Andrej Karpathy on the Decade of AI Agents: 'Summoning Ghosts, Not Building Animals'

The Future of AI Agents

• 💡 Andrej Karpathy argues that fully functional AI agents will emerge over the next decade, not this year, requiring sustained progress.
• 🎯 Current agent technologies, while impressive, lack crucial features like continuous learning, multimodality, deep cognition, and practical computer interface use.
• 🔑 A major bottleneck is the absence of continuous learning, meaning agents restart at each session, preventing genuine adaptability and growth over time.

Distinguishing AI from Biological Intelligence

• 👻 Karpathy posits that AI is "summoning ghosts, not building animals", meaning AI models reflect human knowledge distilled from internet data, not innate, evolved animal intelligence.
• 🧬 Unlike animals with millions of years of innate knowledge and instincts, AI agents are built by distilling publicly available data, making them ethereal rather than biologically grounded.
• 🧠 Large-scale pre-training is seen as a crude proxy for evolution, assembling initial cognition but limiting the ability for continuous, lifelong learning akin to biological systems.

Current Limitations and Architectural Evolution

• 📚 In-context learning is described as pattern completion within the prompt window, acting like working memory rather than true long-term learning or consolidation.
• ⚠️ Reinforcement Learning (RL) is criticized for its sparse and noisy reward signals, which reinforce entire sequences indiscriminately, unlike human reflection on effective steps.
• 🛠️ While architectural improvements like sparse attention and multimodal training will occur, the fundamental structure of large neural networks trained with gradient descent is expected to remain for the next decade.

Practicalities of AI Adoption

• 📈 AI's most tangible automation currently impacts knowledge work with structured inputs and outputs, like coding, but struggles with complex, messy workflows in other domains.
• 🚗 Drawing parallels with self-driving cars, Karpathy highlights that high capital investment, latency constraints, and legal liabilities significantly slow widespread adoption despite technological feasibility.
• 🐢 Karpathy expresses skepticism about a sudden "intelligence explosion", viewing AI progress as an extension of centuries of continuous computational and productivity improvements, not an abrupt discontinuity.

Empowering Humans in the AI Era

• 🧑‍🏫 Karpathy emphasizes the critical role of education in empowering humans to understand, control, and influence evolving AI systems.
• ✅ He advocates for adaptive, personalized learning—inspired by successful one-on-one tutoring—to unlock human potential alongside AI technologies.
• 🚀 The vision is a phased transition, starting with human experts assisted by AI tutors, gradually moving towards AI systems taking on more substantive instructional roles as their capabilities mature.