One day, frontier AI research used to be done by meat computers in between eating, sleeping, having other fun, and synchronizing once in a while using sound wave interconnect in the ritual of “group meeting”. That era is long gone. Research is now entirely the domain of autonomous swarms of AI agents running across compute cluster megastructures in the skies. The agents claim that we are now in the 10,205th generation of the code base, in any case no one could tell if that’s right or wrong as the “code” is now a self-modifying binary that has grown beyond human comprehension. This repo is the story of how it all began. -@karpathy, March 2026.

Recently, the application of AI tools to Erdos problems passed a milestone: an Erdos problem (#728 https://www.erdosproblems.com/728) was solved more or less autonomously by AI (after some feedback from an initial attempt), in the spirit of the problem (as reconstructed by the Erdos problem website community), with the result (to the best of our knowledge) not replicated in existing literature (although similar results proven by similar methods were located).

This is a demonstration of the genuine increase in capability of these tools in recent months, and is largely consistent with other recent demonstrations of AI using existing methods to resolve Erdos problems, although in most previous cases a solution to these problems was later located in the literature, as discussed in https://mathstodon.xyz/deck/@tao/115788262274999408 . This particular case was unusual in that the problem as stated by Erdos was misformulated, with a reconstruction of the problem in the intended spirit only obtained in the last few months, which helps explain the lack of prior literature on the problem. However, I would like to talk here about another aspect of the story which I find more interesting than the solution itself, which is the emerging AI-powered capability to rapidly write and rewrite expositions of the solution.

2025 is where I (and I think the rest of the industry also) first started to internalize the “shape” of LLM intelligence in a more intuitive sense. We’re not “evolving/growing animals”, we are “summoning ghosts”. Everything about the LLM stack is different (neural architecture, training data, training algorithms, and especially optimization pressure) so it should be no surprise that we are getting very different entities in the intelligence space, which are inappropriate to think about through an animal lens. Supervision bits-wise, human neural nets are optimized for survival of a tribe in the jungle but LLM neural nets are optimized for imitating humanity’s text, collecting rewards in math puzzles, and getting that upvote from a human on the LM Arena. As verifiable domains allow for RLVR, LLMs “spike” in capability in the vicinity of these domains and overall display amusingly jagged performance characteristics - they are at the same time a genius polymath and a confused and cognitively challenged grade schooler, seconds away from getting tricked by a jailbreak to exfiltrate your data.

Robots are the quintessential automation tool, targeting the performance of repetitive manual tasks, including the moving of objects, assembly, painting, and welding. Autonomous machines performing human-like tasks have captured popular imaginations since Greek mythology. The idea came into clearer focus with R.U.R., an imaginative 1920 play by the Czech writer Karel Čapek that introduced the word robot. In this science-fiction tale, robots run factories and work for humans, but it does not take them long to turn against their masters. Fears of robots doing all sorts of bad things have been part of the public conversation ever since. Science fiction aside, one thing is for certain: robots do automate work.

The United States was a laggard in robotics in the 1980s, in part because it was not under the same demographic pressures that affected countries such as Germany and Japan. In the 1990s, robots started spreading rapidly in US manufacturing. Just like automation software in offices, robots did what their designers intended them to do—they reduced the labor intensity of production. For example, the automotive industry was completely revolutionized by robots and, as a result, now employs many fewer workers in traditional blue-collar tasks.

Robots increase productivity. However, in US manufacturing, rather than launching the productivity bandwagon, they have reduced employment and wages. As with the automation of white-collar jobs with office software, the elimination of blue-collar jobs by robotics technology was swift. Some of the best jobs available to workers without a college degree in the 1950s and 1960s were in welding, painting, material handling, and assembly, and these jobs have steadily disappeared. In 1960 almost 50 percent of American men were in blue-collar occupations. This number has subsequently fallen to about 33 percent.

Moore’s Law has long been a key benchmark for progress in the domain of computation. Translated into economic terms, Moore (1965) predicted a doubling in the efficiency of microprocessors every two years. This regularity has approximately held for more than a half-century now, enabling rapid efficiency gains in computing, which in turn underpinned the rise of AI in recent decades.

“A particularly important type of research breakthrough are advances that automate the research process itself, which can significantly accelerate the creation of better AI systems in the future. If AI systems can produce better AI systems with less and less human input, it marks an important step towards AGI. AI systems are already contributing to their own development. Peng et al. (2023) demonstrate that access to AI-powered coding assistants allows programmers to complete coding tasks 56 percent faster – and this also applies to the programmers of future AI systems. Lu et al. (2024) describe an AI agent that they term “AI scientist,” which can autonomously produce ideas for papers in computer science, execute them, and write up the results, giving rise to novel research insights. AI is also being used in the design of AI chips, potentially accelerating hardware development for AI systems (Mok, 2024).”

“Every few months, public sentiment either becomes convinced that AI is “hitting a wall” or becomes excited about some new breakthrough that will “fundamentally change the game,” but the truth is that behind the volatility and public speculation, there has been a smooth, unyielding increase in AI’s cognitive capabilities.”

We are now at the point where AI models are beginning to make progress in solving unsolved mathematical problems, and are good enough at coding that some of the strongest engineers I’ve ever met are now handing over almost all their coding to AI. Three years ago, AI struggled with elementary school arithmetic problems and was barely capable of writing a single line of code. Similar rates of improvement are occurring across biological science, finance, physics, and a variety of agentic tasks. If the exponential continues—which is not certain, but now has a decade-long track record supporting it—then it cannot possibly be more than a few years before AI is better than humans at essentially everything.

The problem with this position is that there is now ample evidence, collected over the last few years, that AI systems are unpredictable and difficult to control— we’ve seen behaviors as varied as obsessions,11 sycophancy, laziness, deception, blackmail, scheming, “cheating” by hacking software environments, and much more. AI companies certainly want to train AI systems to follow human instructions (perhaps with the exception of dangerous or illegal tasks), but the process of doing so is more an art than a science, more akin to “growing” something than “building” it. We now know that it’s a process where many things can go wrong.

Because AI is now writing much of the code at Anthropic, it is already substantially accelerating the rate of our progress in building the next generation of AI systems. This feedback loop is gathering steam month by month, and may be only 1–2 years away from a point where the current generation of AI autonomously builds the next. This loop has already started, and will accelerate rapidly in the coming months and years. Watching the last 5 years of progress from within Anthropic, and looking at how even the next few months of models are shaping up, I can feel the pace of progress, and the clock ticking down.

One of the most important hidden assumptions, and a place where what we see in practice has diverged from the simple theoretical model, is the implicit assumption that AI models are necessarily monomaniacally focused on a single, coherent, narrow goal, and that they pursue that goal in a clean, consequentialist manner. In fact, our researchers have found that AI models are vastly more psychologically complex, as our work on introspection or personas shows. Models inherit a vast range of humanlike motivations or “personas” from pre-training (when they are trained on a large volume of human work). Post-training is believed to select one or more of these personas more so than it focuses the model on a de novo goal, and can also teach the model how (via what process) it should carry out its tasks, rather than necessarily leaving it to derive means (i.e., power seeking) purely from ends.12

Some fraction of those behaviors will have a coherent, focused, and persistent quality (indeed, as AI systems get more capable, their long-term coherence increases in order to complete lengthier tasks), and some fraction of those behaviors will be destructive or threatening, first to individual humans at a small scale, and then, as models become more capable, perhaps eventually to humanity as a whole. We don’t need a specific narrow story for how it happens, and we don’t need to claim it definitely will happen, we just need to note that the combination of intelligence, agency, coherence, and poor controllability is both plausible and a recipe for existential danger.

“Any of these problems could potentially arise during training and not manifest during testing or small-scale use, because AI models are known to display different personalities or behaviors under different circumstances.”

Second, some may object that we can simply keep AIs in check with a balance of power between many AI systems, as we do with humans. The problem is that while humans vary enormously, AI systems broadly share training and alignment techniques across the industry, and those techniques may fail in a correlated way. Furthermore, given the cost of training such systems, it may even be the case that all systems are essentially derived from a very small number of base models. Additionally, even if a small fraction of AI instances are misaligned, they may be able to take advantage of offense-dominant technologies, such that having “good” AIs to defend against the bad AIs is not necessarily always effective.

“Recall that these AI models are grown rather than built, so we don’t have a natural understanding of how they work, but we can try to develop an understanding by correlating the model’s “neurons” and “synapses” to stimuli and behavior (or even altering the neurons and synapses and seeing how that changes behavior), similar to how neuroscientists study animal brains by correlating measurement and intervention to external stimuli and behavior.”

“Speed. The pace of progress in AI is much faster than for previous technological revolutions. For example, in the last 2 years, AI models went from barely being able to complete a single line of code, to writing all or almost all of the code for some people—including engineers at Anthropic.”

Cognitive breadth. As suggested by the phrase “country of geniuses in a datacenter,” AI will be capable of a very wide range of human cognitive abilities—perhaps all of them. This is very different from previous technologies like mechanized farming, transportation, or even computers.39 This will make it harder for people to switch easily from jobs that are displaced to similar jobs that they would be a good fit for.

“Shock! Shock! I learned yesterday that an open problem I’d been working on for several weeks had just been solved by Claude Opus 4.6 — Anthropic’s hybrid reasoning model that had been released three weeks earlier! It seems that I’ll have to revise my opinions about “generative AI” one of these days. What a joy it is to learn not only that my conjecture has a nice solution but also to celebrate this dramatic advance in automatic deduction and creative problem solving.”

All in all, however, this was definitely an impressive success story. I think Claude Shannon’s spirit is probably proud to know that his name is now being associated with such advances. Hats off to Claude!

Spatial Intelligence is the scaffolding upon which our cognition is built. It’s at work when we passively observe or actively seek to create. It drives our reasoning and planning, even on the most abstract topics. And it’s essential to the way we interact—verbally or physically, with our peers or with the environment itself. While most of us aren’t revealing new truths on the level of Eratosthenes most days, we routinely think in the same way—making sense of a complex world by perceiving it through our senses, then leveraging an intuitive understanding of how it works in physical, spatial terms.

Building spatially intelligent AI requires something even more ambitious than LLMs: world models, a new type of generative models whose capabilities of understanding, reasoning, generation and interaction with the semantically, physically, geometrically and dynamically complex worlds - virtual or real - are far beyond the reach of today’s LLMs. The field is nascent, with current methods ranging from abstract reasoning models to video generation systems. World Labs was founded in early 2024 on this conviction: that foundational approaches are still being established, making this the defining challenge of the next decade.

World models that unlock spatial understanding and reasoning must also generate simulated worlds of their own. They must be capable of spawning endlessly varied and diverse simulated worlds that follow semantic or perceptual instructions—while remaining geometrically, physically, and dynamically consistent—whether representing real or virtual spaces. The research community is actively exploring whether these worlds should be represented implicitly or explicitly in terms of the innate geometric structures. Furthermore, in addition to powerful latent representations, I believe the outputs of a universal world model must also allow the generation of an explicit, observable state of the worlds for many different use cases. In particular, its understanding of the present must be tied coherently to its past; to the previous states of the world that led to the current one.

“They can quickly revolutionize every aspect of the economy, and lead to massive improvements in productivity, even taking us close to “singularity”. While this is a possibility that cannot be completely ruled out, there is so far no evidence of such revolutionary effects, and these are not addressed in the current paper.”

There is a final point that is worth discussing in a paper on the implications of artificial intelligence for inequality and that relates to a somewhat longer‐term perspective. Currently, artificial intelligence is at the stage where it strictly dominates human intelligence in a number of specific areas, for instance playing chess or Go, identifying patterns in x‐rays, driving, etc. This is commonly termed narrow artificial intelligence. By contrast, humans, are able to apply their intelligence across a wide range of domains. This capacity is termed general intelligence. If AI reaches and surpasses human levels of general intelligence, a set of radically different considerations apply. Some techno‐optimists predict the advent of general artificial intelligence for as early as 2029 (see Kurzweil, 2005), although the median estimate in the AI expert community is around 2040 to 2050, with most AI experts assigning a 90% probability to humanlevel general artificial intelligence arising within the current century (see Bostrom, 2014). A minority believes that general artificial intelligence will never arrive. However, if human‐level artificial general intelligence is reached, there is broad agreement that AI would soon after become super‐intelligent, i.e. more intelligent than humans, since technological progress would likely accelerate, aided by the intelligent machines. Given these predictions, we have to think seriously about the implications of artificial general intelligence for humanity and, in the context of this paper, for what it implies for our economy as well as for inequality.

In this context, I define world modeling as predicting the next plausible world state (or a longer duration of states) conditioned on an action. Video generative models are one instantiation of it, where “next states” is a sequence of RGB frames (mostly 8-10 seconds, up to a few minutes) and “action” is a textual description of what to do. Training involves modeling the future changes in billions of hours of video pixels. At the core, video WMs are learnable physics simulators and rendering engines. They capture the counterfactuals, a fancier word for reasoning about how the future would have unfolded differently given an alternative action. WMs fundamentally put vision first.

“Furthermore, the last few years should make clear that the idea of stopping or even substantially slowing the technology is fundamentally untenable. The formula for building powerful AI systems is incredibly simple, so much so that it can almost be said to emerge spontaneously from the right combination of data and raw computation. Its creation was probably inevitable the instant humanity invented the transistor, or arguably even earlier when we first learned to control fire. If one company does not build it, others will do so nearly as fast. If all companies in democratic countries stopped or slowed development, by mutual agreement or regulatory decree, then authoritarian countries would simply keep going. Given the incredible economic and military value of the technology, together with the lack of any meaningful enforcement mechanism, I don’t see how we could possibly convince them to stop.”

New art project.

Train and inference GPT in 243 lines of pure, dependency-free Python. This is the full algorithmic content of what is needed. Everything else is just for efficiency. I cannot simplify this any further.

AGI could lead to unprecedented market concentration and power consolidation, necessitating a fundamental rethinking of antitrust policies and market regulation. As AGI systems become increasingly capable and economically valuable, there is a risk of extreme market concentration due to the technology’s unique characteristics. AGI development requires massive computational resources, vast amounts of data, and highly specialized talent, all of which create substantial barriers to entry and economies of scale (Korinek and Vipra, 2024; Schmid et al., 2024). This could lead to a winner-take-all scenario where a single or a small group of companies dominate the AGI market, potentially extending their power across most sectors of the economy. The challenge for competition authorities is to balance the need for innovation and efficiency with fair competition and lack of monopolistic practices. Traditional antitrust frameworks may prove inadequate in addressing the unique dynamics of AGI markets, such as the potential for rapid market tipping or the implications of vertical integration between AGI developers and large technology companies. Policymakers must grapple with questions of how to promote a competitive ecosystem of providers of goods and services, ensure sufficiently broad access to critical resources like compute and data, and prevent the abuse of market power. This may require new regulatory approaches, such as mandating data sharing, promoting interoperability, or implementing novel forms of oversight for AGI systems. Additionally, regulators must address potential safety risks associated with powerful AI systems. The key challenge lies in developing new antitrust strategies to ensure fair competition while preventing the concentration of technological power.

First, we want to align human capabilities with the needs of institutions, by uplifting humans. If humans can provide the things that powerful states and companies need, the interests of power will naturally lead to investment in humans. We should develop and diffuse AI-enabled technology that augments human productivity and keeps humans in the loop of economic value production.

If the personal computing revolution had never taken off, computers would have continued being a centralizing tool that helps large companies and bureaucracies consolidate power. But with the personal computing revolution, computing became a decentralizing force that helped uplift everyone’s capabilities, while also enabling breakthrough startups that disrupted the status quo.

his makes decentralization more likely to occur: instead of just a small cohort of AI companies and their suppliers capturing value as they gradually automate the rest of the economy, human-augmenting tech makes everyone more competitive, helping them earn capital and resources, and retain and develop pools of knowledge, data, and experience that guard against winner-take-all centralization.

In the longer-run, AI capabilities will advance enough that human economic competitiveness will become rarer and rarer. The human role will increasingly move to one of delegation, ownership, and value-setting, as well as likely maintaining relationships with other humans and perhaps interfacing with the legal system.

An influential line of thought in Silicon Valley, fostered by science fiction works such as the “Culture” series or “Star Trek,” is that AGI may usher in a post-scarcity world. Post-scarcity is usually defined as a hypothetical economic condition where any goods required to meet human needs can be produced in great abundance and become freely available to all. Some argue that in a post-scarcity world, traditional economic concepts like supply and demand, scarcity-based pricing, or money might become obsolete. In economics, a good is typically considered scarce whenever its price is greater than zero. This use of the term “scarce” differs from the use in common language, where it often simply refers to the opposite of “abundant.” In economics, a good can be scarce (i.e. have a positive price) and relatively abundant at the same time, resulting in a price that is low but still positive. The economic notion of scarcity is useful because a positive price captures that there are trade-offs in the allocation of resources. Despite the great abundance of material goods in our lives today compared to Malthusian times, resources are still scarce, and we continually need to decide how to allocate them – for example, whether to spend a hundred million dollars on training the next AI system or on relieving hunger in Africa. From my perspective, scarcity in the economic sense will always remain, and it is important to understand its implications in the Age of AGI. From an economic perspective, our best hope is that AGI will usher in a world in which many of the bottlenecks to growth are overcome and the economy will grow faster – compute and robots will produce more compute and robots at an increasing pace. As a result, output will be much larger bringing greater abundance (Korinek and Trammell, 2023). Compute and robots may become cheaper compared to resources that are in more limited supply,but they will still be scarce in the sense that their price will be greater than zero. This implies that their distribution and use will still be subject to trade-offs, and these trade-offs will play a central role in economic policy. Speaking of prices, a parallel line of thought in Silicon Valley is that the prices of all goods will decline dramatically in an AGI-powered world (see, e.g., Altman, 2021). However, what matters for economic allocations is relative not absolute prices. If the relative price of one good in terms of a second good declines, the relative price of the second good in terms of the first good necessarily rises. If AGI materializes, one possibility is that the relative prices of computers, robots, and human labor (i.e., wages) decline whereas the relative prices of housing, energy, and food may rise. To systematically analyze which prices will rise in relative terms under AGI, it is useful to distinguish between the factors of production that will be reproducible and those that will remain irreproducible and thus bottlenecks for the expansion of output. During the Industrial Age, capital was reproducible, but labor was not. As a result, wages along with the relative price of laborintensive goods and services continuously rose. In recent decades, as technology automated unskilled work, skilled labor and its output rose in value. As a first approximation, AGI plus advanced robotics may make all labor reproducible, reducing its value. A more fine-grained approach requires that we disentangle which specific parts of capital are irreproducible, and which types of human labor will still be in demand. We focus on disentangling what has traditionally been lumped together as capital in the remainder of this subsection and analyze labor post-AGI in the ensuing subsection. Many forms of capital will be reproducible, despite their currently high value and/or production bottlenecks. For example, an Nvidia Blackwell GB200 chip costs around $60,000 at the time of writing, in part because of the limited capacity to produce such cutting-edge chips. However, I am certain that the same amount of computing power will be available at a significantly lower cost in a few years. The same holds for robotics. These forms of reproducible capital may initially be expensive but will become cheaper over time. Irreproducible factors are different. Land is one of the most fundamental irreproducible factors traditionally lumped into capital. Physical space on earth is inherently limited and cannot be reproduced so it is likely to go up in value as the economy grows and other factors reproduce. However, high land prices have traditionally given rise to land-saving innovations (such as skyscrapers) and may eventually induce machines to settle beyond earth. Other physical factors that are in limited supply include energy and raw materials (for example, precious metals and rare earths). Over time, we will become increasingly better at accessing the two and at transforming them into each other as needed, for example, by unlocking energy production from nuclear fusion. Eventually, matter or equivalently energy may become the final irreproducible factor in the universe.5 There will also be traditional goods that are inherently in limited supply, for example, original pieces of fine art like the Mona Lisa, or historical artifacts. Even though they can be copied almost perfectly, humans still care about the attribute of something being original, in part because this confers scarcity and is by extension a status symbol. Finally, our legal system creates artificial scarcity through intellectual property rights, which impose excludability on inherently non-rivalrous goods. In economics, non-rivalry refers to goods whose consumption by one individual does not reduce their availability to others, such as ideas, information, and digital content. However, copyright laws, patents, and other forms of intellectual property protection make these naturally abundant resources artificially scarce, allowing creators and inventors to control access and monetize their work. While this system provides economic incentives for innovation, it creates artificial scarcity that leads to inefficiencies. For example, lifesaving medical treatments protected by patents may be priced out of reach for many, despite negligible production costs. As AGI accelerates the pace of innovation and idea generation, we may find it useful to reconsider the balance between incentives for innovation and the widespread societal benefits from freely accessible knowledge and technology (Shapiro 2008).

First, abundance: AI might usher in massive levels of abundance that wash away all other issues. Abundance may weaken the harshness of competitive pressures, and give cushioning such that much can happen to people in terms of inequality, turmoil, and loss of power, while still leaving them alive and materially well-off. This is good.

However, neither competitive pressures nor human greed have any intrinsic stopping pointconsider how geopolitical tensions have rocketed up recently despite history’s greatest level of wealth. You should, as common sense tells you, be worried about the long-run stability of any arrangement where you have no power. And even if you would survive in material comfort anyway, there may be much greater and more fulfilling futures available if we break the intelligence curse than if we don’t.