In this extensive interview, AI researcher Andrej Karpathy offers a pragmatic and nuanced perspective on the state of artificial intelligence. He argues that true agentic AI (AGI) is still a decade away, cautioning against industry hype by highlighting the significant "demo-to-product" gap, exemplified by the slow progress of self-driving cars. Karpathy critiques current reinforcement learning (RL) as a deeply inefficient paradigm, famously describing it as "sucking supervision through a straw." A central theme is his analogy of LLMs as "ghosts"—entities trained by imitating vast amounts of internet data—rather than "animals" shaped by evolution, which possess innate, compressed algorithms for learning. He discusses the cognitive deficits of current models, such as their over-reliance on memorization and their tendency to "collapse" into non-diverse outputs. Finally, Karpathy shares his vision for his new venture, Eureka, an educational project aimed at empowering humanity by creating effective "ramps to knowledge" in a future increasingly shaped by AI.
Andrej Karpathy’s Pragmatic Take on AI: From "Ghosts" to GDP and Education
0:48 The Decade of Agents, Not the Year: Karpathy pushes back against the hype of immediate, powerful AI agents. He argues it will be a "decade of agents" because current models have significant cognitive deficits, lacking continual learning and robustness, making them unsuitable for complex roles like a human intern.
8:41 Building "Ghosts," Not Animals: Karpathy distinguishes the AI development process from biological evolution. Animals evolve with a huge amount of "baked-in" hardware and learning algorithms compressed into their DNA. In contrast, AI models are trained by imitating human-generated internet data, resulting in "ghosts" or "spirits" that are fundamentally different from evolved animals. He calls pre-training a "crappy evolution."
13:35 The Need for a "Cognitive Core": He argues that LLMs are held back by their over-reliance on memorized knowledge from the internet. The goal should be to distill a "cognitive core"—the pure algorithms for intelligence, problem-solving, and in-context learning—while stripping away the vast, hazy memory.
40:53 Reinforcement Learning is "Terrible": Karpathy offers a sharp critique of RL, stating it's an inefficient learning method where a single final reward is used to update an entire sequence of actions—a process he calls "sucking supervision through a straw." This noisy method wrongly reinforces every step in a successful trajectory, unlike human reflection. He notes that better process-based rewards are hard to implement because LLM judges are easily tricked by adversarial examples (e.g., "dhdhdhdh").
51:59 The Problem of Model "Collapse": When models are trained on their own synthetic outputs (e.g., reflections on a text), they "collapse" into a very narrow distribution of ideas. They lose diversity and entropy, which is why simply training on synthetic data fails. He humorously notes that humans also collapse over their lifetime, becoming less creative and more set in their ways.
1:07:13 AGI Will Blend into 2% GDP Growth: Karpathy predicts that AGI will not cause a sudden, sharp economic explosion. Instead, like computers and the internet, its impact will be a gradual diffusion of automation that blends into the existing ~2% annual GDP growth trend we've seen for centuries. He views AI as a continuation of a long history of recursive self-improvement in technology.
1:18:24 Superintelligence and Loss of Control: He views superintelligence as an extrapolation of automation. The most likely long-term outcome is not a single rogue AI, but a gradual societal loss of understanding and control as we delegate more and more tasks to a complex ecosystem of competing, autonomous AI entities.
1:43:43 Lessons from Self-Driving Cars: Having led Tesla's AI for self-driving, Karpathy uses it as a core example of the "demo-to-product gap." Achieving the first 90% (a demo) is easy compared to the "march of nines" (99%, 99.9%, etc.) required for a reliable, safe product. Each "nine" of reliability requires a constant, massive amount of work.
1:57:08 The Future of Education (Project Eureka): Karpathy reveals his new focus is on education to empower humanity. His project, Eureka, aims to build the "Starfleet Academy"—an elite institution for technical knowledge. The ultimate goal is to create an AI tutor that can perfectly understand a student's knowledge and provide appropriately challenging material, making learning feel effortless and desirable, much like people go to the gym for fun and self-improvement today.