Cliff Sosin: Intelligence as a Commodity

A technician reran a weather forecasting model after rounding off all of the initial conditions by a single decimal place. The two runs started almost identically. By the fourth or fifth simulated day, one said it would be seventy degrees and sunny. The other said twenty degrees and pouring rain (05:13). The error hadn't stayed small. It had compounded, step by step, until the forecasts had nothing in common. That accident is where chaos theory comes from, and the name people gave it was the butterfly effect: a wing flap in one place, a hurricane somewhere else (05:53).

Investor Cliff Sosin shares this story in conversation with Patrick O'Shaughnessy on Invest Like the Best, because he thinks it explains something most people get backward about artificial intelligence. The popular image has intelligence as an escalator: amoeba, monkey, person, Einstein, then a vertical line to something god-like (01:52). Sosin thinks that picture is a mistake. "What you should really think about intelligence as is more like a ratio, like the roundness of a ball," he says, not an unbounded number like the height of a tree

What you should really think about intelligence as is more like a ratio, like the roundness of a ball. (Cliff Sosin, [02:19])

This is not a case against artificial intelligence. Sosin calls himself AI-pilled (22:22). His point is narrower and, for investors, more useful: good-enough intelligence is getting cheap fast, and it will answer the large majority of questions people actually ask it (04:04). That is exactly why he thinks the valuable thing to own going forward isn't intelligence itself. It's whatever intelligence can't touch.

Owning What AI Can't Replace

Sosin's working method is to assume every competitor, every company, every customer already has access to excellent intelligence, and then ask what's left to compete on. His answer: the complements to intelligence. "The complements to intelligence are data, attention, love, affection, whatever sort of emotional attachment," he says, along with physical proximity, scale, and, more cynically, regulatory capture (08:35). None of those things get commoditized just because a smarter, cheaper model shows up.

If intelligence becomes cheap and ubiquitous, what you wanna own are the complements to intelligence. (Cliff Sosin, [08:14])

He walks through Meta as the cleanest case. He argues the company's core business has quietly shifted from social networking to personalized television, built on an endless supply of short videos matched to what each person likes to watch (09:35). Meta earns about $250 billion a year, and spends roughly the same amount (09:54). Sosin's reductionist read is that nearly all of that spending, on data centers, on inference, on engineers writing algorithms, is really just spending on intelligence. If the price of intelligence falls by a factor of one hundred over the next five years, as he expects, Meta's costs shrink while its two durable assets, the data it holds on billions of people and the attention it already commands, stay exactly as valuable (10:35, 11:30). He contrasts that with Grindr, the dating app, which he says has plenty of data and attention but has been limited by a small team and budget, the kind of constraint cheap intelligence could actually loosen (12:13).

The same logic makes Sosin skeptical of the frontier AI labs as pure businesses. He traces a pattern: a new model launches, can only do easy tasks well, then climbs the ladder to harder and more valuable work, until the easy layer behind it gets commoditized by cheaper competitors (13:46). Programming was a huge market for the frontier model. The question is what's next, and whether there are enough high-value, low-competition problems left to climb toward. He doubts intelligence alone unlocks something like curing cancer, because that still requires touching reality through lab experiments, which take time no matter how smart the model directing them is (15:37). He points to Cogent Communications, a small network company that now routes thirty to forty percent of global internet traffic through one of its minor divisions, as a warning about where this can end: the cost of moving information collapsed, demand grew for a long time, and then the market itself started shrinking even as prices kept falling (18:31). Jevons paradox, the idea that cheaper resources get used more, only holds for part of that curve, he says. Push the price low enough and the market can shrink instead of grow (17:44).

That skepticism extends to the AI supply chain generally, chips, data centers, and the rest. Sosin says he's been willfully ignorant of these stocks because forecasting inference demand or future chip costs is something he can't do within several orders of magnitude, and even a correct forecast doesn't tell an investor who in the value chain actually captures the profit (24:04). He prefers a different historical lens:

Instead of betting on which railroad tycoon is gonna make money and how many railroads we need, I'm trying to think about, okay, once you have railroads, I'd love to have the foresight to figure out the CPG brands come next. (Cliff Sosin, [27:30])

Physical scale is the complement he returns to most. He names TSMC, the Taiwanese chip manufacturer, as a business whose moat is built from decades of process knowledge that is "not even legible to an AI," the kind of tacit, physical know-how that resists replication even by very capable software (29:45). He runs the same thought experiment on something mundane: garbage collection. Once everyone has an AI agent negotiating on their behalf, he argues, people stop being too lazy to switch providers, and a street with one low-cost hauler can roll up all its neighbors, turning scale into an even bigger advantage (31:53). He expects this to sharpen prices broadly: instead of a flat $60-a-month bill that quietly rises five percent a year unnoticed, people may see garbage collection, and lots of other sticky-priced services, move day to day as agents compete on actual marginal cost (37:09).

A Concentrated Bet on the Idea

Sosin's own portfolio, rough justice, is about five positions, with eighty percent in Carvana, the online used-car retailer, and Hilton Grand Vacations, and five percent in Meta (42:34). He defends the concentration not on conventional diversification math but on conviction: Carvana's physical infrastructure and brand are differentiated enough, he says, that there are few scenarios where the business goes badly wrong, and that confidence is what makes a large position tolerable (55:20). It's a bet that mirrors his larger thesis. Carvana already built the kind of full-stack, vertically integrated business that Sosin thinks cheap AI coordination will make possible across other industries, from pet care to hospitality, by solving the principal-agent and complexity problems that used to force companies to specialize narrowly (48:34).

His advice to the companies he owns is almost procedural. Start with hackathons for everyone, not just engineers, to reset what people believe is possible (42:59). Then have the CEO demand read-only access to every internal system through an AI assistant, which sounds simple and turns out to be a multi-year fight over permissions, before write access and a change-management process can even be discussed (43:27). He is candid that this sequence can take years, while a faster-moving competitor could win the market first (44:39).

Sosin is careful not to oversell certainty. He holds two contradictory stories at once: AI build-out could be wildly underbuilt given how much he personally uses it, or wildly overbuilt because most people won't adapt as fast as he has (52:13). He thinks ubiquitous intelligence likely pushes corporate margins down and consumer surplus up, since so much of what companies currently charge for is just crystallized intelligence that gets cheaper to produce (48:00). He even expects his own profession to get automated, starting with short-timescale trading before it reaches the long-term judgment calls he makes himself (54:10). What he's confident about is narrower and, in a noisy debate about AI, oddly calming: intelligence is going to get much, much cheaper, and the money will go to whoever owns what stays scarce anyway.