Intelligence Gets Cheap: Cliff Sosin on Owning the Complements

Contents

1. What this episode is about
2. 1. Intelligence is the roundness of a ball, not the height of a tree
3. 2. One rounding error, one model sunny and the other pouring
4. 3. Assume everyone has great intelligence, then ask what’s left
5. 4. The demand frontier is diamond-shaped
6. 5. Reality has speed limits, and mice still take a year
7. 6. Agents make prices move again
8. “Can I still buy the AI names?”
9. ”Will the model eat what I’m good at?”
10. ”80% in one name — isn’t that insane?”
11. Sources worth your time
12. The one thing worth taking with you

Hearing of a thing is better than not hearing of it; seeing it is better than hearing of it; knowing it is better than seeing it; and doing it is better than knowing it. Learning comes to rest only in the doing.

—— Xunzi, “The Effect of the Confucian” (Warring States; my own translation)

On 27 October 2026, Invest Like the Best EP.497 paired Patrick O’Shaughnessy with fund manager Cliff Sosin. His claim is that intelligence has a ceiling, which is why intelligence will get both very good and very cheap — and the investment consequence sits on the other side, in the complements: data, attention, affection, physical proximity, scale. Meta is his worked example: roughly $250 billion of revenue a year and roughly $250 billion of spending a year, most of which buys intelligence. If the price of intelligence falls by a factor of a hundred over five years, the cost side collapses while the data and the attention stay. For that to hold, one thing has to be right: good-enough intelligence commoditizes, and the uses for elite intelligence never get numerous enough to hold a high price.

What this episode is about

Cliff Sosin runs CAS Investment Partners and is the kind of manager who puts the whole fund into a handful of names — Patrick says the shape out loud on air: about five positions, one of them 80%. The two of them talk on the phone fairly often, and this episode is one of those calls recorded again, so it moves like two people at a table pushing one idea as far as it will go. The starting point is the line Cliff used on that call: it’s just intelligence. From there the argument runs out through chaos theory to garbage collection companies and ends up at how his kids’ school should teach.

1. Intelligence is the roundness of a ball, not the height of a tree

Most people picture intelligence as a line going up forever: amoeba, monkey, person, Einstein, then zoom, God. Cliff thinks that is an intellectual mistake. Intelligence is a gadget we use to predict the universe. We hold a thin scattering of facts, and intelligence colors in the gaps between them. So as intelligence rises, the coloring gets better — but it is bounded like a ratio, like how round a ball can get, not like how tall a tree can grow.

In mathematics you can color all the way in. Most of the interesting questions, though, are complex dynamic systems where no closed-form answer exists. You can simulate, and a simulation carries error that compounds with every step. His limit statement is the elegant one: if you want to simulate the whole universe, what you need is the whole universe.

2. One rounding error, one model sunny and the other pouring

He backs this with the origin story of chaos theory. A man ran a weather model on a computer that filled a room, liked the output, reset it and ran it again to check. The first days matched, then the runs diverged, and by day three, four, five of the forecast they described different worlds. He assumed a bug and spent a long time looking for it. The answer was that the technician loading data for the second run had rounded every pressure and temperature initial condition off by one decimal. You would expect a small, steady error; instead it compounded through every iteration, and partway into the forecast one model said 70 degrees and sunny while the other said 20 degrees and pouring. Cliff’s use of it is restrained: you can know everything about pressure and temperature and still not write down a solvable equation that tells you a hurricane is coming. You can only model it, and the model is grained.

3. Assume everyone has great intelligence, then ask what’s left

This is the most portable tool in the episode. When Cliff looks at a company he assumes everybody has excellent intelligence, then asks what remains. The pushback he gets is some version of you don’t understand how hard this is, how many smart people thought about it for how long — and his answer is: fine, but that’s all just intelligence.

Follow it through and what you want to own are the complements: data, attention, love, emotional attachment, physical proximity, scale, and the one he calls a little cynical, regulatory capture. Meta is the worked case. The core business now is personalized TV rather than social media, and he mentions discovering he loves dark British humor. So where does the $250 billion go? Data centers are intelligence-producing tools, buying inference is buying intelligence, paying smart people to write algorithms is intelligence, writing code is intelligence, surveying videos is intelligence-adjacent. The cost side caves. On the revenue side, creators with intelligence at a hundredth the cost make better content, advertisers make better ads, and Meta targets them better. His conclusion is that they probably don’t take costs to zero — they sell a few more ads.

He also names a company he says plainly that he does not own: Grindr. Plenty of attention and plenty of data, held back by a finite team and budget. What if it had ten thousand of Google’s best engineers for five years?

4. The demand frontier is diamond-shaped

Patrick argues the bull case: frontier intelligence isn’t a commodity, and the labs can supply the cheap tokens too. Cliff answers in two layers.

The first is the path he has watched models take through the stack. Models improve, climb a level, and each step unlocks a larger addressable market — but behind them, once the evals saturate, those tasks commoditize. Nobody runs a customer-service chatbot on a frontier model. So there is a wave of tasks that genuinely need the frontier, and everything behind it does not. The second layer asks how far that wave travels. The bull version is that people will pay anything for a model that cures cancer, and Cliff folds it back into his first point: curing cancer means touching reality, and more IQ doesn’t buy that.

Hence the diamond: wider and wider at first, then increasingly rarefied. His question is the one I liked most — how often do you really need a 300 IQ? His analogue is fiber. Moving information used to be expensive and we moved little of it; costs fell, volume exploded, the market grew for a long time; then the market stopped growing while costs kept falling, and the market shrank. He points out that Cogent Communications routes 30 to 40% of all IP traffic through a division that isn’t even its largest, and the whole company is worth a few billion dollars. Jevons paradox is a claim about one spot on an elasticity curve, not a law.

5. Reality has speed limits, and mice still take a year

Patrick asks about automated-lab companies. Cliff agrees closing the loop in the physical world accelerates learning, then adds the limit, and his phrasing is good: you’re really smart, so you must understand — being smart has limits.

Concretely: wire an automated biology lab to a computer, let it generate ideas, run experiments, iterate. Wonderful. But if the experiment requires breeding mice and mice need a year to grow up and die, that experiment takes a year. The lab is only so big, so you may have 5,000 good ideas and get to run 20. The model picks better, so maybe your odds are 1% — possibly ten times the average researcher, which is remarkable, and still 1%.

6. Agents make prices move again

Cliff calls price a miracle because it is a signal wrapped in an incentive: it tells you about scarcity and cost while also making you want more or less of the thing. Even with unlimited intelligence, he says, you’d be hard-pressed to design anything better.

So what do agents add? They sharpen markets. A lot of prices today are set lazily — deposit rates are the classic case. His example is garbage collection: you don’t remember what you pay a month, so a cheaper, better hauler has to spend real effort to get you to bother switching. Agents can just negotiate — I’ll save you fifty cents a month; will the garbage go away; yes — and you switch. The first-order effect is that everyone moves to the low-cost provider. The second-order one is more interesting: it would be more efficient for everyone on a street to share a hauler, since one truck does the run, but nobody ever coordinated it. Now the hauler who already has two houses can go to every other house and take the street. It wins volume at thinner margins and the customers keep the surplus.

“Can I still buy the AI names?”

I get asked this every day, and the answer here is uncomfortable. Cliff says he has been willfully ignorant about the supply side and admits he may be worse off for it. The difficulty he lays out is a chain: forecast how much more inference people will use, forecast how much cheaper and better inference gets, and even if you get both right, you still have to sit at the bottleneck to capture the rents, and then beat your competitors in that layer. All four or nothing.

Patrick pushes back: if supply and demand are this unknowable, why are the best entrepreneurs alive putting their lives into the supply buildout? Cliff reverses the causality — a buildout of this size for a technology this important could not happen without a strong price signal and a flood of talent, so your causality may be backwards. His own choice is to skip guessing which railroad tycoon wins and how many railroads we need, and instead think about what comes after the railroads; his example is consumer brands. He also expects it to be overdone, because of lead times and the winner’s curse the last person to stop will regret it.

Using that chain as a checklist beats listening to stories. Which of the four links does your holding actually have an answer for? If the answer is that it sits inside a very large theme, the theme is what you own. My own book is concentrated in a few software names, and my first reaction to this stretch was to go back and ask the dumb question: assume everyone has great intelligence — what do these companies still have? For one of them I have an answer. For the other two my answer is thin, and I’m still sitting with that.

”Will the model eat what I’m good at?”

His education segment answers this almost by accident. He asks what your own intelligence is worth in a world of infinite model intelligence. You might say nothing, but that doesn’t feel right to him, and he locates why: the more intelligence and the more crystallized intelligence you have, the better you are at telling a model what you want.

His self-deprecating version is specific. He couldn’t hold a deep conversation about designing frontier quantum physics experiments, because he knows nothing about quantum physics; the model could try hard and would still be explaining in crayon and Nerf. The compressed version: you can’t ask it to assume a stochastic process unless you know what a stochastic process is.

So his view of school has gone bimodal. On one side, foundational skills get taught away from the models, the reason being that without multiplication it’s hard to understand what matrix multiplication is, and without understanding what the model is doing you can’t direct it. On the other side, when an English teacher complains that kids cheat on their Shakespeare essays, his response is that the teacher isn’t being ambitious enough: give them something a hundred times harder, let them use the models to the hilt, then interrogate them on why it reads the way it does and make them defend it.

”80% in one name — isn’t that insane?”

Patrick asks him to his face, and he gets the funniest line of the episode back: I’m underweight. Then the substance. That body of research, he says, looks at changes in securities prices and spends less time thinking the way an owner of a business would; and he is comfortable holding this much because he judges the company’s infrastructure and brand differentiated enough that few scenarios end badly. The honest cut is his own: that literature is what people looking for average or slightly better than average results should follow, and it is entirely right for them — it just isn’t what he’s after. It’s a trade between expected return and knowingly taking an unusual risk. The people who did really well historically were concentrated, and the history is also littered with the ones who died that way; Patrick supplies the words survivorship bias.

What I take from it is that the question gets rewritten into a condition you can write down: which scenarios make this go badly, how many are there, can you see them? If you can write them, the position can be heavy. If you can’t, the heavy position is a bet on a feeling. And one line I’ll remember: if the stock ran a long way and he felt compelled to diversify, that would be fine with him too.

Sources worth your time

• Invest Like the Best EP.497, 27 October 2026, hosted by Patrick O’Shaughnessy, joincolossus.com
• Edward Lorenz, “Deterministic Nonperiodic Flow” (1963) — the source of the rounding error and the butterfly
• William Stanley Jevons, The Coal Question (1865), the original statement of Jevons paradox
• Ben Thompson’s aggregation theory, which he uses to explain why the 1990s “the internet will crush corporate margins” reports missed
• Cogent Communications’ public filings, if you want to check the traffic share against the market cap yourself

The one thing worth taking with you

The idea I’m keeping is this: being smart has a ceiling, and the rest of your progress gets bought by touching reality. No amount of IQ skips the mouse that needs a year to grow, which is bad news for those of us who like to finish things inside our heads.

Something I’ve tried: take one thing you’re currently “thinking through before acting” — changing jobs, whether to learn a craft, saying the thing to the person you’ve been avoiding — and today run the smallest real test of it. Make the call, book the one meeting, make one of it and show it to one person. You tend to find that three months of thinking hands you less than one test does, and the test hands you the kind of information you couldn’t have thought your way to. A thing I’d put off for weeks got settled last month by a seven-minute phone call, after I’d written two pages of analysis about it.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.