Patrick O'Shaughnessy

Cliff Sosin is comfortable keeping most of his portfolio in Carvana because he thinks about the stake the way an owner thinks about a company, not the way a study of price swings tells a typical investor to diversify. Patrick O'Shaughnessy describes the portfolio as roughly 80 percent Carvana, a vacations holding, and about 5 percent Meta, with one position at 80 percent, and Sosin agrees. Asked whether an AI sizing tool would put less in Carvana, he says he is underweight and that it should be 90 percent.

He would not hold this much if he did not believe Carvana's infrastructure and brand left few scenarios that work out really badly. He compares it to finding Walmart, Home Depot, or Faste

Cliff Sosin says that once people hand routine choices to an AI agent, the customer moves to the lowest-cost provider. He puts decisions on a spectrum. An insured bank deposit can be delegated: take the highest rate. A favorite sweater cannot. The agent can find options, and the person still has to choose. Garbage collection, in his stylized example, sits near the deposit. A cheaper hauler still struggles to win a switch today, because many people do not know what they pay and will not spend attention on it. An agent can hear an offer of 50 cents a month, confirm the trash still gets picked up, and switch.

A second effect is coordination on a street. One truck serving every house is cheaper

Cliff Sosin says the businesses that benefit when intelligence gets cheap own the complements to intelligence, not the intelligence itself. Assume everyone already has very good intelligence, and ask what is left. The complements he names are data, attention, affection, physical proximity, scale, and, more cynically, regulatory capture. A product that is only a hardened piece of intelligence can be hurt, because a general model does that work. A company held back by the cost of intelligence can be helped once that cost falls. A company whose profits depend on customers being too tired to shop around can be hurt.

Meta is his example. He calls the core business personalized TV. It used to be

Cliff Sosin argues that children should learn foundational skills apart from models so they can direct a model. You cannot ask a model to treat a problem as a stochastic process unless you already know what a stochastic process is. A model could teach the term, but figuring it out that way adds a lot of friction. You do not have to work out the math, but you do have to know what it is.

Foundational skills build on one another, and children have to be able to do them without models. The point is not that later work will happen without models. If they cannot multiply, matrix multiplication is very hard to understand. They need not perform matrix multiplication especially well. They do need a

Cliff Sosin tells two stories about how cheap, ubiquitous intelligence affects wealth and income distribution. One story, already widely discussed, holds that income concentrates: people made money on labs, data, and NVIDIA, with increasing returns to small groups.

The counterargument starts with rent takers. A software company that charges bowling alleys gets smoked. Losing the rent lowers the cost of bowling and raises living standards for everyone who bowls. Spending on bowling is a small share of a high income and a larger share of income for people who make less, so the effect flattens income. When the little rent pieces go away, corporate profits decline. Money the average American gi

Cliff Sosin says American capital markets redeployed 3% of GDP into a highly speculative AI venture on the turn of a dime. Worry about a decline in American dynamism ran into a market that moved capital at a scale and speed no other country has matched. He allows the fitting share might be two or six.

An inability to move investment into productive resources can bind productivity growth, especially when technology is advancing quickly. Capital spending is tied to people doing different work. Every day, people wake up and go to Pennsylvania to build a data center. The buildout depends on infrastructure that moved an enormous amount of capital.

Taking risk on speculative ventures and slicing

Cliff Sosin says a person working with a model now beats either one alone. Computer-aided chess had a long stretch of freestyle chess, where the human plus the program out-competed the human alone and out-competed the computer alone, because each side thought differently and the mix was synergistic. Investing is in that era.

A person still understands things a model misses by having lived a life, and not everything is on the internet. Using a model and then judging makes better judgments while the gap remains, which he thinks will last a while. Later the gap can shrink until it does not matter. Models will not predict the future, because the future is intrinsically unpredictable, but the mo

Cliff Sosin says AI infrastructure gets pushed until it is overbuilt because a strong price signal is what makes a buildout of a huge technology possible at all. Super smart, talented people are focusing their lives on the project, and they will make, and deserve, lots of money. Without that money, people would not keep pushing.

American society looks like a household saving money for a house. Electricians, plumbers, and other resources go into the AI infrastructure, so costs rise elsewhere in the economy. The work keeps going because nobody really knows the right stopping point, and it continues until the buildout is overdone. Lead times and a winner's curse mean the last person to stop wi

Cliff Sosin says the limit arrives after the papers are finished. Models read all the papers and figure out connections that a person who had read all three papers and thought about them would have seen, though nobody had. All kinds of great results fall out of the reading. At some point the mining of published work is done, and what is left is a hypothesis about how to think about things outside what people already know. Even a super smart model will not necessarily be better at guessing the hypothesis. The model will still guess better than the smartest person, but the smartest person's odds are super low anyway.

The model can come up with great ideas for experiments, and the experiments

Cliff Sosin says a market shrinks once further price cuts stop creating enough demand, and a price of zero takes the market size to zero. Jevons paradox is only a narrow spot on a U-shaped demand curve, and markets in general take that shape. When a product is terrible, there is no demand. As the product gets better, people will pay a lot but will not buy very much. As the price goes down, the market grows.

Moving information once cost a huge amount, and people moved very little of it. A series of innovations cut that cost a lot, the amount of information moved rose sharply, and the market grew for a long time. Growth then stopped. The cost kept falling and the market kept shrinking. Cogent

Cliff Sosin says even superintelligent AI cannot predict most of what people want to know, because complex systems amplify tiny errors until a forecast falls apart. Intelligence fills in a thin scattering of facts. Mathematics can reach a closed form. Weather, mixing cream into coffee, and how groups of people interact cannot. A simulation carries error in the starting conditions or in each truncated step, and that error builds.

The weather model behind chaos theory is the concrete case. A forecaster ran it, reset it, and ran it again. The second run matched at first and then diverged. A few days out, one forecast was 70 degrees and sunny and the other was 20 degrees and pouring. The only c

Christina Cacioppo says total pay for the top person at Vanta is probably ten times what that seat paid five years ago, and she thinks it should be higher. Her rule is that you never pay the top people enough. You should always try, and you will not get there. Comp bands are a joke, because the survey data is already wrong. The number that matters is whatever those people are being offered now. What she is paying for is a high-context person who knows the company, the customers, and the domain, makes reasonable decisions, fixes the bad ones, and keeps moving.

The rate is still accelerating. The brake would be investors pushing back on dilution and stock-based pay. The industry had that argu

Christina Cacioppo says Vanta is eating the AI token cost and taking a lower margin. Customers notice a markup and compare it with the model itself, so marking tokens up is a comparison the product loses. About 25 percent of customers are already in the flow where an agent goes off and does the work, and she wants that at 75 percent. Retention and the experience come first. Price can wait until the usage is real.

The end state she wants is outcome pricing, getting paid for the result. Everyone she talks to wants that, and she likes the alignment from her years teaching microeconomics. Patrick O'Shaughnessy separates a closed support ticket, an output, from the outcome the customer actually

Christina Cacioppo says Vanta lost the billboard spot the company was known for because the deal was a placement order, not a multi-year contract. She had told podcast listeners the contract was locked up for years. It was not.

The billboard got famous, and people kept asking how Vanta did billboards. Referrals to one agency contact added up to millions of dollars of business. Years later, at a dinner, a startup friend said they had just signed the same spot. The supposed contract was written in crayon: there was no real contract. The contact called it a placement order. The other company paid more, so the spot was swapped out underneath Vanta. She went to that founder ready to pay almost a

Christina Cacioppo says security chiefs are much more likely to get budget for an enterprise risk overhaul today because the request is attached to AI. Over the past couple of years, Vanta has spent more time with government at the federal, state, and local levels inside the U.S. and outside the U.S. There is an appetite for modernization. It is a big stated priority of the current U.S. administration, and the same appetite exists outside the U.S. AI created more FOMO in government officials. They may have always known they did not have modern processes or modern software, but the gap feels more of an imperative now.

The conversation around AI has changed the priority and raised the priorit

Christina Cacioppo says AI models are far more persistent than people, so issues that look low severity on their own get chained into a real finding, and teams have to fix problems as fast as they find them.

Models will try nine quadrillion things where a person would give up after 9000. Some of what they surface are zero days people had not found before, or had not found in public. The bigger change is chaining 17 benign or only benign-ish issues into one real result. An OpenAI talk at RSA about Hugging Face showed the same pattern: persistence, then joining items a person would call low severity and plan to fix in 90 days. Someone would still wait 90 days and do other work first. Machines

Christina Cacioppo learned that a lot of the French software companies in Paris are using Chinese models predominantly. She was with customers in France a couple of weeks ago, and the pattern surprised her. She first thought the reason was nationalism, maybe an anti-American feeling, and she suggested that out loud. Someone corrected her very quickly: absolutely not.

The companies had all switched to Chinese models when Trump closed Fable off to non-Americans. They realized there is a huge business continuity risk. They cannot trust that if they build a business on one American model, the model will always be available. So they stopped doing that.

French software companies in Paris describ

Christina Cacioppo says the hard part of adopting AI that has paid off is simply getting people to use it, with personal use coming first.

Building something for the company on your own does not matter. What matters is knowing where the models are good, where they are bad, and what a harness means, and doing that check at least every two weeks because the tools change so quickly. A visceral sense of what works and what does not is what lets a leader answer when someone says they will fire all of the SDRs because they built one agent, and press on where that claim is true and where it is false. The aim is to get as many other people as possible into the same hands-on work.

A year ago, peopl

Christina Cacioppo says a business growing 80 percent with 80 percent margins returns more cash over a long stretch than one growing 160 percent with 30 percent margins, yet the faster business is valued at three or four times the multiple, and the gap can reach ten times. She calls living through that cognitive dissonance.

Patrick O'Shaughnessy sets up the market side: Vanta is still growing very fast, with the high margins that looked like the best model four years ago, and nothing about that got worse, yet the multiple on a business like it is a third or a quarter of the old level. She still wants to do what customers want, which may or may not be token pricing that pushes a company towa

Christina Cacioppo started Vanta about eight years ago to help companies build and earn trust. Building trust means starting security programs and compliance programs, or maturing the ones a company already has, then earning credit from customers through audits such as a SOC 2, ISO 27001, or HIPAA, plus security questionnaires and trust centers. The bet was that a security company could grow by helping companies put security in place because customers were already asking for it.

Vanta was initially known for SOC 2. A billboard on the 101 led her to joke that the company was mostly known for making billboards. The work started very compliance and audit centric. The audit was always just one