Cliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497]
What this covers
The conversation is between the host and Cliff Sosin, who founded CAS Investment Partners in 2012 and keeps roughly 80% of his portfolio in Carvana. It opens with a weather-forecasting story: a model rerun diverged by the third to fifth day because a technician rounded the initial pressure and temperature data by one decimal place. He uses it to explain sensitive dependence in complex dynamic systems, then asks what that means for investors when intelligence becomes close to free.
The middle of the episode applies that frame to specific businesses. He treats Meta as personalized television and looks at the frontier labs as businesses whose tasks commoditize once evaluations saturate. He recasts Jevons paradox as a narrow case of price elasticity, with markets that are U-shaped, and cites Cogent Communications as a market that shrank as the cost of moving information fell. He argues that many software tools are crystallized intelligence and are therefore exposed to general-purpose models. He also challenges Ben Thompson's aggregation theory: if agents read every search result, ranking first stops mattering. The garbage-collection example shows agents pushing households toward the lowest-cost provider, which undercuts businesses that depend on consumer inertia.
The later sections cover why he avoids the AI build-out, drawing on lessons from past capital cycles and the physical moat of TSMC. Several of his claims are flagged as contested. Aggregate corporate margins may fall, though he notes that 1990s internet forecasts have a poor base rate. Cheap intelligence could flatten or steepen income distribution. Personal expertise matters more, not less, for directing models. He also defends holding 80% of a portfolio in one stock, on the grounds that diversification research mostly addresses average investors seeking average results.
Cliff argues intelligence has hard limits because most real-world systems are complex and chaotic and can only be understood through physical experiments, so as intelligence gets cheap, value accrues to the complements of intelligence (data, attention, physical proximity, scale) rather than to building its supply.
- Intelligence interpolates between sparse facts, so it cannot beat chaotic systems like weather where small errors compound
- Progress beyond the set of predictions intelligence can make requires experiments that touch reality, which take time
- Cheap, ubiquitous intelligence commoditizes supply, leaving scarce data, attention, proximity, and scale as the durable advantages
Investors should own complements to intelligence such as data, attention, and scale.
- Meta's core business is personalized TV, and its roughly $250 billion of annual spend is largely on intelligence-related inputs like data centers, inference, algorithms, and engineers, so falling intelligence costs should eventually reduce its cost base.
“And a really reductionist view is that they spend it all on intelligence.”
- If intelligence becomes cheap and ubiquitous, investors should own the complements to intelligence, namely data, attention, love and emotional attachment, physical proximity, scale, and in some cases regulatory capture, because these are not replicated by cheap intelligence.
“what you want to own are the complements to intelligence. And so the complements to intelligence are data, attention, love, affection”
- Meta's data about user behavior and its attention base compound each other and are not destroyed by ubiquitous intelligence, so cheaper intelligence may let Meta sell more ads rather than just cut costs.
“All of this is downstream of what they possess, which is not destroyed by ubiquitous intelligence, which is data and attention.”
- Grindr has lots of attention and data but is constrained by its finite team and budget, so access to abundant intelligence could let it build many more products, though Cliff does not own the stock.
“it turns out they've got lots of attention and they've got lots of data, but their ability to build all the things”
- A useful investing method is to assume everyone has great intelligence and then ask what is left as the source of value, because pushback that the problem is too hard is just an argument about intelligence.
“I've found it helpful to assume everyone has really great intelligence and to ask what's left.”
- Companies that can only deploy a finite amount of intelligence, due to cost or size, are at a disadvantage against competitors who can deploy far more, so cheap intelligence can magnify the advantage of larger players, as in the example of 10,000 of Google's best engineers.
“they only a certain size company and they don't have access to 10,000 of Google's best engineers for ten years”
Intelligence is bounded by chaos, unknowability, and slow real-world experimentation, including discovery limits.
- A rerun of a weather forecasting model diverged by the third to fifth day because a technician rounded initial pressure and temperature data by one decimal place, and the small error compounded into a completely different forecast, showing that many systems are unpredictable beyond a horizon regardless of intelligence.
“the technician who had loaded the data into the computer in the second run had rounded off at some decimal, one decimal less”
- Thinking of intelligence as an unbounded scalar that grows without limit, like the height of a tree, is an intellectual mistake; intelligence is better understood as a ratio, like the roundness of a ball, with real limits on what it can do.
“you're thinking about intelligence as an unbounded scalar — the height of a tree — isn't correct. What you should really think about intelligence as is more like a ratio, like the roundness of a ball.”
- Intelligence works by interpolating between a thin scattering of known facts; in mathematics it can reach perfection, but for most interesting questions, which are complex dynamic systems, intelligence cannot guess its way to a closed-form answer and simulations accumulate error.
“But for most of the interesting stuff — that is, most of the questions in the world — the mathematical term is they're complex dynamic systems.”
- Once super-smart reasoning has made the predictions it can make, which is a relatively small set of the interesting things one wants to know about the world, further progress requires gathering more information through experiments that touch reality, and that takes time.
“To get past that, you then need to go out and get more information about the world, and that just takes time. You have to do experiments, you have to touch reality.”
- Unless one simulates the whole universe, which requires the whole universe itself, models of chaotic systems will be grained and error-prone, so intelligence is fundamentally limited by unknowability even though simulation helps for many approximations.
“ultimately, in the limit, if you want to simulate the whole universe, what you actually need is the whole universe.”
- Closing the loop with automated biology labs will accelerate learning, but experiments requiring breeding mice that take a year to mature cap throughput, so even a 1 percent success rate on the best ideas may be ten times the average researcher while still being low odds.
“if that experiment requires breeding mice and mice need a year to grow up and die, that experiment's going to take time.”
- Models will find discoveries hidden across existing papers, such as an insight that follows from combining three published papers that nobody had connected, producing a wave of discoveries from the existing corpus of knowledge.
“the models read all the papers and they'll figure that out.”
- Education is becoming bimodal, and students need foundational skills learned without models, because to direct a model effectively one must understand the underlying concepts, since a model cannot be told to assume a stochastic process without knowing what it is.
“to direct a model to do your will, which I think is going to be an important skill in the future, you do have to understand something about the world”
- A 1 percent success rate can still be large relative to average researchers, but the total pace of discovery remains limited by lab capacity, meaning automated research will be awesome but will not suddenly reveal a new world humans cannot comprehend.
“this is not like suddenly the model looks around and says, you guys missed it.”
- If students cheat on Shakespeare essays with models, teachers should make assignments far harder and more ambitious that use models fully, and require students to bring work back and justify it, so they own the thinking.
“Give them something 100 times harder and more ambitious and have them use models to their fullest extent.”
Cheap agents erode incumbent software, rent-taking intermediaries, platform power, and corporate margins.
- Businesses that rely on consumer inertia and bounded rationality will be hurt when AI agents continuously optimize on consumers' behalf, because agents will switch customers to the lowest-cost provider rather than letting them stay out of inertia.
“But with an agent, you would think the agent will continuously optimize for them.”
- Many software tools are crystallized, hardened projections of intelligence, so a general-purpose AI model threatens their value, and certain software companies face risk as intelligence becomes a general tool.
“what many software tools are, are crystallized and hardened pieces, like projections of intelligence.”
- When every customer has an agent, garbage collection companies constantly rerun financials to find their marginal cost and bid for customers, and agents switch to the lowest-cost provider, so prices move frequently and accurately rather than rising 5 percent annually through inertia.
“Your garbage collection will vary by day, and one week it'll be whatever.”
- Every person will have a team of 200 IQ agents working for them full-time, and every company will provide similar teams to employees, optimizing small decisions that people are too lazy to optimize themselves.
“We're all going to have a team of 200 IQ people working full-time in our pocket”
- The cost of moving information fell and volume rose for a long time, but after the market stopped growing, falling costs kept shrinking the market, and today Cogent Communications routes roughly 30 to 40 percent of IP traffic while being a few-billion-dollar market cap company.
“nowadays, Cogent Communications, I think, routes 30 or 40% of all IP traffic across the internet”
- A low-cost garbage collector with two trucks on a street can offer everyone lower prices, and with agents, households switch simultaneously, which lets the low-cost producer win more volume at lower margins, creating a better business.
“the guy who has two people on my street can go to everyone else on the street and say, hey, I'll offer you a lower price”
- Corporate margins in aggregate will likely go down as agent-driven consumer surplus rises, though the base rate from 1990s internet forecasts suggests such predictions can be wrong.
“I think everything we're talking about would argue down.”
- Aggregation theory assumes consumers only look at the first search result, but if agents review all search results, being ranked first no longer matters and quality, price, and service become more salient, undermining platform power.
“if your agent looks at everything on search, then being ranked first doesn't matter.”
- When cheap software displaces rent-taking intermediaries like the software provider charging bowling alleys, consumers benefit because a larger share of their income goes to goods they buy, and money shifts away from firms like Bank of America, which flattens the income distribution.
“The capitalists lose, the average consumer wins. So that's leveling.”
- AI agents and cheap code can reduce coordination costs and principal-agent problems, enabling full-stack vertical businesses such as pet food through insurance and veterinary care, which could be higher margin and more efficient than fragmented horizontal industries.
“you could build the whole vertical stack.”
- Cheap intelligence could flatten income distribution by eliminating rent-taking intermediaries and commoditizing professional intelligence work, but it could also steepen it because smart people with initiative get more leverage, so the effect on inequality is uncertain.
“I can tell two stories.”
- Prices are sticky in the economy and many prices are lazily set, such as deposit rates, so the market is less competitive than in theory, and agents will reduce that stickiness.
“There's a lot of stickiness to prices. It's a well-known macroeconomic phenomenon, this price stickiness.”
Cheap, good-enough intelligence commoditizes, leaving small price-sensitive markets for elite intelligence.
- The rise of Chinese models shows the cost of intelligence collapsing at every step, so even if labs hold a cost position on tokens for tasks behind the frontier, that market may be small and competitive.
“If you were to look at the rise of Chinese models in every step. At a minimum, the cost has been collapsing.”
- Demand for AI is diamond-shaped: uses widen with coding and then narrow into rarefied niches, so there are few use cases for super elite intelligence and many for very good intelligence, which is highly competitive and commoditized.
“There are very few use cases for super elite intelligence. There are lots of use cases for very good intelligence.”
- The cost of intelligence is falling rapidly, and good-enough intelligence will become close to free for everyone and will solve the vast majority of questions where people need intelligence, even though it will never literally be free.
“really good enough intelligence is rapidly falling in price. And it turns out that really good enough intelligence will solve the vast majority of questions”
- As models improve, they unlock larger markets, but once evaluations saturate, tasks like customer service chatbots commoditize and no one uses a frontier model for them, so frontier models are needed only for a shrinking set of hard tasks.
“once evals or whatever get saturated, you would never use a frontier model for a chatbot in a customer service application.”
- Jevons paradox is just a narrow example of price elasticity; markets are generally U-shaped, where demand grows as price falls only up to a point, and further price cuts can shrink the market toward zero.
“In general, all markets are U-shaped.”
- Consumer demand for AI is very sensitive to price: Cliff would use AI tutoring videos or fitness form feedback at small cost but not at $5, so total consumer spending on AI is intrinsically limited.
“I don't think I would do that if it was going to be $5.”
Price signals drive infrastructure overbuild, and AI supply and demand cannot be forecast.
- A huge infrastructure build-out would be impossible without a strong price signal and many talented people, so causality runs both ways and the fact that top entrepreneurs are building supply reflects the price signal.
“it would be impossible for there to be a huge infrastructure build-out of a huge and important technology if there weren't a strong price signal and a ton of talent going at it.”
- Capital cycles overbuild because the society keeps pushing until it overdoes it, since lead times and the winner's curse mean the last person to stop investing will regret it, and the rapid technological dynamism makes this harder than past infrastructure buildouts.
“invariably the last person to stop will regret it.”
- Price is a miracle of coordination because it is both an incentive that changes buying behavior and a signal that conveys relative value, scarcity, and production cost, and it is hard to find a better mechanism even with unlimited intelligence.
“Price is a miracle because it's a signal wrapped in an incentive.”
- Cliff cannot forecast AI supply or demand within several orders of magnitude, and the value chain, bottleneck position, and competition make it very hard to know where to invest in the AI build-out.
“I can't forecast supply or demand within several orders of magnitude.”
Models increasingly displace human judgment in forecasting, investing, and competitive chess-like contests.
- In computer-aided chess there was a long period when humans working with computers outperformed either alone because they think differently and are synergistic, and investing is currently in that era, though humans will eventually be competed out.
“When computer-aided chess came along, there was this long span of time where freestyle chess was better”
- As the gap between human judgment and models narrows, adding human judgment to a model's distribution of outcomes will make forecasts worse, similar to adding one's opinion to a weather forecast, though this day is not yet here and models may judge management and price capital in the future.
“adding a human to the forecast would be like adding my judgment to whether or not it's going to rain in seven days.”
- AI is likely to encroach on investing first at short timescales, where verification loops are fast and many results accumulate, and then move to longer timescales, because a long-horizon strategy like a 20-year track record cannot be verified quickly.
“I think it's going to start at shorter timescales and work its way up to longer timescales. Because one of the challenges is that there's no verification loop.”
- Investors across timescales will increasingly delegate to models, and quantitative strategies can be enriched by encoding human qualitative analysis into a quantitative framework, as multi-strategy funds already do, adding dimensions of freedom to success.
“you can have a fully quantitative strategy, and then you can enrich it with what you think is qualitative human-designed analysis”
America's capital markets, risk culture, and institutions uniquely enable the AI buildout.
- Fraud and bad behavior will likely be discovered when the water goes out, but given the magnitude of the AI buildout, it appears cleaner than the railroad buildout, which had much waste and fraud.
“If you read about the railroad buildout or something, the amount of waste and fraud.”
- Rates are higher because a society making a large investment in the future must consume less today, so it must pay people to save; the plumber building a data center is future consumption rather than present consumption.
“rates are higher because we're inducing people to save for the future. And the way you do that is you pay them.”
- America's capital markets are underrated because any person can buy a diversified stake in well-run companies instantly and cheaply, and they redeployed roughly 3 percent of GDP into AI quickly, a capacity no other country has matched.
“I think we just redeployed, what is it, 3% of GDP right into this highly speculative venture”
- America dominates risk-taking because immigrants who came were the most risk-seeking, that trait may have a genetic component, and this selection was reinforced by bankruptcy laws, labor laws, and venture capital, which created a cultural and institutional feedback loop.
“So basically the people who came to the United States were the subset of people who were most interested in taking risk”
- If a moon base is built, the United States would likely be the one to build it because it has been persistently innovative, even though China is brilliant and numerous and Europe exists.
“So if there's a moon base, I wouldn't be shocked if it's ours”
Concentrated bets on great companies are justified despite survivorship bias in history.
- Cliff can tell both the story that AI demand will consume massive compute forever and the story that the build-out is overbuilt, because his own usage is extreme while others adapt slower and per-token costs collapse, and he holds both with equal certainty.
“And so maybe actually we're totally overbuilt. I can tell both stories with equal certainty.”
- Holding 80 percent of a portfolio in one stock is reasonable when the company's infrastructure, brand, and business are so differentiated that few scenarios work out badly, because diversification research mostly applies to average investors seeking average results.
“Were it not for my sense that Carvana's infrastructure and brand and everything else is so differentiated that there are few scenarios where it works out really badly”
- Historical studies of concentrated winners such as Walmart and Home Depot are distorted by survivorship bias, since history is littered with concentrated failures, though those who held the great winners would have done well with concentration.
“Yeah. Or the history is jumbled.”
New AI-era companies need wedges, and incumbents must phase in agent access to systems.
- It is the best time to start a company because capital is cheap for experimental ventures and starting without baggage allows fast progress, but since easy things are easy for everyone, a new company needs a wedge such as lock-in or proprietary data to produce enduring economics.
“Yeah, it's the best time to start a company until tomorrow.”
- To transform a company with AI, the CEO should push hackathons across the company and demand that a general AI agent get read-only access to every internal system, which is a multi-year process that then expands to write permissions and change management.
“I want my Grok Bot, Meta Muse agent, whatever, to be able to access every system in the company on a read-only basis”