Cliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497]

Summary

• The investment question is what remains scarce when intelligence becomes cheap and ubiquitous. Sosin would own “the complements to intelligence” — data, attention, physical proximity, scale, and emotional attachment — while scrutinizing software that packages intelligence and businesses whose margins depend on consumer inertia. The premise is powerful but bounded: more IQ cannot remove the need to “touch reality.”
• Meta is Sosin’s clearest example of an existing franchise strengthened by cheaper intelligence. His deliberately reductionist model pairs $250 billion of revenue with $250 billion of annual spending treated as intelligence; prices falling by “a factor of 100 or something over the next five years” would improve the economics of “personalized TV.” Better creative, recommendations, and advertising can monetize the same scarce data and attention rather than simply shrink costs.
• Frontier intelligence can keep improving without guaranteeing durable frontier-lab profits. O’Shaughnessy argues that private models and superior distillation could preserve lab advantages; Sosin grants that uncopyable quality and cost leadership would make a great business. His competing hypothesis is a “diamond-shaped demand frontier”: enormous demand for increasingly commoditized good-enough intelligence, but potentially few uses for elite capability.
• The AI build-out can transform the economy while remaining extremely difficult to underwrite. Sosin cannot forecast inference supply or demand “within several orders of magnitude,” and both bottlenecks and technology can shift before capital earns its return. He prefers identifying the CPG brands that follow the railroads, while admitting that lasting shortages and overbuilding are stories he can tell “with equal certainty.”
• Consumer agents could compress inertia-based profits while reinforcing genuine cost advantages. Garbage collection illustrates the loop: a 50-cent monthly saving triggers switching, route density rises, and the winner becomes cheaper still, potentially capturing much more volume at lower margins. “Price is a miracle because it’s a signal wrapped in an incentive”; agents should sharpen that mechanism.
• Sosin’s roughly 80% Carvana position reflects an explicit trade-off between expected return and unusual risk. He believes its infrastructure and brand leave few scenarios of severe failure. In the portfolio-sizing discussion, he says he is “underweight”; O’Shaughnessy jokes Carvana “should be 90%.” Sosin accepts that AI may eventually outcompete investors, but questions portfolio-sizing evidence without a long verification history.
• Cheaper intelligence could reduce economic rents without settling the direction of margins or inequality. Full-stack businesses might become more efficient and more profitable, while consumers gain from lower software, banking, and professional-service costs. Against those leveling forces, agents could magnify returns to initiative and ambition: “I seem to very distinctly have two hands on everything.”
• Education still needs human knowledge even when models can do the work. Sosin wants unaided foundations that let students direct models intelligently, alongside AI-assisted assignments “a hundred times harder and more ambitious.” Students must defend the resulting work and demonstrate that they understand it.

Deep dive

1. Intelligence cannot outrun missing information

• Sosin rejects the amoeba-to-Einstein-to-God extrapolation: intelligence resembles “the roundness of a ball,” rather than an unbounded height. It interpolates between sparse facts; eventually, more intelligence exhausts what existing information supports. Mathematics can admit perfect answers, while complex dynamic systems require observations and experiments. “You have to touch reality.”
• His weather-model example supplies the mechanism: rerunning a forecast after rounding initial conditions by one fewer decimal initially produced similar answers, then entirely different weather by the third, fourth, or fifth day. Errors compound through successive simulation steps. “If you wanna simulate the whole universe, what you actually need is the whole universe.”
• Automated labs can accelerate learning without abolishing physical timelines. If an experiment requires mice that need a year to grow up and die, it still takes time. A lab might have 5,000 ideas but capacity for 20; a model-selected experiment might have only a 1% success rate, even if that were ten times the average researcher’s odds. “Being smart has limits.”

2. Meta illustrates the value of owning intelligence’s complements

• The investment test is to assume everyone has excellent intelligence and ask “what’s left.” With “a team of 200 IQ people working full-time in our pocket,” agents can pursue a $400 annual saving consumers currently ignore. Software packaging “crystallized and hardened” intelligence, and businesses monetizing customer lethargy, face pressure. Complements include data, attention, emotional attachment, physical proximity, scale, and regulatory capture.
• Meta’s core business, in Sosin’s framing, is “personalized TV.” His deliberately reductionist picture puts revenue at $250 billion and annual spending, including CapEx, at $250 billion, treating that spending as intelligence purchased through computing and engineering. If intelligence prices fall “by a factor of 100 or something over the next five years,” costs could fall dramatically; he warns against taking the simplification literally.
• His upside scenario gives creators intelligence at 1/100 the cost and 10,000 times the effectiveness, alongside better recommendations, ads, and advertiser conversion. Meta might reinvest rather than eliminate spending, sell more ads, and win time from other activities. Its durable assets are “data and attention”: cheap intelligence does not make people stop enjoying funny 20-second videos.
• Grindr supplies a smaller-company variant: attention and data already exist, but a finite team and budget constrain what it can build. Sosin’s thought experiment supplies 10,000 of Google’s best engineers for five years. More intelligence could unlock more of that franchise. He twice stresses that he does not own Grindr.

3. Frontier labs face both commoditization and uncertain demand

• O’Shaughnessy’s countercase is that private frontier models may pull away, resist distillation, and let labs dominate cheap tokens too. Sosin concedes the economics if those conditions hold: superior models plus uncopyable, much cheaper adequate intelligence would make “a great business.” But he does not know whether labs can sustain that advantage; their financials are private, and bullish friends may have better information.
• His observed pattern is a moving frontier: better models unlock tasks, but those tasks commoditize behind them; customer-service chatbots need not use frontier models. He points to rapid commoditization at each step, including the rise of Chinese models, and collapsing costs. Even if labs lead in cheap tokens, the shrinking commodity market may limit the opportunity.
• Coding was a huge market. Beyond it, he considers a “diamond-shaped demand frontier,” with increasingly rare uses for elite intelligence. “How often do you really need a 300 IQ,” he asks, while remaining unsure that higher IQ alone delivers cancer cures.
• Against O’Shaughnessy’s Jevons-paradox argument, Sosin describes a market that first grows as a product improves and gets cheaper, then can shrink when further price drops no longer generate enough demand; at zero price, he says, market size goes to zero. Communications is his analogy: traffic growth eventually stopped offsetting price declines, and the market shrank as costs kept falling. He thinks one Cogent Communications division routes 30% or 40% of internet IP traffic, yet the company’s market capitalization is only a few billion dollars. Jevons describes “a particular spot on a curve.”
• O’Shaughnessy invokes a friend’s maxim, “Intelligence creates its own demand,” and asks whether models could generate new questions. Sosin grants, “Certainly, maybe.” Reading across disconnected papers will unlock discoveries, but eventually new hypotheses require experiments. Physical throughput then constrains progress; he does not assume thought alone supplies unlimited new, valuable applications.
• Sosin says his own model use has become commoditized: he has a Groqbot route work among his subscriptions on a server, and no longer knows which model is doing the work.

4. An essential build-out can still be difficult to underwrite

• Sosin cannot forecast inference supply or demand “within several orders of magnitude.” Even getting volumes right leaves questions about falling unit costs, bottlenecks along the value chain, and competition within each segment. His demand example is a five- or ten-minute, 3Blue1Brown-style mathematics explainer for his daughter: wonderful at very low cost, but something he does not think he would buy for $5. He also uses his Muse app to get workout-form pointers for very little money, but says he would not do so at a meaningful price.
• O’Shaughnessy’s pushback is that the smartest entrepreneurs are dedicating themselves to supply. Sosin reverses the causality: a huge infrastructure project requires strong price signals and extraordinary talent; their participation does not establish where an outside investor earns returns. He would rather identify what follows construction: “once you have railroads,” anticipate the CPG brands.
• Sosin likens society to a household saving for a house: resources dedicated to AI infrastructure mean less consumption today. The build-out absorbs electricians and plumbers, raising costs elsewhere. Because profits attract investment and projects have lead times, he expects expansion to continue until it overshoots: “the last person to stop will regret it.” Technology changes during construction, further complicating forecasts. He can describe both chronic shortages and overbuilding “with equal certainty.”
• Even Nvidia’s designs ultimately become “flat files that they send to Taiwan”; Sosin raises the possibility that intelligence could commoditize that work. TSMC’s process knowledge has more physical, tacit, accumulated attributes, though he disclaims expertise. Its difficulty meeting demand also pushes others to replicate its moat: “That’s not a great fact.”

5. Agents can strengthen scale while competing away inertia

• Delegability matters. For bank deposits, Sosin could almost entirely defer to an agent: given FDIC insurance, seek the highest rate. Choosing a favorite sweater sits at the opposite end; an agent can source options, but personal judgment remains essential. The distinction determines how aggressively optimization can alter customer behavior.
• Garbage collection shows how switching can reinforce scale. An agent might accept a 50-cent monthly saving after confirming that “the garbage will go away.” A collector with two customers on a street could recruit the rest, gaining route density and lowering costs further. The result could be much more volume at much lower margins: “everyone moves to the low-cost producer, who in turn gets to be the lower-cost producer.”
• Asked whether intelligence could replace price, Sosin calls it “a signal wrapped in an incentive.” Agents instead sharpen markets by reducing sticky, lazy pricing. A $60 monthly garbage bill followed by an unnoticed 5% increase to $63 could give way to frequently changing prices as providers recalculate marginal costs and customers continuously compare offers.

6. Companies need sustained organizational work to deploy cheap intelligence

• Sosin’s advice starts with everyone vibe coding, including directors and the CEO. Carvana calls its hackathons “pizza parties.” The immediate software output matters less than teaching employees what is possible and changing their expectations about what they can accomplish.
• Next, the CEO should demand read-only agent access to every company system containing information the CEO could otherwise request. Wider use then requires individual permissions; requests to enter data require write permissions; changing software requires a process for managing changes. O’Shaughnessy calls this “years of work.” Sosin calls it “impossibly hard,” making the key question whether the incumbent has enough time before a new entrant moves ahead.
• Starting without that baggage is attractive: “it’s the best time to start a company until tomorrow.” Agents allow much more work with fewer people, but easy execution is also available to competitors. Sosin asks where enduring economics emerge — lock-in, proprietary data, or another wedge — while acknowledging that obvious paths attract competition and exceptional outcomes can begin with ideas that initially make little sense.

7. Margins and inequality have forces pushing both ways

• Sosin initially sees pressure toward lower aggregate corporate margins and more consumer surplus. O’Shaughnessy notes that investors said the same thing about the internet in the 1990s. Sosin’s qualification is aggregation economics, associated with Ben Thompson: agents inspecting all search results could weaken the importance of ranking first and make quality, price, and service more salient.
• Cheap intelligence might also support more complicated, vertically integrated businesses. His example is full-stack pet care spanning food, insurance, hospitals, and veterinarians. Lower coordination costs and fewer principal-agent problems could make combinations viable that were previously too complex. Carvana is an example built before this became easier. Such businesses might produce both higher margins and greater consumer surplus.
• Distribution has similar crosscurrents. Removing bowling-software and banking rents benefits consumers, with savings larger relative to lower incomes; commoditized professional intelligence reduces wage premiums while broadening access to expertise. Conversely, agents become a “bicycle for the mind on steroids” for ambitious people with initiative, potentially increasing their returns. Sosin’s disclaimer: “I seem to very distinctly have two hands on everything.”

8. Capital markets finance a future that constrains present consumption

• America’s markets, Sosin argues, redirected roughly 3% of GDP into a speculative project “basically on a dime”; whether 2% or 6% would be right is unknown. He expects fraud eventually, but sees impressive governance relative to historical railroad construction. The everyday counterpart is buying $100,000 of stock in well-run, well-motivated companies almost instantly and at little cost, then selling when done.
• His pet theory for America’s advantage is an interplay of genetics, cultural evolution, and institutions. He suggests that people who came to the United States were a risk-seeking subset, and that risk-taking has a meaningful genetic component; risk-taking norms then helped foster bankruptcy laws, labor laws, and venture capital that reinforce the tendency. This is his theory, not a settled conclusion.
• Disclaiming macroeconomic expertise, he offers one possible contribution to higher rates: society must consume less today to build future productive capacity. A plumber constructing a data center is not producing present consumption. Paying higher interest induces the saving needed to support that reallocation; it is one part of a complex explanation.

9. AI challenges investing skill before resolving long-horizon portfolio sizing

• Sosin expects investors may eventually be “competed out,” but sees the present as freestyle chess: human and computer together outperform either alone. Personal experience and information absent from the internet still enrich his judgment. That gap should persist for a while, in his view.
• Eventually, a model’s distribution of outcomes might become the best available forecast without making the future predictable. Adding human judgment could then resemble Sosin adjusting a seven-day weather forecast: “However you fudge it, it’s just gonna make it worse.” He does not think that transition will affect his career much because it will take time.
• O’Shaughnessy describes a friend whose AI sizes a human-selected portfolio and has substantially outperformed his own construction. Sosin calls it clever but questions the verification horizon: these are not 20-year results. He expects encroachment to begin at seconds and move through minutes, hours, weeks, and months, while accepting that quantitative allocation enriched by qualitative judgment makes sense.
• When O’Shaughnessy suggests AI might weight Sosin’s portfolio differently, Sosin says, “I’m underweight.” O’Shaughnessy jokes, “Underweight Carvana. Should be 90%.” Later, Sosin defends his roughly 80% Carvana position by saying its infrastructure and brand leave, in his assessment, few scenarios of severe failure. Diversification research often measures security-price changes; he thinks as a business owner. He notes that diversification would have missed the full upside of discovering Walmart, Home Depot, or Fastenal early, while acknowledging the many concentrated investors who failed. His choice is a trade-off between “expected returns and knowingly taking an unusual risk.”

10. Education needs unaided foundations and far more ambitious assignments

• Sosin’s educational model is bimodal. Teach foundational skills without models so children can later direct them: multiplication supports understanding matrix multiplication, and knowing what a stochastic process is enables better instructions even when AI does the mathematics. Domain knowledge makes communication effective; without quantum-physics knowledge, explanations to him remain in “crayon and Nerf,” far from designing leading experiments.
• On the other side, Shakespeare-essay cheating signals insufficient ambition. O’Shaughnessy supplies the remedy — “Give them something harder to do” — and Sosin pushes for work “a hundred times harder and more ambitious,” using models fully. Teachers should question students about why the result is as it is and require justification. The aim includes teaching children that “they can do anything.”