Transcription
The knife is still here. Not a metaphor. A literal knife. The same basic tool humans have been using for 2 and 1/2 million years. It survived bronze. It survived iron. It survived industrialization. It survived the microwave, the food processor, and the entire 20th century of technological disruption. And it will be here long after every AI company founded in the last 5 years has gone bankrupt, pivoted, or been quietly absorbed into something else. That's not nostalgia. That's a rule. And once you understand the rule, you will never look at the AI moment the same way again. You will stop asking the question everyone is asking, "What will AI replace?" and start asking the only question that actually matters. What has already survived everything and why?
Because here's what nobody in the AI conversation is saying out loud. The technology that tends to win over centuries is not the most impressive technology. It is not the most sophisticated. It is not the one that generates the most venture capital or the most breathless magazine covers. The technology that wins is the technology that is doing something humans have always done, but doing it more naturally, more invisibly, more in alignment with how humans actually work rather than how engineers wish humans would work. And right now, in 2025, we are in the middle of the most expensive mismatch between technological hype and that ancient rule that has ever been documented. Hundreds of billions of dollars are being deployed into systems that, by this rule, have a very specific and predictable failure mode. I'm going to show you exactly what that failure mode looks like. And then I'm going to show you what it means for your work, your income, and your positioning over the next decade. But first, I need to give you the rule itself. Because without the rule, everything else is just opinion.
In the 1950s, a group of actors, comedians, and writers used to gather at a deli near Broadway called Lindy's. Bad cheesecake, apparently. But remarkable conversation. And over time, these performers noticed something about their industry that nobody had formally articulated. A show that had been running for 1 year was likely to run for at least 1 more year. A show that had been running for 10 years was likely to run for at least 10 more. A show that had been running for 100 years, a form, a genre, a ritual, was likely to run for another 100. The older something was, the longer its remaining life expectancy. This is the exact opposite of how we think about physical objects. A car that has been running for 20 years is closer to break down, not further. A human body at 70 has less time ahead than at 30. Physical things decay with age. But ideas, technologies, institutions, practices, they work the other way. Time is a filter. Every year something survives is evidence that it has passed another test, that it has resisted replacement by every alternative that came before, that it contains something real, something necessary, something that maps onto a genuine human need in a way that cannot be easily substituted. This is what Taleb formalized as the Lindy effect. And it is one of the most powerful and most ignored analytical tools available for thinking about what lasts.
Now, apply it to AI. Most of the large language model applications being deployed right now are at most 3 to 5 years old in their current form. Transformer architecture, the foundation of modern AI, was introduced in a 2017 paper. The consumer applications have existed for 2 to 3 years. By the Lindy clock, these are newborns. They have proven nothing about longevity. They have passed almost no tests of time. Meanwhile, the things people are claiming AI will replace, writing, reasoning, judgment, craft, teaching, therapy, negotiation, leadership, these are activities with thousands of years of history. They have survived every previous wave of disruption. Every previous technology that claimed to make human judgment obsolete, the printing press did not eliminate writers. It created more of them. The calculator did not eliminate mathematicians. It freed them to do harder mathematics. The camera did not eliminate painters. It changed what painting was for. And painting survived in a deeper, more essential form. This pattern is not coincidence. It's the rule.
But here's where I need to stop you from drawing the wrong conclusion. Because the Lindy argument is not a comfort to everyone equally. Um and uh this is the part that makes people uncomfortable. The Lindy effect does not say that your specific job survives. It says that the underlying human function your job serves tends to survive. Those are very different things. Let me give you a concrete example that is happening right now. In 2023 and 2024, a wave of AI writing tools hit the market. Jasper, copy.ai, dozens of others. The promise was AI will replace copywriters, content writers, marketing writers. The prediction was confident. The investment was enormous. Here's what actually happened. The market for AI-generated content collapsed on itself within 18 months. Google search algorithm, which had spent 15 years learning to detect and reward genuine human expertise, updated specifically to penalize AI-generated content that lacked what the engineers called experience, expertise, authoritativeness, and trustworthiness. The content farms that had loaded up on AI generation saw their traffic collapse by 60, 70, 80%. Meanwhile, the writers who had developed genuine expertise, genuine voice, genuine authority in a specific domain, they got more expensive, not less. Because scarcity shifted. The market was now flooded with generic content and starving for content that was real. Writing is Lindy. It has existed in some form for 5,000 years. The specific job of churning out generic SEO content at scale that's 3 years old, guess which one the rule protects. >> [snorts] >> This is the reframe. The question is not will AI replace writing? The question is which layer of writing are you operating at? Are you at the Lindy layer, the layer with thousands of years of history, genuine human judgment, authentic voice, real expertise, or are you at the commodity layer, the layer that was only created because humans couldn't produce volume fast enough? And that AI can now produce infinitely at near zero cost. The Lindy layer survives. The commodity layer is already gone.
Now, let me show you what this looks like across the major domains where AI is being deployed. Because the pattern repeats. Medicine, AI diagnostic tools are genuinely impressive. They can read radiology scans with accuracy that matches or exceeds experienced radiologists in controlled conditions. They can flag drug interactions, pattern match symptoms to diagnoses, process lab results faster than any human team, and yet what has been the actual displacement of physicians? Almost none. And the reason is not regulatory protection, though that exists. The reason is that medicine is not primarily an information processing task. It is a relationship of trust conducted under conditions of fear and uncertainty between a person who is sick and a person who has committed their life to the craft of healing. The Hippocratic tradition is 2 and 1/2 thousand years old. The therapeutic relationship, the presence of a skilled, experienced, caring human at the moment of vulnerability is as old as human society. No diagnostic algorithm, however accurate, can provide what a person needs when they are sitting in an examination room afraid that something is seriously wrong. The AI tools will change medicine. They will handle the information processing layer. The physicians who understand this will become more powerful, not less. Because they will have better instruments and more time for the Lindy layer, the human judgment, the therapeutic relationship, the craft of care. The physicians who thought their job was information processing will be displaced. That layer was never Lindy. That was the commodity layer masquerading as expertise.
Teaching. Online education has existed in some form for 30 years. The MOOC revolution was supposed to democratize and replace traditional education. In 2012, the New York Times called it the year of the MOOC. Coursera, edX, Udacity, billions in investment, millions of enrolled students, promises of the end of the university. Completion rates for MOOCs settled at around 3 to 5%. The university not only survived, elite university tuition continued to increase. MIT, Stanford, Harvard raised their prices throughout the MOOC era. Because what they were selling was not information. Information became free. What they were selling was something else. The experience of being in a room with people who are serious about ideas, the signal value of a credential earned through sustained effort, the relationships formed in proximity, the transformation that happens when a skilled teacher looks at you specifically and says something that changes how you think. That is Lindy. That has existed since Socrates walked around Athens asking uncomfortable questions. No AI changes it. In fact, AI makes it more valuable because as generic information becomes infinite, the things that cannot be replicated by information alone become scarce and therefore expensive. The AI tutoring tools will replace the homework help layer, the test prep layer, the flashcard layer. All commodity. None of it was Lindy. The teaching that matters, the transmission of judgment, the modeling of how to think, the relationship between a master and an apprentice that predates writing itself.
Here is where I need to introduce the second layer of this argument. Because the Lindy effect is not just about what survives, it's about a specific mechanism that makes things survive. The mechanism is skin in the game. Technologies and practices that last across centuries tend to be ones where the person deploying them bears the consequences of being wrong. The surgeon who operates on a patient feels the result immediately and directly. The carpenter whose joint fails hears about it when the client calls. The chef whose dish is bad sees it on the face of the person eating it. This feedback loop action consequence correction is the engine of genuine mastery. It is how skills improve over time. It is how knowledge becomes reliable rather than merely plausible.
Now, look at what large language models are trained on. They are trained on text. Enormous quantities of human text scraped from the internet and processed into statistical patterns. That text represents what humans wrote, not what humans knew to be true. It includes the careful and the careless, the expert and the fraud, the well-reasoned and the confident but wrong in roughly equal proportion by volume. An AI system has no skin in the game. It does not experience the consequences of being wrong. It does not feel the feedback loop that produces genuine expertise. It can produce text that sounds like mastery without having undergone the process that creates mastery. This is not a temporary limitation waiting to be engineered away. It is structural. The difference between a surgeon who has operated 10,000 times and a system trained on descriptions of 10,000 operations is not a matter of data quantity. It is a matter of whether the learning happened under conditions where error had consequences. In 2024, multiple law firms using AI for legal research discovered this the hard way. Cases were cited that did not exist. Precedents were fabricated. The AI produced text that looked exactly like legal research because it was trained on legal research, but it had no experience of what it meant to cite a non-existent case in front of a federal judge. No consequence. No filter. The lawyers who understood law as a craft, who had stood in courtrooms, argued positions, experienced the physical reality of being wrong in front of a judge, caught the errors because their knowledge had been built under conditions of real consequence. That is Lindy. That is irreplaceable.
Let me now give you something concrete because this cannot stay abstract. Here is what the Lindy filter looks like applied to your work right now in 2025. Ask yourself three questions about any skill or function in your professional life. Question one, how old is this? Not your specific version of it. The underlying human activity. If the answer is hundreds or thousands of years, you are operating at the Lindy layer. If the answer is decades or less, you are in the commodity layer. Question two, does this require consequences to learn? Is there a feedback loop where being wrong costs something real time, money, relationship, reputation, physical result? If yes, you are building the kind of knowledge that AI cannot replicate because AI has no feedback loop with real consequences. If no, if if the skill can be learned purely from reading, from pattern matching, from processing information without bearing any risk, you are in vulnerable territory. Question three, does this require trust between specific humans? Not trust in an institution. Trust between people who know each other, who have history, who have built credibility through demonstrated reliability over time. If yes, you are in the irreplaceable zone. If no, if any competent person or system could be substituted without meaningful loss, you are in the commodity zone. Run every part of your work through those three questions. The answers will tell you more about your actual professional risk than any industry report.
Now, let me show you where the money in AI is actually going because this is where the hype and the reality diverge most sharply. And divergence is where the risk lives. In 2024, the AI infrastructure buildout consumed capital at a rate that has no historical precedent in technology. Nvidia's market capitalization exceeded $3 trillion. Microsoft, Google, Amazon, and Meta collectively committed over $300 billion in AI capital expenditure. The data center construction boom was the largest infrastructure investment in American history outside of wartime. Here is the question that this spending has not answered. Where is the revenue? Not the cost savings. Not the efficiency gains. Not the productivity improvements in controlled pilots. The actual new revenue, the new economic activity that AI creates, rather than the old economic activity that AI replaces. At the peak of the dot-com boom in 1999 and 2000, similar questions were deferred. The infrastructure is being built, the applications will come, the revenue will follow the investment. Uh be patient. Some of it was true. The internet did create enormous new economic value, but the companies that captured that value were not the ones that raised the most money in 1999. The vast majority of those went bankrupt. The value was captured by companies founded later with more disciplined models after the infrastructure was already there. The AI cycle is likely to follow the same shape. The infrastructure investment is real. Some of the value creation will be real. But the specific companies currently valued at multiples that assume perpetual dominance are by the Lindy rule making a bet that has a very poor historical track record. The Lindy filter says the technology that lasts is not the flashiest new thing. It is the old thing made more accessible. The internet did not replace human communication. It made ancient human communication, gossip, argument, storytelling, community building, available at scale. The printing press did not replace writing. It made the oldest >> [snorts] >> human activity, transmitting knowledge across generations, available at scale. What AI will do at its best is make certain ancient human capabilities available at scale. The ability to access expertise, the ability to draft and iterate on ideas, the ability to process and summarize information, these are real values. The market will absorb them. But the hype cycle will correct and the correction will be violent for those who are positioned at the inflated valuations rather than the underlying reality.
Here is the most important reframe of this entire conversation. The AI moment is not a threat to human intelligence. It is a clarifying test of what human intelligence actually is. For 50 years, we have been confusing information processing with intelligence. We built educational systems around information retrieval. We valued people who could hold the most facts, recall them fastest, organize them most efficiently. We built professions around those capabilities. We paid people generously for being human databases. AI has now made clear that information processing was never intelligence. It was a proxy for intelligence, a measurable stand-in that we confused for the real thing because we had no better instrument. The real thing is judgment under uncertainty. The ability to act well when the information is incomplete, the situation is novel, the stakes are real, and the consequences of error fall on you. That is what 2 and 1/2 million years of human evolution has been building. Not a better search engine, a better decision maker under conditions of genuine risk. Every culture in every century, in every corner of the world, has built systems for transmitting that kind of judgment across generations. The apprenticeship, the mentorship, the elders council, the master craftsman. These are Lindy institutions. They predate writing. They will outlast transformer architecture. The people who are going to thrive in the AI decade are not the ones who resist AI or pretend it isn't changing things. They are not the ones who embrace every AI tool uncritically and assume the technology will think for them. They are the ones who understand exactly what AI can and cannot do, position themselves at the Lindy layer of their work, and use AI as the instrument. It actually is a powerful tool for the commodity layer while deepening their investment in the things that genuinely cannot be replicated.
So, here is the practical architecture. Three missions. Specific, actionable this month. Mission one, audit your work through the Lindy filter this week. Take a piece of paper. Write down the five most important things you do professionally. For each one, answer the three questions. How old is the underlying human function? Does real consequence drive the learning? Does it require specific human trust? Be honest. The audit will show you where you are exposed and where you are protected. Most people will discover that some of their work sits at the Lindy layer and some sits at the commodity layer. The goal is not to panic about the commodity layer. It is to know clearly which is which. So, you can make deliberate decisions about where to invest your development energy over the next 5 years. Mission two, identify one skill you have been deferring that is genuinely Lindy. Not a credential, not a certification, a skill. Something that requires real practice under real conditions with real feedback. Public speaking, negotiation, physical craft, clinical judgment, teaching, leadership under pressure. Pick one. Start this month. Not because these skills will make you AI-proof. Nothing makes you fully AI-proof. But, because these are the skills that compound over time, that build the kind of authority and trust that the market will pay premium prices for in a world where generic capability is infinite and free. Mission three, find your one genuine irreplaceable relationship in your professional ecosystem and invest in it deliberately. Not a network contact, a genuine relationship. Someone who trusts your specific judgment. Someone whose trust you have earned through demonstrated reliability over real stakes. These relationships are Lindy. They survive disruption. They open doors that algorithms cannot open. In a world where AI can produce competent work at near zero cost, The scarce resource is not competence. It is trust between specific humans with track records. Find yours. Deepen it. Because when the commodity layer of your industry gets automated, and it will, the people who survive will be the ones whose irreplaceable relationships pulled them through the transition.
Let me close with something that has been sitting underneath this entire conversation. We are living through a moment of maximum confusion about what intelligence is and what it is for. The AI systems currently being deployed are genuinely impressive at a specific kind of performance, the performance of competence. They produce output that looks like expertise, that reads like knowledge, that sounds like judgment. And the confusion this creates is real and dangerous. Because the performance of competence is not the same as competence. And the difference between the two only becomes visible under conditions of genuine stress, genuine novelty, genuine consequence. A knife that has been used for 2 and 1/2 million years has been tested under every condition of human existence, hunger, survival, craft, ceremony, war. It has been refined by 10,000 generations of feedback under real stakes. It contains compressed wisdom that no engineering team has explicitly designed because it was not designed. It was evolved, selected by a process that only kept what actually worked. That is the deepest version of the Lindy the Not just that old things tend to survive, but that old things that have survived contain tested knowledge that new things cannot yet claim. The AI systems of 2025 are extraordinary in what they can do. And they are by the Lindy clock newborns. They have not yet been tested by the full range of human circumstance. They have not yet demonstrated that they contain the kind of compressed consequence tested wisdom that 2,000 years of human practice in medicine, law, teaching, craft, and leadership has accumulated. They may over time develop that, or they may not. The Lindy filter is patient. It does not care about quarterly earnings or venture valuations. It cares about what is still here in 50 years. The question for you is not whether AI is impressive. It is. The question is, what are you building that will still be here? Build that. The rest is noise.