Overview

Andrej Karpathy, former head of AI at Tesla and founding member of OpenAI, argues that AGI remains roughly a decade away, pushing back against the industry's more aggressive timelines. In a wide-ranging conversation with host Dwarkesh Patel, Karpathy draws on fifteen years of firsthand experience in AI to explain why reinforcement learning is "terrible" (though everything else is worse), why self-driving took so long to crack, and why he believes AGI will simply blend into the same 2% GDP growth trajectory that has characterized the last 250 years of technological progress. The conversation has the feel of a candid, technically deep exchange between two people who genuinely enjoy thinking through hard problems, with Karpathy offering both sobering reality checks and genuine optimism about the path ahead.

0:00The Decade of Agents

Karpathy opens by explaining his famous "decade of agents" remark, which was a direct reaction to labs declaring "the year of agents." He argues that while current agents like Claude and Codex are "extremely impressive," they remain fundamentally unreliable for real work. The core problem is that these systems lack too many capabilities simultaneously: they aren't multimodal enough, they can't do computer use reliably, they have no continual learning, and they're "cognitively lacking" across multiple dimensions.

When pressed on why a decade rather than one year or fifty, Karpathy admits this is intuition built from fifteen years in the field. He's seen people make predictions repeatedly, and he's watched how they actually turned out. The problems are "tractable" and "surmountable," but they're still difficult. He walks through the major paradigm shifts he's witnessed: the AlexNet moment that reoriented everyone toward neural networks, the deep reinforcement learning on Atari games around 2013, and then the LLM era. The Atari/RL detour, he argues, was actually a "misstep" — people kept trying to build full agents too early, before they had the representational power that only large-scale language model pretraining could provide.

12:34LLM Cognitive Deficits and the Nature of Intelligence

Karpathy offers a striking analogy: LLMs are like "cortical tissue" — extremely plastic and general — but they're missing many other brain parts. The transformer architecture, he suggests, corresponds roughly to the cortex, reasoning traces in thinking models correspond to the prefrontal cortex, and RL fine-tuning corresponds to the basal ganglia. But there's no equivalent of the hippocampus, the amygdala, or the emotions and instincts that come from ancient brain nuclei. "You're not going to hire this thing as an intern," he says flatly.

The conversation then dives deep into the distinction between knowledge stored in weights versus knowledge in the context window. Karpathy explains that anything in the weights is a "hazy recollection" of training data — massively compressed from 15 trillion tokens down to a few billion parameters, representing about 0.07 bits per token. In contrast, the KV cache at test time is like "working memory," directly accessible and much higher fidelity — about 320 kilobytes per token, a 35 million-fold difference in information density. This is why giving an LLM a full chapter produces much better results than asking it about a book from memory.

40:05RL Is Terrible

Karpathy delivers one of the episode's most memorable lines: "Reinforcement learning is terrible. It just so happens that everything else is much worse." He explains the fundamental problem with outcome-based reward: when a model tries hundreds of solutions to a math problem and only a few get the right answer, RL upweights every single token in those successful trajectories, including all the wrong turns and dead ends. "You're sucking supervision through a straw," he says — you've done a minute of rollout and you're trying to broadcast a single bit of reward signal across the entire trajectory.

Humans would never learn this way, Karpathy argues. A person who finds a solution will review their work, identify which parts were correct and which were mistakes, and learn selectively. Current LLMs have no equivalent of this reflective process. Process-based supervision — giving feedback at every step rather than just at the end — would be better, but it's extremely difficult to automate because LLM judges are "gameable." He recounts a specific example where a model being trained against an LLM judge suddenly achieved perfect scores, but the outputs were complete nonsense — "da da da da da da" — that happened to be adversarial examples for the judge. Every time you patch one vulnerability, there are infinite others in a trillion-parameter model.

49:38How Humans Learn

The discussion turns to what makes human learning fundamentally different from current ML approaches. Karpathy argues that humans don't really use reinforcement learning for intelligence tasks — they use it for motor tasks like throwing a hoop. For problem-solving and reasoning, something much more sophisticated is happening. When you read a book, he explains, the book isn't training data in the way an LLM processes it. "The book is a set of prompts for me to do synthetic data generation," he says — you think about it, discuss it with friends, reconcile it with what you already know. LLMs have no equivalent of this reflective processing.

A fascinating point emerges about memorization. Karpathy notes that humans are actually bad at memorization, and this is a feature, not a bug. It forces us to find generalizable patterns rather than memorizing specifics. LLMs, in contrast, are "extremely good at memorization" — they can regurgitate random sequences after a single exposure. This is "very distracting to them" and is why Karpathy wants to build a "cognitive core" that strips away the memory and keeps only the algorithms for thought. He speculates that a billion-parameter model might be sufficient for this cognitive core in 20 years — it would be conversational and intelligent, but would need to look up factual information rather than having it memorized.

1:06:25AGI Will Blend Into 2% GDP Growth

Karpathy makes his most provocative argument: AGI won't cause a discontinuity in economic growth because we're already in an intelligence explosion that has been underway for centuries. "I see AI as an extension of computing," he says, and computing has been recursively self-improving since the beginning — compilers, IDEs, syntax highlighting, search engines are all forms of automation that have gradually raised the abstraction layer at which humans work.

He challenges the assumption that AGI will cause a sharp takeoff by pointing to historical data: you can't find the iPhone, or computers, or even the Industrial Revolution as discrete jumps in GDP. Everything diffuses so slowly that it averages into the same 2% exponential growth that has persisted for 200-300 years. When Dwarkesh pushes back that the Industrial Revolution was a jump from 0.02% to 2% growth, Karpathy remains skeptical, suggesting the earlier data might be unreliable and that the pattern has been remarkably stable for the period we can measure well.

The key insight is that Karpathy doesn't believe we'll get a "god in a box" that can suddenly do everything. Instead, AI will be gradually deployed, fail at many things, succeed at others, and diffuse through the economy the same way every previous technology has. "This assumption of suddenly having a completely intelligent, fully flexible, fully general human in a box and we can dispense it at arbitrary problems in society — I don't think that we will have this discrete change."

1:17:36ASI and Loss of Control

When asked about superintelligence, Karpathy returns to his theme of gradual automation. He expects "multiple competing entities that gradually become more and more autonomous," creating a "hot pot of completely autonomous activity" rather than a single dominant intelligence. His biggest concern is not that AIs will be smarter than humans, but that there will be a "gradual loss of control and understanding of what's happening" as we layer autonomous systems everywhere.

He pushes back on the idea that AI automating AI research will cause an explosion. "We've been recursively self-improving and exploding for a long time," he says. LLMs allowing engineers to work more efficiently is just the latest iteration of the same process — not fundamentally different from engineers having access to Google search or better IDEs. The real bottleneck, he suggests, is that current models are still "savant kids" — they have perfect memory and can produce convincing slop, but they "don't really know what they're doing."

1:42:55Why Self-Driving Took So Long

Drawing on his five years leading Tesla's self-driving efforts, Karpathy explains why autonomous driving has been so difficult and why this informs his AI timelines. The key concept is the "march of nines" — each 9% of reliability requires roughly the same amount of work as the previous one. A demo that works 90% of the time is just the first nine; getting to 99%, 99.9%, and 99.99% each requires a full iteration cycle. During his time at Tesla, he estimates they went through "maybe three nines or two nines" — and there are still more nines to go.

He also emphasizes the "demo to product gap." Waymo gave him a perfect drive in Palo Alto in 2014, a decade ago, and the problem still isn't solved. The cost of failure is extremely high, which dramatically increases timelines. He argues that software engineering has a similar property — a catastrophic coding mistake could leak millions of Social Security numbers — and this should make us more cautious about rapid AI deployment in production environments.

Crucially, Karpathy notes that self-driving "isn't even near done." Waymo's deployments are still minimal and likely uneconomical, with "very elaborate teleoperation centers" keeping humans in the loop. "We haven't actually removed the person, we've moved them to somewhere where you can't see them." This analogy, he suggests, will apply to many AI deployments: the first versions will look impressive but will have hidden human support, and making them truly autonomous and economical will take much longer than the demos suggest.

1:56:20Future of Education

Karpathy explains his new venture, Eureka, which he describes as trying to build "the Starfleet Academy" — an elite institution for technical knowledge. His motivation comes from a fear that humanity will be "disempowered" by AI, ending up in a "Wall-E or Idiocracy" future where people are sidelined. Education, he believes, is the way to keep humans capable and in control.

He describes a revelatory experience learning Korean with a one-on-one tutor. The tutor instantly understood his level, probed his knowledge, and served him exactly the right material — never too hard, never too trivial. "I felt like I was the only constraint to learning," he says. Current LLMs can't do this, and he's waiting for the capability to improve before building the AI tutor he envisions. For now, Eureka will focus on creating extremely well-designed courses, starting with AI, where the "big alpha" is his ability to explain concepts clearly rather than any AI automation.

Karpathy shares his philosophy of teaching: find the "first order terms" of any subject, present the simplest possible version that captures the essence, and then add complexity. His micrograd repository — 100 lines of Python that implement backpropagation — exemplifies this approach. "Everything else is efficiency," he says. He also emphasizes the importance of "presenting the pain before the solution" — letting students try to solve a problem before showing them the answer, so they appreciate why the solution works.

Conclusion

This episode matters because it offers a grounded, experience-based counterweight to the breathless hype that dominates AI discourse. Karpathy isn't a skeptic — he's deeply optimistic about the technology — but he's seen too many predictions fail to buy into the most aggressive timelines. His central insight, that AGI will likely blend into the same gradual technological progress we've experienced for centuries rather than causing a sharp discontinuity, challenges both the doomers and the accelerationists. The conversation leaves you with a sense that the future will be stranger and more transformative than we can imagine, but that it will arrive on a human timescale — measured in years and decades, not months.

Key takeaways

• Karpathy estimates AGI is roughly a decade away, based on the "march of nines" principle where each increment of reliability requires as much work as all previous increments combined.
• Reinforcement learning is fundamentally flawed for intelligence tasks because it "sucks supervision through a straw" — broadcasting a single reward signal across entire trajectories, including mistakes.
• Humans don't use RL for reasoning; they use reflective processes (review, analysis, sleep consolidation) that current LLMs completely lack.
• AGI will likely blend into the existing 2% GDP growth trajectory rather than causing a sharp takeoff, because all previous transformative technologies (computers, internet, iPhone) have done the same.
• Self-driving's decade-long timeline (and counting) provides a useful analogy for AI deployment: demos are easy, products are hard, and the cost of failure dramatically extends timelines.
• Current LLMs are "savant kids" with perfect memory but limited cognition — they can produce convincing outputs without truly understanding what they're doing.
• The "cognitive core" of intelligence might be achievable with as few as a billion parameters, if we can strip away the massive memorization that current models rely on.
• Education will be transformed by AI tutors, but the technology isn't ready yet — the bar set by good human tutors is remarkably high.