories like with equal certainty.”
Cliff Sosin, on whether AI infrastructure is underbuilt or overbuilt

“Give them something a hundred times harder and more ambitious and have them use models to their fullest extent.”
Cliff Sosin, on students using AI to cheat on essays

Watch the full episode of Invest Like the Best here: Why Free Intelligence Won’t Change Everything.

Cliff Sosin: Why Free Intelligence Won’t Change Everything for Investors