Cliff Sosin, founder of CAS Investment Partners, argued in an October 27, 2026 podcast that AI has hard limits in predicting complex dynamic systems like weather or economies, making intelligence cheap and abundant. He advocates owning complements to intelligence—data, attention, physical scale—rather than intelligence itself. He disclosed ~80% of his portfolio in Carvana, with other positions including Hilton Grand Vacations and Meta (about 5%). He likened Meta to 'personalized TV,' saying its data and attention assets won't be eroded by cheap AI. Sosin also said he avoids AI infrastructure due to extreme difficulty forecasting supply and demand.
At a Glance Cliff Sosin, founder of CAS Investment Partners, discussed his views on the predictive limits of AI and investment opportunities in a podcast. He argues that even the most intelligent AI has hard predictive constraints and explains why 80% of his portfolio is concentrated in Carvana. Sos
In the podcast, Cliff Sosin (Founder of CAS Investment Partners) discusses his core thesis on the limits of AI prediction: he believes even the most intelligent AI has hard predictive limits because most real-world problems belong to complex dynamic systems that cannot be solved through pure reasoning alone. Based on this judgment, he believes the truly valuable assets to own are complements to intelligence—data, attention, physical scale, regulatory barriers, etc.—rather than intelligence itself. He discloses that approximately 80% of the portfolio is concentrated in Carvana, with other holdings including Hilton Grand Vacations and Meta. Sosin says he avoids investing in AI infrastructure, arguing that supply-demand forecasting in this area is too difficult and that rapid technological iteration can quickly commoditize any lead. He compares Meta to "personalized television," believing its data and attention assets will not be eroded by cheap intelligence. He also discusses how AI agents will drive customers toward the lowest-cost suppliers, the physical speed limits of automated experimentation, and the dominant role of U.S. capital markets in risk-taking.
Sosin argues that a common mistake people make is to view intelligence as a boundless scalar ("from amoeba to monkey to Einstein to God"). He proposes an alternative framework: intelligence is more like the roundness of a ball—it is a ratio, not a height. Intelligence is a tool used for interpolative prediction between known facts. As intelligence improves, one's ability to fill in the gaps strengthens, but for complex dynamic systems, intelligence itself has hard limits.
> "What you should really think about intelligence as is more like a ratio, like the roundness of a ball."
Sosin uses the historical case of weather forecasting to illustrate this limitation: early weather models produced entirely different results upon repeated runs, simply because a technician rounded off the initial conditions by one fewer decimal place on the second input. This small error was amplified through iterative calculations, ultimately leading to drastically different predictions—"70°F and sunny" versus "20°F and a blizzard"—which is the butterfly effect.
> "In the limit, if you want to simulate the whole universe, what you actually need is the whole universe."
Many important problems (weather, human interactions, economic systems) fall into this category of complex dynamic systems. You can approximate them with models, but models inevitably contain some degree of granularity error, and after many iterations, that error accumulates to an unacceptable level. This means that even with "perfect intelligence," it is impossible to make accurate predictions for all interesting questions.
Sosin believes that what AI can truly do, and is already doing, is dramatically reducing the cost of intelligence. The result is that everyone will have "a 200 IQ team working full-time in their pocket." Sufficiently good intelligence will become nearly free, and this will solve the vast majority of everyday problems that require intelligence.
Sosin's logical chain is: if intelligence becomes cheap and ubiquitous, then what you truly want to own are the complements to intelligence — those assets that intelligence cannot replace and that instead appreciate in value as intelligence proliferates.
Complements to intelligence that Sosin lists include:
Sosin defines Meta's core business as "personalized television" — users enjoy watching content that aligns with their interests (such as the "British dark humor" he himself discovered).
From the cost side: Sosin notes that Meta's annual revenue is approximately $250 billion, and it spends nearly all of it. Using a highly simplified perspective, he argues that Meta's expenditures are essentially all purchases of "intelligence" — data centers (tools for producing intelligence), inference computing (intelligence itself), highly paid engineers (intelligence), code (intelligence), and video services (a close complement to intelligence). If the price of intelligence falls by a factor of 100 over the next five years, Meta's cost structure will undergo a fundamental transformation.
From the revenue side: Sosin believes creators will be able to use cheap intelligence to produce higher-quality content; Meta can better recommend videos; advertisers can create more precise personalized ads and build smoother conversion funnels.
Core moat: Sosin believes Meta possesses data and attention — neither of which will be destroyed by cheap intelligence, but will instead appreciate as intelligence improves. Users will not stop watching an interesting 20-second video just because intelligence becomes cheaper.
> "All they buy is intelligence. And if you think forward with the fact that the price of intelligence is falling by a factor of 100 or something over the next five years, first principles, their costs kind of go away."
Sosin uses Grindr as a counterexample (he does not hold it himself): Grindr has a large amount of attention and data, but is constrained by its company size and cannot leverage intelligence as effectively as Meta. If it could obtain "the resources of Google's best 10,000 engineers working for 5 years," it could do much more. Sosin believes small companies may be limited by their inability to fully deploy intelligence, but companies that own data and attention still have opportunities.
Sosin observes that the evolution path of model deployment starts with models of limited capability that can only handle a small portion of tasks; as models improve, they move up the stack, unlocking larger addressable markets. However, behind each tier, once evaluation metrics become saturated, you would no longer use a frontier model for simple tasks — these areas become clearly commoditized.
> "Once evals or whatever gets saturated, you would never use a frontier model for a chatbot in a customer service application. You'd be an idiot."
Sosin believes the demand for intelligence takes a diamond shape: initially, as models improve, application scenarios broaden; but beyond a certain point, the market served by further intelligence upgrades becomes increasingly narrow and rare.
> "How often do you really need a 300 IQ?"
His reasoning:
In response to the view that "demand will surge as intelligence becomes cheaper," Sosin points out: The Jevons paradox is merely a description of a specific point on the price elasticity curve, and it does not necessarily hold forever. He uses the history of fiber optics as an analogy: after communication costs dropped dramatically, the volume of information transmitted surged, and the market grew for a long time; but later, the market stopped growing, costs continued to fall, and the market actually shrank. Today, Cogent Communications carries 30-40% of internet IP traffic, yet its market capitalization is only a few billion dollars — a very small industry.
> "Jevons paradox is a statement about a particular spot on a curve, which doesn't necessarily have to be true."
Regarding the view that "intelligence can discover new problems on its own," Sosin responds: To answer new questions, you ultimately need to touch reality. A model can read all papers and discover overlooked connections — this would generate many new findings. But at some point, you need to conduct experiments, and experiments are constrained by physical speed. A model can propose better experimental designs, but the experiment itself takes time (e.g., breeding mice takes a year).
> "If you had a model and it was going to think more about physics, it could certainly think lots of interesting thoughts about physics, but eventually it's going to need to do experiments."
Sosin admits to "deliberately ignoring" AI infrastructure investments, citing the following reason:
> "I can't forecast supply or demand within several orders of magnitude."
Sosin compares the current wave of AI infrastructure investment to a family saving up to buy a house: society channels vast resources (electricians, plumbers, etc.) into building data centers, driving up costs in other parts of the economy. This process continues until overbuilding occurs. Because construction has a lead time and there is a "winner's curse," the last person to stop will regret it.
> "Invariably the last person to stop will regret it."
Unlike historical infrastructure investments, AI technology itself is evolving rapidly—both at the chip level and the model level, and could be commoditized by AI itself. Sosin points out that NVIDIA essentially sends flat files to TSMC—though these files are extremely difficult to produce, they can theoretically be replicated. If one steps back from the frontier, manufacturing costs drop significantly.
Sosin believes TSMC's moat is more durable because its manufacturing process is highly physical, with many details opaque even to AI—"someone who has worked in the system for years knows it must be done this way." However, TSMC also faces a challenge: because it cannot meet all demand, it effectively incentivizes the entire world to try to replicate its moat.
Sosin categorizes consumer decisions along a spectrum:
Sosin uses garbage collection services to illustrate the impact of AI agents:
First-order effect: Currently, people are reluctant to switch garbage collection companies, even when cheaper and better options exist. AI agents can automatically compare prices and switch, directing customers toward the lowest-cost provider.
Second-order effect: Suppose on a single street, if all households use the same garbage collection company, that company can offer a lower price due to route optimization. But in the past, coordination was difficult. Now, a provider with two customers can proactively offer a lower price to other households, creating a virtuous cycle of scale and cost — the low-cost provider wins more customers, thereby becoming even lower-cost.
> "The first order effect is that you move to the low cost provider."
Sosin predicts that as AI agents continue to optimize, prices will become more dynamic. Garbage collection fees may no longer be a fixed $60 per year, but could change weekly based on actual costs. Each garbage collection company would continuously recalculate its marginal cost and price in real time for each customer.
Sosin compares the current investment environment to the era of computer-assisted chess: humans and computers working together outperform either acting alone. Humans possess information that models lack (life experience, non-internet-based information), while models have computational advantages. The synergy between the two leads to better judgment.
Sosin believes that over the long term, as models gain access to more and more information, the added value of human judgment will gradually diminish. When a model's prediction of the probability distribution for the future is already optimal, adding human judgment only reduces accuracy—much like adding a personal opinion to a seven-day weather forecast.
> "That day may come."
The host mentioned an experiment by a friend: selecting the stocks themselves but delegating portfolio construction (position sizing) to AI. The result was that the AI's position sizing significantly outperformed the human version. Sosin's response:
> "A lot of that research, I think, is more looking at price changes in terms of securities prices and less focused on thinking about it the way someone who owns a company might think about it."
Sosin cites a classic expression: Price is a signal wrapped in an incentive. A rise or fall in price incentivizes you to buy more or less, while simultaneously conveying information about relative value, scarcity, and production costs. He argues that even with unlimited intelligence, it would be difficult to find a better coordination mechanism than price.
Sosin believes that if everyone has a continuously optimizing AI agent, price stickiness will be significantly reduced. Currently, many prices (such as deposit rates) exhibit inertia, and competition is less robust than theory would suggest. The continuous optimization of AI agents will allow prices to reflect real supply and demand more frequently and accurately.
Sosin points out that the U.S. has reallocated approximately 3% of its GDP to highly speculative AI infrastructure in an extremely short period, something no other place in the world can achieve. This capital allocation capability itself is a key constraint on productivity growth.
> "It is amazing that we can redeploy resources at this scale, this quickly."
Sosin believes that although large-scale construction is often accompanied by waste and fraud (as seen in the railroad era), this AI buildout has been "unusually clean." A large number of participants are trying to do the right thing, and the governance mechanisms are commendable.
Sosin's "pet theory": People who come to the U.S. are a subset of those most willing to take risks, because immigration itself is a high-risk behavior. Risk-taking has a genetic component and is more prevalent in the U.S. This tendency, in turn, fosters a culture and institutions that support risk-taking (reasonable bankruptcy laws, labor laws, venture capital), creating a self-reinforcing cycle.
> "If there's a moon base at some point, you're going to be bullish on that community's ability to do really well economically."
Sosin recommends that enterprises start with the board and CEO, having everyone participate in a hackathon — not just developers. The goal is not to produce usable software, but to let everyone understand what is possible, thereby changing expectations.
CEOs should ask their IT heads: Give my AI agent read-only access to all company systems. Any information I can obtain by asking someone should be accessible via a prompt. While this is difficult (involving countless systems), it is the starting point.
Start with read-only access, then gradually expand to write permissions, and then to permissions to modify systems. Sosin acknowledges this will take years, so the key is whether your business has enough time to complete this transformation or whether it will be disrupted by new entrants starting from scratch.
Sosin argues that personal intelligence is not useless in the AI era. The higher one's intelligence and the richer one's knowledge, the more effectively they can communicate with the model and guide it to do what they want. If you don't understand quantum physics, the model can only explain it to you "with crayons and Nerf balls," unable to reach the frontier.
> "You can't ask it to assume it's a stochastic process unless you know what a stochastic process is."
Sosin advocates for having children learn foundational skills without the model — not because they will never use the model in the future, but because without understanding multiplication, one cannot understand matrix multiplication, and without understanding matrix multiplication, one cannot understand what the model is doing. Understanding is the prerequisite for guidance.
For tasks such as a Shakespeare essay, Sosin believes teachers should set more difficult requirements: have students use the model to complete tasks that are 100 times more complex, and then come back to prove they truly understand the results and can explain "why it is this way." This trains both the ability to use the model and critical thinking.
Sosin proposes two opposing narratives:
Concentration Narrative: AI enables a small number of entities (labs, NVIDIA, etc.) to reap enormous returns, making income distribution more concentrated.
Flattening Narrative:
> "Truman asked for a one-handed economist. I seem to very distinctly have two hands on everything."
Sosin leans toward the view that overall profit margins will decline, with more value shifting to consumer surplus. However, he also notes that similar predictions were made in research reports during the 1990s internet boom, and they turned out to be wrong—due to the aggregation economy effect (Ben Thompson's aggregation theory). But in the age of AI agents, aggregation theory may break down: if an agent can search all results rather than just the first one, being ranked first no longer matters, making quality, price, and service far more critical.
Sosin believes AI could make more complex vertical integration businesses feasible. In the past, coordination costs and principal-agent problems pushed industries toward horizontal structures. Now, with AI agents, it is possible to build a full-stack pet care business (from pet food to insurance to veterinary services), reducing intermediaries and improving efficiency. Carvana is a precedent for this kind of vertical integration, built even before AI became widespread.
| Position | Guest View or Disclosed Action | Key Rationale & Data |
|---|---|---|
| Carvana | ~80% position, highly concentrated holding | Sosin believes Carvana's infrastructure and brand are highly differentiated, with few failure scenarios; Sosin believes its vertically integrated model is a replicable precedent for the AI era |
| Meta | Holds ~5% position | Sosin believes Meta possesses the two complementary goods of intelligence: data and attention; Sosin notes its $250 billion in revenue is almost entirely spent on "buying intelligence," and a decline in the price of intelligence will significantly reduce costs |
| Hilton Grand Vacations | Holds (position not disclosed) | One of the few positions in Sosin's portfolio |
| Grindr | Not held, used only as a case study | Possesses attention and data, but is constrained by company size and unable to fully deploy intelligence |
| NVIDIA | Not held, holds a skeptical view | Sosin believes NVIDIA essentially sends flat files to TSMC, theoretically replicable; Sosin believes rapid technological iteration may commoditize its lead |
| TSMC | Holding status not disclosed | Sosin believes TSMC's physical manufacturing process constitutes a more durable moat, but cannot meet all demand, incentivizing global replication of its moat |
| Cogent Communications | Used as an analogy | Sosin uses this as an example to illustrate that even with massive demand, the market can be small |
1. Intelligence is the roundness of a ball, not the height of a tree (Sosin's view): Sosin believes intelligence is not an unbounded scalar but a tool with hard limits. For complex dynamic systems, even perfect intelligence cannot make accurate predictions.
2. Own the complements to intelligence, not intelligence itself (Sosin's view): Sosin believes when intelligence becomes cheap and ubiquitous, assets that intelligence cannot replace—such as data, attention, physical scale, and regulatory barriers—will become more valuable.
3. AI agents will drive customers to the lowest-cost provider (Sosin's view): Sosin believes the delegable portion of consumer decisions will be automatically optimized by agents, breaking past inertia-based pricing and brand loyalty, creating a virtuous cycle of "lower cost → more customers → even lower cost."
4. Price is a signal wrapped in an incentive (Sosin, quoting): Sosin quotes this view, believing even with infinite intelligence, it is hard to find a better mechanism for economic coordination than price. AI agents will make prices more responsive and dynamic.
5. Demand for intelligence is diamond-shaped, not infinitely growing (Sosin's view): Sosin believes there are very few use cases for super-elite intelligence, while there are many use cases for "good enough" intelligence, but they are highly competitive and commoditized. Jevons' paradox describes a point on the curve, not a permanent rule.
6. Touching reality is a hard constraint on intelligence (Sosin's view): Sosin believes a model can propose a better experimental design, but the experiment itself is limited by physical speed (e.g., breeding mice takes a year). Automated experiments are good, but their speed is finite.
7. The US capital market's ability to rapidly reallocate resources is its core advantage (Sosin's view): Sosin believes the ability to quickly reallocate roughly 3% of GDP to AI infrastructure is something no other place in the world can replicate.
8. Education should cultivate the ability to "communicate with models" (Sosin's view): Sosin believes children need to master foundational concepts without relying on a model, not so they can avoid using models in the future, but so they can effectively guide models to perform more complex tasks.