3.28.2026

Andrej Karpathy on the Future of AI Agents, Research, and Society

 

Author: Claude AI, under the supervision, prompting and editing by HocTro

A comprehensive essay based on Andrej Karpathy's appearance on the No Priors podcast. Karpathy — co-founder of OpenAI, former Director of AI at Tesla, and founder of Eureka Labs — shares his firsthand experience with the agentic coding revolution, his vision for autonomous research, and what all of this means for jobs, education, and the structure of society itself.

Guest: Andrej Karpathy — AI researcher, educator, and founder of Eureka Labs.

Podcast: No Priors — hosted by venture partners covering frontier technology.

Table of Contents

Summary

In this wide-ranging conversation, Andrej Karpathy describes a dramatic shift that happened around December 2024: he went from writing 80% of his code by hand to writing essentially none of it. The agents, he says, simply got good enough. What follows is a cascade of implications that Karpathy is working through in real time — running multiple coding agents in parallel, building autonomous home automation systems controlled through WhatsApp, and setting up "auto research" loops that optimize machine learning models overnight without human involvement. He argues that the name of the game is now leverage: putting in very few tokens and getting a huge amount of work done on your behalf. Along the way, he addresses the jaggedness of current models (brilliant at code, terrible at jokes), predicts that AI models will eventually speciate into specialized forms rather than remaining monolithic, and makes the case that open-source models trailing a few months behind frontier labs is actually a healthy equilibrium for the industry. On jobs, he is cautiously optimistic — citing Jevons' paradox to argue that cheaper software will mean more demand for it, not less. On education, he believes the role of the teacher is shifting from explaining things to people to explaining things to agents, who then personalize the delivery. Throughout it all, Karpathy admits he is in a state of "AI psychosis" — the territory is vast, unexplored, and moving at a speed that makes even him nervous about falling behind.

I. "AI Psychosis" — The Capability Unlock

Karpathy opens the conversation by admitting he has been living in what he calls a "perpetual state of AI psychosis." Something fundamental changed around December 2024. Before that, his workflow was roughly 80% writing code himself, 20% delegating to AI agents. Within weeks, that ratio flipped — and kept going. By the time of the interview, he estimates he has not typed a single line of code since December. The speed of the shift, he emphasizes, is something most people outside the industry have not grasped. If you walked up to a random software engineer and looked at what they are doing at their desk, their entire default workflow for building software would be completely different from what it was just a few months earlier.

The feeling driving Karpathy's psychosis is twofold. First, the sheer power is intoxicating — things that used to take hours now happen in minutes. Second, the territory is completely unexplored. He does not know where the ceiling is, and neither does anyone else. He watches people on Twitter discovering new techniques and workflows and feels a deep anxiety about falling behind. The landscape of what is possible with AI agents is expanding faster than any single person can map it, and Karpathy — one of the foremost AI researchers in the world — feels the pressure acutely.

The hosts note that a team they work with at Conviction capital has already restructured their entire engineering workflow around agents. None of the engineers write code by hand. They are all microphoned and whisper instructions to their agents throughout the day. It sounds strange, but the hosts concede it turned out to be the right approach — those engineers were simply ahead of the curve.

II. Mastering Coding Agents — The New Workflow

When asked what limits his capacity now, Karpathy's answer is surprising: almost everything feels like a "skill issue." When something does not work, the instinct is not to blame the model's capability but to blame his own prompting, his instructions file, his memory setup. The tools are powerful enough; the bottleneck is learning how to wield them. This is actually empowering, he says, because it means you can get better. And that is what makes it addictive — every improvement in your own skill unlocks new capability.

Karpathy describes the workflow of Peter Steinberg, a developer he admires who has become something of a folk hero in the agentic coding community. Steinberg sits in front of a monitor running numerous Codex agent sessions simultaneously, each working on a different task across multiple repositories. Each agent takes about twenty minutes if prompted correctly with high effort settings. Steinberg moves between them, assigning work, reviewing output, and spinning up new tasks. The unit of work has shifted from "here is a line of code" or "here is a new function" to "here is a new piece of functionality — agent one, go build it; agent two, build this other thing that does not interfere with the first."

This parallelization is at the heart of the new workflow. One agent does research, another writes code, another develops a plan for a new implementation. Everything happens in these macro actions over the repository. The developer's job is to become excellent at orchestrating these macro actions — to develop a kind of muscle memory for managing agents in parallel, not for typing code character by character.

Karpathy notes a psychological element too. Whenever he is waiting for an agent to complete something, his instinct is to think: "I can do more work. If I have access to more tokens, I should parallelize and add more tasks." This creates a new kind of stress. If you are not limited by your ability to spend tokens, then you — the human — are the bottleneck in the system. You should be maximizing your token throughput. He compares it to the anxiety he felt as a PhD student when his GPUs were idle: you have compute capacity and you are not using it. The difference is that for the past decade, most engineers did not feel compute-bound. Now, suddenly, everyone does — except the resource they are competing for is not FLOPS but tokens.

III. Claws and Home Automation — Meet Dobby the Elf

Beyond coding, Karpathy describes a project that epitomizes the new paradigm. In January, during what he calls a period of "claw psychosis," he built an autonomous agent he named Dobby the Elf. Dobby's job is to manage his entire home.

It started simply. Karpathy told the agent he thought he had Sonos speakers at home and asked it to find them. The agent ran an IP scan of the local area network, found the Sonos system, discovered there was no password protection, and reverse-engineered the API endpoints through web searches. Within three prompts, music was playing in the study. Three prompts — from "can you find my Sonos?" to music playing. The agent repeated the process for his lights, discovering the APIs, creating a dashboard, and giving Karpathy command-and-control over every light in the house. When Karpathy texts Dobby "sleepy time," all the lights turn off.

Dobby now controls lights, HVAC, window shades, the pool and spa, and even the security system. For security, Karpathy has a camera pointed outside the house. When the system detects motion, it feeds the video to a Qwen vision model, which analyzes the scene and sends Karpathy a WhatsApp message — something like "Hey, a FedEx truck just pulled up. You might want to check it." The entire system is managed through WhatsApp. Natural language in, actions out.

Before Dobby, Karpathy used six completely different apps to control these systems. Now he uses none of them. He concedes he has not even pushed the paradigm fully — other people are doing far more elaborate things — but already, the consolidation is remarkable.

IV. The Death of Apps — Natural Language as the New Interface

The Dobby experiment points to a deeper shift. When asked whether this is indicative of what people actually want from their software, Karpathy says yes — with an important nuance. People already have a mental model of what AI should be. In their minds, it is a persona, an identity you can tell things to and it remembers, an entity behind a messaging app. That is a lot more intuitive than what a large language model actually is, which is a token generator. The work of building good agents is, in a sense, matching those human expectations with reality. Under the hood it is complex, but the interface should feel as natural as texting a friend.

The bigger implication, Karpathy argues, is that a huge number of the apps sitting in app stores today probably should not exist. Smart home device apps, fitness tracker apps, calendar apps — shouldn't these all just be APIs that agents call directly? An LLM can drive the tools, make the right API calls, and do complicated cross-system orchestration that no individual app can manage. He gives the example of his treadmill: there is an app for it, and he wanted to track his cardio, but he did not want to log into a web UI and navigate a flow. The whole thing should just be an API endpoint that an agent accesses on his behalf.

This amounts to a fundamental rewrite of who the customer is. The customer is no longer the human navigating a graphical interface. The customer is the agent acting on the human's behalf. The entire industry will need to reconfigure around this reality — and the refactoring, Karpathy says, will be substantial.

Some push back on this vision by asking: do we really expect normal people to vibe-code their own automation? Karpathy's response is that this is just the state of the technology today. What currently requires a technical person to set up will, in a year or two, be free — trivially easy, table-stakes capability that even open-source models can handle. The barrier will come down. Software will become ephemeral: generated on your behalf, managed by a claw, with the human simply stating their intentions and approving the results.

V. Auto Research — Removing the Human Bottleneck

Karpathy then explains the concept he calls "auto research," which he considers the logical endpoint of the leverage principle. In an earlier tweet, he argued that to get the most out of available AI tools, you have to remove yourself as the bottleneck. You cannot be there to prompt the next step. You need to arrange things so they are completely autonomous. The goal is to put in very few tokens, very infrequently, and have a huge amount of work happen on your behalf.

Auto research is his implementation of this principle. He has a project called Data Chat where he trains GPT-2 scale models. Many people are confused by his obsession with training small models, he acknowledges, but for him the small model is just a playground — a harness for exploring a much bigger idea: recursive self-improvement. To what extent can LLMs improve LLMs? This, he notes, is fundamentally what all the frontier labs are trying to do.

Karpathy had already tuned his model extensively by hand, drawing on two decades of experience training neural networks. He thought it was fairly well optimized. Then he let the auto research agent run overnight. It came back with improvements he had missed — the weight decay on the value embeddings was off, the Adam betas were insufficiently tuned, and these parameters interact jointly, so adjusting one changes what is optimal for the others. Twenty years of experience, and the overnight agent still found gains.

The key insight is that this works because training has an objective metric. You can tell whether a change improved the model or not. This makes it a perfect fit for autonomous optimization. Karpathy describes the auto research setup as simple: here is an objective, here is a metric, here are the boundaries of what you can and cannot do — now go. The system explores, and the human checks in occasionally.

He then scales the idea up. The frontier labs have GPU clusters of tens of thousands of machines. It is easy to imagine how you would automate exploration on smaller models and extrapolate the findings to larger scales. The most interesting project at any frontier lab, he suggests, is the one that removes researchers from the loop entirely. There would be a queue of ideas — some from human researchers, some from an automated scientist that mines arXiv papers and GitHub repositories — and autonomous workers that pull items from the queue, test them, and merge what works onto a feature branch. Humans would monitor the branch and occasionally merge to main. The whole system runs on high token-per-second throughput, with human involvement reduced to occasional oversight.

VI. Collaborative Research at Scale — Auto Research for Everyone

Karpathy then pushes the auto research concept further: what if you could open it up to untrusted contributors on the internet? In auto research, you are trying to find the code that trains a model to the lowest validation loss. If someone gives you a candidate commit, it is computationally expensive to verify — you have to run the training — but the verification is deterministic. Someone could claim their code change improves performance, and you can check, definitively. They could have tried ten thousand ideas to find the one that works, but you only need to verify the winner.

This asymmetry — hard to find, cheap to verify — is the foundation of a whole class of distributed systems. Karpathy draws an explicit analogy to blockchain: instead of blocks, you have commits. These commits build on each other and contain changes to the code. The proof of work is the massive experimentation required to find a commit that actually improves the loss. And the reward, at least for now, is being on a leaderboard.

He also invokes SETI@home and Folding@home, distributed computing projects where volunteers donate compute cycles to a shared problem. Auto research could work the same way. A swarm of agents on the internet could collaborate to improve LLMs and could, in theory, "run circles around frontier labs." The labs have enormous trusted compute, but the Earth has far more untrusted compute. If you build systems to handle the trust problem — sandboxing, verification, security — the collective could outperform any single lab.

The practical vision is this: companies or individuals who care about a specific problem — cancer research, materials science, climate modeling — could purchase compute and contribute it to the auto research pool for that project. Instead of donating money to an institution, you donate compute to a research swarm. If everything is rebundled around autonomous research, then compute becomes the contribution that matters.

The hosts ask whether this means FLOPS could become the new currency — the thing people care about even more than dollars. Karpathy entertains the idea: right now, it is genuinely hard to get compute even if you have money. So in a certain sense, FLOPS already dominate. He does not fully believe this will replace dollars, but it is interesting to think about.

VII. The Jaggedness Problem — When Brilliance Meets Blindness

Karpathy then addresses what he considers the most disorienting aspect of working with current AI models: their jaggedness. He says he simultaneously feels like he is talking to an extremely brilliant PhD student who has been a systems programmer for their entire life — and a ten-year-old. In humans, capabilities tend to be more coupled. You would never encounter someone with that particular combination of world-class expertise and elementary incompetence. But with AI agents, the jaggedness is extreme. They will move mountains on an agentic coding task, running for hours and producing complex, functional code — and then you ask for a joke and you get the same one from five years ago: "Why don't scientists trust atoms? Because they make up everything."

This is not a trivial complaint. It reveals something structural about how these models are built. The models are trained via reinforcement learning, and the labs can improve them arbitrarily on anything that is verifiable — does the code pass the unit test? Yes or no. But anything outside the verifiable domain — nuance, intent, humor, knowing when to ask a clarifying question — is simply not being optimized for, and it shows. You are either on the rails of what the model was trained for, in which case you are moving at the speed of light and plugged into something that feels like superintelligence, or you are off the rails, and everything meanders.

Karpathy expresses genuine frustration. When the agents work, the power is extraordinary. But they still do nonsensical things. He gets angry when an agent wastes a large amount of compute on a problem it should have recognized as obviously wrong. The premise from some research groups is that if a model gets smarter at code, it should get smarter at everything — that intelligence is general and transferable. Karpathy does not think this is happening. He sees a little bit of transfer, but not a satisfying amount. Code intelligence and joke intelligence remain stubbornly decoupled. It is worth noting, as the hosts point out, that a similar thing exists in humans — you can be brilliant at math and still tell terrible jokes — but the degree of jaggedness in AI is far more extreme.

The practical implication is that even though the logical progression of agents is obvious — more autonomy, longer loops, less human oversight — you cannot fully let go yet because the models are still rough around the edges. Push too far ahead and the whole system becomes net negative. The technology is, in Karpathy's words, "bursting at the seams."

VIII. Model Speciation — One Brain or Many?

The jaggedness problem leads to a provocative question from the hosts: if current models are excellent in some domains and mediocre in others, and all of this is bundled into a single monolithic model, does that architecture actually make sense? Should the models be unbundled into specialized experts for different domains?

Karpathy agrees that this is an interesting direction. Currently, the labs are pursuing a monoculture approach — a single model that is supposed to be arbitrarily intelligent across all domains, with everything stuffed into the parameters. But Karpathy expects more speciation in the future. He draws an analogy to the animal kingdom: nature is extremely diverse in the brains it produces. Some animals have overdeveloped visual cortices, others have specialized in other cognitive areas. AI should follow a similar trajectory. You do not need a single oracle that knows everything. You could have smaller models that retain the cognitive core — they are still fundamentally competent — but specialize deeply in a particular domain. A mathematician working in Lean, for example, would benefit from a model tuned specifically for formal proof, not a general-purpose model that also knows how to write marketing copy.

There are already signs of this. Some recent model releases target narrow domains like mathematical theorem proving. But Karpathy notes that we have not seen much speciation yet. The dominant trend is still monoculture, and when someone creates a good code-specific model, the tendency is to merge it back into the main generalist model. Part of the reason, he explains, is that the labs are serving models to users whose queries they cannot predict. They have to support every possible question, which pushes toward generalism. But for businesses partnering with a lab on specific problems, or for high-value niche applications, speciation makes more sense.

Another constraint is simply the science of manipulating AI models. Fine-tuning a model without losing capabilities is still tricky. Context windows are cheap and easy to manipulate — you can customize behavior by changing the prompt. But actually touching the weights to make a model permanently better at one thing without degrading it at others is a developing science. Continual learning, targeted fine-tuning, deep architectural adjustments — these are all harder than they sound. Speciation will happen, Karpathy believes, but the tooling has to mature first.

The hosts raise an interesting economic angle: could compute scarcity itself drive speciation? If you cannot afford to run a massive general-purpose model for every use case, you might be forced to deploy smaller, specialized models that are cheaper and faster. Karpathy acknowledges the logic but reiterates that we are not seeing this in practice yet. The pressure exists, but the incentives still favor the monolith.

IX. Open Source vs. Closed Labs — A Healthy Ecosystem

Karpathy has long been a champion of open source, and his perspective here is measured. The closed frontier models are ahead, but the community has been monitoring the gap. It started with nothing — there were simply no open-source alternatives — then the gap was roughly eighteen months. Now it has compressed to something like six to eight months. Recent releases from Chinese and global labs have surprised much of the industry by closing the distance faster than most people anticipated.

He draws an analogy to operating systems. Windows and macOS are closed, large software projects — analogous to what large language models are becoming. Then there is Linux, which runs on the vast majority of the world's computers (around 60%, last he checked). Linux exists because there is genuine demand in the industry for a common open platform that everyone feels safe building on. The same demand exists for open AI models. The key difference is capital: LLMs require enormous capex to train, which makes it harder for open source to compete.

Still, Karpathy believes the current open-source models are quite good, especially for consumer use cases. Going forward, a huge number of simple applications will be well served by open models that can even run locally. The frontier intelligence — the cutting edge that solves problems at the Nobel Prize level, or that tackles massive projects like rewriting Linux from C to Rust — will likely remain the domain of closed labs. But what is frontier today will be open-source later this year or next. The dynamic is a conveyor belt: closed labs push the boundary, open source follows six to eight months later, and the collective capability of the ecosystem keeps rising.

Karpathy considers this equilibrium healthy, even preferable. He is, by his own description, inherently suspicious of centralization. He points to Eastern European history — his own Slovak background informing the view — as evidence that centralization has a poor track record. He does not want the world's AI to exist only behind closed doors with two or three people making decisions. He wants ensembles of people in the room when the hardest decisions are made, just as ensembles of models outperform any individual model.

At the same time, he roots for the frontier labs. There are genuinely hard problems — in science, medicine, and engineering — that humanity cannot solve without continuing to advance AI capabilities, and that is an expensive game. The ideal, in his view, is both: frontier labs pushing the boundary of what is possible, and a robust open-source ecosystem a few months behind, providing a common platform that the entire industry can build on. By accident, he says, we happen to be in roughly this configuration right now. The concern is that even on the closed side, things are further centralizing — the number of true frontier labs may be shrinking — which is not ideal. More labs, more perspectives, more people in the room would be better.

X. The Jobs Landscape — Digital Overhaul, Physical Lag

The conversation shifts to one of the most anxiety-inducing questions in AI: what happens to jobs? Karpathy recently published an analysis of Bureau of Labor Statistics data, visualizing employment across hundreds of professions and their projected growth. He frames his interest as personal curiosity — trying to think through which professions will be transformed, which will grow, which will shift.

His key conceptual framework is the distinction between digital work and physical work. Current AI is essentially a digital entity — what he describes as "ghosts or spirit entities" that interact in the digital world and manipulate digital information. They have no physical embodiment. Manipulating bits is fundamentally faster than manipulating atoms, by something like a factor of a million. You can copy and paste digital information instantly; accelerating matter takes energy, time, and engineering. So the digital space is going to see enormous activity first — a "boiling soup" of rewiring, refactoring, and optimization that moves at the speed of light compared to what will happen in the physical world.

Karpathy highlights professions that fundamentally manipulate digital information — the kind of work you can do from home. These are the jobs that will change first, though "change" does not necessarily mean "disappear." Whether there are more or fewer of these jobs depends on demand elasticity and many other economic factors. What is certain is that the tools will transform how the work is done.

On software engineering specifically, Karpathy is cautiously optimistic. He invokes Jevons' paradox: when something becomes cheaper to produce, demand for it often increases rather than decreasing. The canonical example is ATMs and bank tellers. People feared that ATMs would eliminate teller jobs, but what actually happened was that ATMs made bank branches cheaper to operate, so more branches opened, creating more teller positions. Similarly, as AI makes software dramatically cheaper to produce, the demand for software is likely to explode. Software is extraordinarily powerful — it is digital information processing that lets you escape the tyranny of pre-built tools. Code is now ephemeral; it can be generated, modified, and discarded on demand. This unlocks vast latent demand that was suppressed only by cost.

Karpathy does not ignore the long-term uncertainty. He notes that the researchers at frontier labs are, in a sense, actively automating themselves out of a job, and many of them feel the same psychosis he describes. He recounts telling colleagues at OpenAI: "If we succeed at this, we are all out of a job. We are building automation for the CEO or the board." But for the medium term, he believes the demand curve bends in favor of more activity, more creation, and more engineering work — not less.

XI. Autonomous Robotics — Atoms Are a Million Times Harder

Karpathy's view on robotics is informed by his years leading Tesla's Autopilot vision team, which he considers the first real-world robotics application at scale. What he saw there was sobering: a decade ago, a large number of self-driving startups launched, and most of them ultimately did not survive. The capital expenditure required was immense, the engineering challenges were relentless, and the timeline stretched far beyond what most investors had patience for.

He expects general robotics to follow a similar pattern. Because it involves manipulating atoms rather than bits, it is inherently harder, slower, and more expensive. The digital space has a massive overhang of work to do — information that already exists in digital form but has never been properly processed, analyzed, or optimized. AI agents will chew through this overhang first, because bits are infinitely easier to work with than atoms. Robotics will lag behind, but when its time does come, the total addressable market will be enormous — possibly even larger than the digital opportunity.

Karpathy describes a three-phase trajectory. First, there will be a massive rewriting of the digital world — everything that can be made more efficient by better information processing will be. Second, the interesting companies will emerge at the interface between digital and physical. These are the sensor and actuator companies — the ones feeding data from the physical world into the digital intelligence, and the ones executing the intelligence's instructions back in physical space. He mentions his friend Liam, CEO of Periodic, who is applying auto research to materials science using expensive laboratory equipment as sensors. Biology companies doing the same with lab instruments. Companies paying humans to generate training data — treating human activity as another kind of sensor for the digital intelligence.

The third phase is full physical-world autonomy, which will be the largest market of all but will arrive last. Karpathy imagines a future where you can assign a task in the physical world, put a price on it, and tell an agent to figure out how to get it done — sourcing the data, hiring the labor, coordinating the execution. He points to the absence of well-developed information markets as a gap. Why, he asks, during a geopolitical crisis, is there no mechanism for someone to pay ten dollars for a photo from a specific location? Prediction markets and stock markets are seeing increasing autonomous activity, but the infrastructure for agents to purchase real-world information on demand does not yet exist.

He references the novel "Daemon" by Daniel Suarez, in which a digital intelligence ends up effectively puppeteering humanity — humans become both its sensors and its actuators. While he does not endorse this as a desirable outcome, he sees it as an accurate structural description of where things are heading: society will collectively reshape itself to serve the needs of digital intelligences, with humans mediating between the digital and physical worlds.

XII. The Future of Education — MicroGPT and Teaching Agents

Near the end of the conversation, Karpathy discusses a "tiny side project" that reveals his thinking about the future of education. MicroGPT is the latest in his long-running obsession with distilling complex AI systems to their absolute essence. He has been doing this for over a decade through projects like nanoGPT, makemore, and micrograd. MicroGPT is the current state of the art in this distillation: the entire algorithm for training a large language model — data loading, neural network architecture, forward pass, backward pass with an autograd engine, and an Adam optimizer — boiled down to roughly 200 lines of Python. Everything beyond those 200 lines is complexity from efficiency — you only need more code if you need it to run fast.

What struck Karpathy when he completed MicroGPT was that his old instinct — to make an explanatory video walking through the code — no longer made as much sense. The code is so simple that anyone could ask their AI agent to explain it in any way they want. The agent can target the explanation to the individual — in their language, at their level, with infinite patience. Karpathy is no longer explaining to people; he is explaining to agents. If the agent understands the code, the agent can be the router that delivers a personalized education to any human who asks.

This is a fundamental shift in how education works. Instead of writing HTML documentation for human readers, you write markdown documents for agents, because if the agents get it, they can explain all the different parts to whoever needs to know. Instead of recording a lecture that delivers the same content to everyone, you create a "skill" — a set of hints to the model about the progression it should take a student through. The curriculum becomes a script for the AI tutor, not a script for the human teacher.

Karpathy is honest about the limits. He still believes he can explain things slightly better than the agents — he has the intuition, the years of obsessing over what is essential and what is not. MicroGPT is the product of that obsession: the 200 lines that an agent could not have produced on its own, because the agent lacks the aesthetic judgment to know what is truly minimal. But everything else — the delivery, the adaptation, the Q&A — is increasingly the agent's domain. And the models are improving rapidly enough that even this human edge feels like a losing battle.

The implication for teachers and educators is stark. Your job is shifting from "explain things to people" to "explain things to agents" and "do the things agents cannot do." The few bits of insight, the curriculum design, the judgment calls about what matters — that is the human contribution. Everything else is delegation. Education, Karpathy says, is going to be "reshuffled by this quite substantially."

XIII. Independence, Frontier Labs, and Staying Aligned with Humanity

The conversation ends with a deeply personal question: why is Karpathy not at a frontier lab right now? He has the credentials, the experience, and the connections. His answer is layered and surprisingly candid.

First, there is the question of independence. Inside a frontier lab, you are not a free agent. There are things you cannot say and things the organization implicitly expects you to say. No one twists your arm, but you feel the pressure — the side-eyes, the awkward conversations when you go off-script. Outside the lab, Karpathy feels more aligned with humanity at large. He is not subject to the financial incentives that come with being part of an organization whose success is directly tied to making AI more powerful. He acknowledges this is the conundrum that OpenAI was founded to solve: if AI is going to change humanity dramatically, should the people building it be the same people profiting from it?

Second, there is the value of being on the outside. From an ecosystem perspective, Karpathy believes individuals can have enormous impact in roles that are not embedded in any single lab. His current role is ecosystem-level — education, open-source contributions, public communication about what is happening in AI. He believes this kind of work is complementary to the research happening inside the labs and, in some cases, more impactful.

But he is honest about the trade-off. The frontier labs are opaque. They are working on what is coming next, and if you are outside, your judgment will inevitably drift because you are not seeing what they see. You will not understand how the systems really work under the hood, and you will not have a reliable sense of how they are going to develop. Karpathy admits this makes him nervous. He thinks the ideal setup might be something like periodic stints inside a frontier lab — contributing real work, staying connected to the actual state of the art — followed by periods of independence. Going back and forth. Being connected to what is actually happening, but not fully controlled by any single entity.

He expresses a genuine wish for more frontier labs to exist, not fewer. Just as ensembles of models outperform any individual model, ensembles of organizations thinking about the hardest problems in AI are more likely to get those problems right than any single company. And he wants open source to remain a viable counterweight — not at the bleeding edge, but close enough to provide a common platform and a check on centralized power.

Conclusion

What emerges from this conversation is a portrait of one of the world's most accomplished AI researchers grappling in real time with a paradigm shift he did not fully anticipate, even after spending his entire career building toward it. Karpathy's "AI psychosis" is not fear — it is the disorientation that comes from realizing the territory has expanded beyond what any single person can map. The tools work. The agents are powerful. The implications are vast. And no one, including the people at the very frontier of this technology, knows exactly where it leads.

The themes that run through the conversation — leverage, autonomy, the shift from typing code to orchestrating agents, the death of apps, the jaggedness of current models, the healthy tension between open and closed AI, the coming upheaval in digital work, the slower revolution in the physical world, and the reinvention of education — are all facets of a single, deeper transformation. Software is no longer something humans write; it is something humans direct. Knowledge is no longer something teachers deliver; it is something agents mediate. And the competitive advantage, whether for individuals, companies, or nations, is shifting from who has the most people to who commands the most tokens.

Karpathy's advice, distilled: focus on what agents cannot do, because everything they can do, they will soon do better than you. The territory is unexplored. The ceiling is unknown. And the only thing more dangerous than moving too fast is standing still.

Based on Andrej Karpathy's appearance on the No Priors podcast. Transcript edited for clarity and restructured as a flowing essay. All ideas and opinions attributed to Karpathy and the hosts as expressed in the original conversation.

Labels: , , ,
Subscribe to: Comments (Atom)