Why Free Intelligence Won’t Change Everything
In a Nutshell
Intelligence has hard limits because complex dynamic systems are fundamentally unpredictable—chaos, incomplete information, and physical constraints prevent any amount of reasoning from yielding perfect foresight. When intelligence becomes cheap, the scarce complements shift to data, physical presence, attention, and organizational control; most companies will struggle to use super-intelligent agents effectively, while a few like Meta are perfectly positioned. Markets will follow a U-shaped demand pattern where falling costs first expand usage, then shrink it once incremental intelligence delivers little extra value, ultimately favoring real-world experimentation and human judgment over pure computation.
These notes were generated by AI and may contain inaccuracies.
The common notion that intelligence evolves linearly from an amoeba to a monkey to a human to Einstein and then to godlike status is a cognitive error. This view overlooks the true limits of what intelligence can accomplish. Rather than treating intelligence as an unlimited quantity like the height of a tree, it should be understood as a ratio, similar to the roundness of a ball.
Intelligence functions as a tool to predict the affairs of the universe. We begin with a scattered collection of facts about a complex universe, and intelligence helps us connect these facts. However, some fields like mathematics are amenable to this approach, while most interesting questions involve complex dynamic systems that cannot be predicted through pure intelligence alone.
Complex dynamic systems, like the weather, cannot be precisely predicted through intelligence. Simulation attempts inevitably contain errors in initial conditions or subsequent steps, and these errors accumulate over time. This makes it impossible to predict outcomes in systems involving mixing cream in coffee, weather patterns, or interactions between groups of people.
The historical discovery of chaos theory emerged from a weather forecasting model that produced completely different results when initial conditions were rounded by one decimal place. This phenomenon, known as the butterfly effect, demonstrates that many systems of interest have characteristics where knowing all variables doesn't allow writing an easily solvable equation. Models remain fairly accurate unless one owns the entire universe, which is practically impossible.
The future will see everyone having access to teams of super-intelligent agents (equivalent to 200 IQ) working full-time in their pockets. These agents will identify small improvements that humans are too lazy to make. Companies will provide similar teams for employees, though many companies may not be positioned to benefit from this capability.
Consumer behavior currently reflects limited rationality and knowledge bases. Consumers accept "good enough" solutions rather than optimizing for maximum savings. With AI agents, continuous improvement in consumer favor becomes possible. Many software companies will face disruption as AI models serve as general-purpose intelligence tools, replacing crystallized and rigid software components.
When intelligence becomes cheap and widely available, the valuable complements become data, attention, love, affection, emotional connections, physical proximity, scope, and organizational control. Some companies' ability to employ intelligence is limited by economic factors, organizational size, or outdated structures.
Meta serves as an example of a company with perfect complements to cheap intelligence. The company's $250 billion in annual revenue and spending goes entirely toward intelligence-related activities: data centers, algorithms, code writing, and video analysis. As intelligence prices drop by a factor of 100 over five years, Meta's costs could approach zero while content quality improves dramatically through better creator tools and more targeted advertising.
Leading AI laboratories face increasing debate about the gap between publicly available models and privately used capabilities. Some argue for escape velocity where pioneering code becomes more valuable and laboratories can deliver the best version of cheap intelligence.
However, the prevailing trend shows rapid commoditization of intelligence. Chinese models demonstrate collapsing costs, and even with superior code, the delivery cost may not justify premium pricing in a competitive market. Programming has been a major use case, but the question remains whether there's room for expansion in subsequent intelligence layers.
The demand for intelligence may follow a diamond-shaped pattern where uses expand initially then diminish, with very few applications requiring super-elite intelligence levels while many uses exist for very good intelligence.
The Jevons paradox suggests that lower costs lead to increased usage, but this represents only one point on a demand curve. Markets typically follow a U-shaped pattern: poor quality products have no demand, improving quality leads to high prices with limited volume, falling prices increase market size, but further price drops eventually fail to increase demand sufficiently, causing market shrinkage.
The fiber optic communications example illustrates this pattern. Transmission costs dropped dramatically, data volumes increased enormously, markets grew for an extended period, then growth stopped despite continued cost reductions. Today, Cogent Communications routes 30-40% of all internet protocol traffic through one division with a market value of only a few billion dollars.
While models can generate many interesting ideas about physics through pure intelligence, they ultimately require real-world experiments to validate hypotheses. Current experimental capabilities remain limited, though this will enable a period of amazing discoveries as models identify connections across research papers that humans missed.
At some point, the limits of knowledge will be reached where even super-intelligent models cannot predict outcomes better than the already slim odds of the smartest humans. Models will generate valuable experiment ideas, but scaling these experiments will remain constrained by practical limitations.
The speaker notes that they used to be able to identify which AI model was better or worse, but this has become impossible due to automatic background optimization. Everything is now built without consuming subscriptions, and the user simply experiences that things work.
The discussion turns to real-world companies creating automated loops in robotics, biology laboratories, automated wet laboratories, and automated testing laboratories. The concept of closing the loop in the physical world is acknowledged as potentially speeding up learning processes.
The speaker explains that intelligence has limits, and therefore the ability to think of solutions alone is limited. Even if an automated biology testing laboratory connected to a computer could generate ideas, conduct experiments, and repeat the process, physical constraints remain. If an experiment requires raising mice that need a year to grow and die, the experiment will take ten years.
Laboratory space limitations mean that even with 5,000 great ideas, only 20 can be chosen. Even if an AI model is ten times better than the average researcher with only 1% probability of success, this remains a 1% success rate. The speaker emphasizes this is not pessimism but recognition of real constraints.
AI can discover things that should have been known but were forgotten in databases, guide research in better directions, and automate processes to speed up learning. However, the model will not suddenly identify a new world that humans cannot understand.
Private markets and investors are heavily focused on supply-side matters—how to build and service intelligence—rather than demand-related matters. The difficulty lies in predicting demand for inference, predicting how much reasoning will improve, and determining how much costs will decrease.
Even if supply and demand could be predicted, the value chain presents additional challenges. Profits require being positioned at a bottleneck, and competitors will eventually arrive at that same stage. The speaker jokes about deliberately ignoring these investment considerations.
The speaker describes using AI to create explanatory videos for their daughter rather than having chatbots solve math problems directly. This would not have been done if it cost $5. The concept of true price elasticity of demand is introduced, noting that AI demand will be very price-sensitive.
The smartest people and most successful entrepreneurs are investing heavily in supply expansion. The focus is on building infrastructure for data centers. The speaker questions why the best entrepreneurs devote their lives to building this infrastructure given uncertainties about supply, demand, and alternative profit opportunities.
The speaker suggests reversing the causal relationship: AI infrastructure development happens because very intelligent and talented people devote their lives to it, not the other way around. These people will make substantial money and deserve it.
Rather than betting on which railway magnate will profit, the speaker focuses on long-term equilibrium outcomes—once railways exist, what comes next? The speaker wishes for more foresight to understand how consumer product brands would emerge.
The current situation is compared to American society saving to buy a house, with resources dedicated to building massive AI infrastructure. Electricians, plumbers, and others are performing these tasks, driving up costs in other economic sectors.
Development will continue until exaggeration occurs because nobody really knows the limits. During this process, people must make huge amounts of money to continue exerting pressure. The last one to stop will certainly regret it due to winner's curse dynamics.
There is tremendous dynamism in technology unlike other historical infrastructure projects. AI could transform products like Nvidia into commodities through iterative self-improvement. Nvidia creates text files sent to Taiwan for manufacturing, which is extremely difficult but theoretically possible with AI assistance.
TSMC is discussed as having better features and having accomplished a difficult project requiring enormous processing power accumulated over years. Their competitive advantage is less amenable to AI development because it is very physical and relies on minute details difficult for AI to understand.
Companies with geographic proximity or large size might benefit from using intelligence to serve customers better or reduce costs. DoorDash and Uber are mentioned as companies with significant local density effects.
With AI agents, many decisions would be entrusted to them for spending optimization. Consumer choices exist on a graduated spectrum: from highest bank deposit interest rates (fully delegable to AI) to choosing a favorite sweater (purely personal judgment).
Garbage collection is presented as a service where AI agents could make decisions. People often don't know their current service costs. With AI agents negotiating, consumers could easily switch to lower-cost providers saving 50 cents monthly.
Previously, neighborhood coordination on single garbage collectors was difficult. With AI agents, the lowest-cost producer can approach all neighbors with better pricing. This creates better businesses with larger sales volume and lower profit margins.
The speaker questions whether the best AI agent could build investment portfolios including companies like Carvana, Hilton Grand Vacations, and Meta as effectively as a human. In the long run, humans will be excluded from markets due to competition.
The current era is compared to computer-assisted chess, where humans working with computers outperform either alone. Humans understand things differently and have information not in models. This collaboration period will continue until the gap shrinks to insignificance.
Models cannot predict the future because it is inherently unpredictable. Price is described as a miracle because it is a signal wrapped in a motivator—it shows relative value, scarcity, and production cost while incentivizing behavior.
Fanta automates compliance, reducing audit preparation time by 82% with 526% ROI reported by customers. Over 16,000 companies use Fanta, including Rump, Harvey, and Snowflake. Ridgeline is the only AI-powered investment management platform combining front, middle, and back-end operations, providing AI agents for automating repetitive tasks.
Dan's three levels of AI application are described: submitting requests and receiving responses, creating monitoring systems that run automatically, and proactive systems that continuously seek ways to use intelligence without limits.
Price stagnation is identified as a macroeconomic phenomenon where many things are priced slowly. With AI agents constantly improving pricing, garbage collection prices would move more frequently and accurately rather than annual fixed increases.
American capital markets redirected approximately 3% of GDP into high-risk AI projects. This resource redeployment capability is unique globally. The infrastructure enabling this capital transfer to data center construction in places like Pennsylvania is remarkable.
The basic concept of people taking risks in speculative ventures and dividing the pie differently is essential and irreducible. While fraud and bad behavior will eventually be uncovered, the scale suggests the process is relatively clean compared to historical railway construction projects.
The idea that anyone can buy $100,000 worth of stock and acquire a portfolio of highly efficient, highly motivated, and well-managed teams that build companies that do most of their work perfectly, almost instantly and at little cost, and then sell them once they're done, is amazing. Even better, all these companies, and other investment methods, contribute to the effective distribution of society's resources.
Ultimately, this is unbelievable. The entire system is being underestimated. Why do you think America has dominated this field? The personal theory is a kind of interaction between genes, cultural evolution, and institutional evolution. Those who came to the United States were among the most willing to take risks, because going to the United States was fraught with danger. Risk-taking has an important genetic component. Therefore, this tendency is more common among Americans. When you are in a place where taking risks is normal, it leads to a build-up of many other things. When risk-taking becomes normalized, it leads to an accumulation of cultural and institutional factors that reinforce it. It is understandable that people may take risks and fail. This leads to sensible bankruptcy laws, labor laws that support it, and venture capital that also supports it. All these things reinforce that, and then lead to success stories. Thus, a complete mix is obtained.
If a base were ever built on the moon, there would be optimism about that society's ability to achieve great economic success. The United States has always been incredibly innovative. So, if there is a base on the moon, it wouldn't be surprising if it was ours. Although the Chinese are intelligent and numerous, the Europeans are also present.
When interest rates are rising, as seen in the long-term chart of the 10-year interest rate in the United States, it looks like this for almost an entire lifetime. When a society builds and invests heavily in the future, it means it will consume less today. All the resources invested in developing artificial intelligence are future-oriented, just as a plumber who builds today does not consume today. He could have consumed enough for the day. Instead, it creates a data center, which represents the consumption of the future. Interest rates are higher because we encourage people to save for the future. The way you do that is by paying them.
The investment portfolio is approximately 80% Carvana, 15% holidays, and 5% Meta. It is essentially five things. One of them represents 80%.
When advising companies on how to use artificial intelligence to transform their businesses, the speech starts with making sure that everyone is interacting with the programming. Starting with the board of directors and the CEO, everyone needs to start building some—they need to organize hackathons. At Carvana, they call them pizza parties. Not just developers, but everyone needs hackathons. These events do not necessarily have to produce highly useful programs, but rather teach everyone what can be done. This will change their expectations of what can be done.
The CEO of a consumer electronics company needs to approach the head of the IT department and say: "I want my Grock bot, or Meta Muse, or any other program, to be able to access all of the company's systems on a read-only basis." That is, anything that can be looked for, for example, if someone is asked, the answer should be obtained immediately. The person will say: "This is really difficult. We have countless problems with that." The response should be: "That's what I mean. I want that." As it is known, it's a complex process and takes place at the level of each system individually, but if access can be obtained, read-only privileges will be gained. This is the starting point.
The starting point gives read-only privileges. Then the next logical thing is that the CEO or whoever else needs to use this everywhere, and everyone will be jealous, and the response will be: "Well, you can have it too, but we need to define your powers." Then, theoretically, it spreads gradually, and now there is read-only permission, and then someone will inevitably say: "I really want to enter data into the system." The response will be: "That's great. We need to define writing privileges." Then someone will say: "I would like to change the system because we rely on experimental programming, and I am confident that we can do it." The response will be: "That's great too. We need to put a mechanism in place to manage changes to the system." It is true that there are four points, but it takes years of work. It is a very difficult task. So, the question is: Do you have enough time in your field to achieve this before a new competitor gets ahead of you?
It is good to have a truly successful business. One of the conclusions seems to be that this is the best time ever to start a business, because of the capital market conditions that allow for pilot projects to be funded at a relatively low cost, and also because, with these tools, if you start without the burden, you can achieve great success very quickly. It is the best time to start a business even tomorrow. It is amazing. Dealing with people is very annoying, especially if you're running a business—they are a huge nuisance. The idea of having all these agents and others enables you to accomplish a lot with fewer people. One of the challenges is always that if you can do something easily, then everyone else can do it too. So, there has to be something that makes what you build produce sustainable value. Is there a path that ends with some kind of monopoly or proprietary data in a world where all customers happily change their providers every day—what gives you any sustainable economy?
If you do something that seems to have a clear path to a sustainable economy, it's bad because everyone is trying to do it. So, the competition is fierce, and the really great results are the ones that seem illogical. You shouldn't do it at all, but then it turns out there's a path the world is going down. The most underrated sources of future competitive advantage are attention, data, and physical proximity. These are examples of things that may be extremely valuable or even more valuable in the future than they are today because of cheap information. There is regulatory takeover. Many advantages are appropriately undervalued.
Everything points to profit margins decreasing. There are many companies that are achieving profit margins due to increased consumer surplus. Everywhere. If you go back and read many of the research reports from the 1990s, when the internet was just starting out, you'll find they pointed to the same thing. So, the baseline rate here might not be accurate. We might be wrong, but the main reason they're often wrong is aggregation economics. A lot of aggregation theories fall apart when you learn that they're based primarily on the idea that consumers only look at the first Google result. But it turns out that if your agent looks at everything in the search, then ranking at the top of the search results doesn't matter, and all the basics like quality, price, and service become much more important.
Another possible side effect is that it allows you to build more complex businesses. Right now, many industries are characterized by a horizontal structure, the complexity of managing your department is too high, making it difficult to handle tasks above and below you without unnecessary complexity. With AI agents and inexpensive, widely available software, you can build a fully vertical structure. Some are already trying to do this in other countries, creating a comprehensive pet care company that includes everything from pet food manufacturing to health insurance, veterinary hospitals, and clinics—all services that were previously multi-tiered. Now, thanks to AI, coordination issues are being addressed, and the problems of the primary agent are being reduced, potentially making this type of company viable. This could ultimately lead to higher profit margins, greater efficiency, and increased consumer surplus. Carvana is an example; it was built before this became so easy, but it could be applied to other sectors, potentially leading to higher profit margins and increased consumer surplus.
There are two stories regarding the impact on wealth and income distribution. On the one hand, many have talked about the idea of concentrating income distribution. There are many people who have made money. A plethora of different labs, data such as that from Nvidia, and the idea that there are increasing returns for small groups of people. But the counter-argument is that if we had cheap, universally available intelligence, two things would happen. The first is that there would be a lot of beneficiaries in the economy. Imagine you're a software company charging bowling allotments. Those people lose everything, which is bad for them, but it lowers the cost of bowling. So, everyone who bowls gets a higher standard of living. This creates a flattening effect because my bowling spending is a small percentage of my income, while people with lower incomes spend a larger percentage of their income on bowling. So, it's a flattening effect on income. If all these small returns disappeared, and corporate profits fell, and the money that the average American pays to Bank of America went away from the bank, then capitalists would lose, and the average consumer would win. That's the balance.
The other side of the coin is that there would be an entire professional class, whose role could be defined as an intelligence class. They take enormous amounts of information and apply it directly to specific circumstances. If these services become universally available, firstly, everyone will have access to the best, which is good for everyone, and secondly, it's especially good for people who can't afford the best right now. Thirdly, those individuals who earned these extra returns because they were smarter will be forced to resort to the next best alternative. Theoretically, this might flatten the curve. The response to that might be: Well, yes, but the returns on initiative, ambition, and intelligence might increase because I can now do more, because these models are like a boosted brain bike. They give smart, proactive people the ability to accomplish more. Perhaps this does indeed increase the returns for this group of people and accelerate the curve.
In many of the situations observed, there are a lot of overlapping currents. On the one hand, there are all these amazing ways AI can be used, using a huge number of tokens, and seeing how even more can be used, and being clearly on the right end of the distribution. If everyone used the same number of tokens, there would be a severe shortage. It would become a massive phenomenon that would last forever, and everyone involved in the build-up would make a fortune. On the other hand, the amount received per token is enormous, like going to the moon, and it turns out being an unusual person in many ways, and people will slowly adapt, and the token price will collapse. So maybe we've actually overbuilt. Both possibilities can be said with equal certainty.
A good friend is conducting a fascinating experiment, where if investing is broken down into choosing the assets to own, and then determining the quantity of each asset chosen, he's doing an experiment where he chooses the assets he wants to own, then manages the construction of his investment portfolio himself, as he used to do, in two ways: first, he then delegates this entire task to an AI system he developed, saying: "Build the portfolio, but stick to the assets you've chosen." He has found that this method far outperforms his own portfolio-building process, which is itself highly sophisticated. His simple argument is: "Well, AI is very good at optimization." They have read every research paper on portfolio building and all that stuff, from determining bet size to the Kelly Criterion, as if they know everything. Perhaps in the future, we will witness this shift where the investor's role is limited to choosing what to own, rather than determining the quantity of each asset.
When thinking about how AI will penetrate this space, it will start on short timescales and then gradually expand to longer timescales. One of the challenges is the lack of a validation mechanism. The friend's ability to confirm this is based on results over years, certainly not 20 years. AI's creep into the markets has already started. It wouldn't surprise if investors found themselves on short timescales, from seconds all the way up to minutes, hours, weeks, and months pass, and they themselves delegate more and more resources to the models. This makes perfect sense, as you can certainly imagine splitting that between trying to enrich the model. Others are already doing things like this, and this is just a more sophisticated version; but you could have a complete quantitative strategy, then enrich it with what you consider to be human-designed qualitative analysis, and then have a quantitative strategy to incorporate the qualitative analysis. Many of these multiple strategies do these things. What they're doing is just a slightly modified version of that. And this makes perfect sense. There are many ways to succeed, and models have added new dimensions of freedom.
Regarding comfort with having an investment position, a lot of research suggests this is an irrational way to build an investment portfolio. Much of that research focuses on price changes from the perspective of stock prices, rather than thinking about it from the perspective of a company owner. Take Carvana, for example. If there wasn't the feeling that Carvana's infrastructure, brand, and everything else about it were so exceptional that there weren't many scenarios where things could go wrong, there wouldn't be comfort owning all that stock. It could be as simple as that. Overthinking that a portfolio approach would have worked, but if companies like Walmart, Home Depot, or Fastenal had been discovered back then, there wouldn't have been benefit from them to their full potential. That's how it feels about Carvana. If you read history, you'll find that all successful companies focused on one area. But of course, history is full of information that makes you not want to read the history of people who haven't read their own. Much of that information is what most people looking for average or slightly above-average results should be doing, which is absolutely true. But that's not what is being looked for. Being a big fan of Carvana. It would never bother if the stock price went up significantly and there was a feeling like diversifying. That would be fine too. It's more about balancing expected returns with intentionally taking on unusual risk.
Regarding education and how institutions have contributed to the development of education in this country that have been really important in the track-dependent story, the thinking is leaning towards the idea that teaching methods for children are becoming more binary in terms of what we're looking for. What is the use of intelligence? Personal intelligence, your own intelligence. In a world with unlimited model intelligence, some might say: "It's useless. It doesn't matter. A person with an IQ of 50 can handle the world just as effectively as someone with an IQ of 150." That doesn't seem entirely true. It doesn't seem logical. But the crux of the matter is that the more intelligent you are, and the more mature and developed your intelligence is—that is, the more educated you are—the better you are at communicating with a model and telling it what you want. There would be a very difficult time having an in-depth dialogue with a model to create groundbreaking experiments in quantum physics because there would be nothing known about the field. The model might try very hard to explain things, but ultimately, it would explain them in a very simplistic way. It would be very difficult to access the latest technologies and make the most of what the model offers. Whereas in a field where there is some expertise, the model and the person can communicate effectively and guide it much better. So, it's very important, to guide the model—which will be a key skill in the future—to understand something about the world. You have to be intelligent, and education helps with that. You can't ask it to assume that something is a random process unless you know what a random process is. It's true that a program can teach you what it is, but this will hinder you considerably instead of allowing you to discover all this for yourself. You don't have to perform the calculations, but you do have to know what they are.
When thinking about education, the idea is to teach children a set of fundamental skills that build upon one another, and ultimately, they should be able to do all of these skills somewhat independently without models. Not because they will use them without models in the future, but because if they don't know how to do multiplication, it will be very difficult for them to understand matrix multiplication. And if they don't know what matrix multiplication is at a sufficient level, they don't have to be experts at it, but they should be able to understand what the model is doing as it's being executed. This is so they can guide the model. There is increasing inclination to believe that a lot of time should be spent with children, away from these models, teaching them the fundamental skills that are considered important. Identifying these skills is difficult, but the idea is that the model can do them, but the child needs to know them in order to guide the model.
Not liking using models because children cheat on their Shakespeare essays. The response should be: "You're not ambitious enough. If the kids are cheating on their Shakespeare essays, let's give them a harder task. Let's give them a task a hundred times harder and more ambitious, and let's make them use the models to their fullest extent. Shouldn't we? Let's have them produce the final work, and let's make sure they haven't just given the model ready-made instructions, but that they've taken responsibility for the work. Ask them why they chose that approach, shouldn't we? And make them justify it. And this is crucial; the kids need to master it, and frankly, they need to be taught that they can do anything."
Talking on the phone from time to time, every time wishing the call was recorded so others could hear it, because there is so much learning. This particular conversation, which is a transcript of a call had several months ago, gives a valuable set of insights into the future of intelligence and what that means for markets and businesses. It's not actually opposing, but rather offering a different perspective from the common one heard, which is the expansion of the Limited, forever short selling, and no upper limit on demand for intelligence.
Keep Invest Like The Best in your library
Save the videos and channels worth coming back to, and find them again in one place.
