Dwarkesh Podcast

Andrej Karpathy — AGI is still a decade away

Dwarkesh Podcast·May 16, 2026

OVERVIEW

This episode features an in-depth conversation with Andrej Karpathy, a prominent figure in AI, about the future of artificial intelligence, particularly focusing on the timeline for Artificial General Intelligence (AGI). Karpathy posits that AGI is still approximately a decade away, a prediction he defends by dissecting the current limitations of AI models and the complex, multifaceted challenges that still need to be overcome. The discussion spans the evolution of AI, the importance of foundational capabilities like continuous learning and multimodality, and a critical analysis of current AI development paradigms, including the role of pre-training, in-context learning, and the limitations of current models in simulating human-like intelligence.

KEY TOPICS

  • The "Decade of Agents" versus the "Year of Agents"
  • Current limitations of AI models (lack of common sense, multimodality, continuous learning)
  • Over-predictions in the AI industry and the realistic timeline for AGI
  • Evolution of AI: from neural networks to agents
  • Missteps in AI development, particularly in reinforcement learning and game environments
  • The "Ghost" analogy for current AI models versus "Animal" intelligence
  • In-context learning and the distinction between learned knowledge and algorithmic patterns
  • Analogies between human and AI learning processes (pre-training vs. evolution, working memory vs. long-term memory)
  • Challenges in replicating human intelligence: emotions, instincts, and the cognitive core
  • Future of AI architectures: transformers, attention mechanisms, and sparse models
  • The role of AI in coding and engineering workflows
  • The importance of high-quality, curated datasets for AI training
  • Educational paradigms for AI and the concept of "Starfleet Academy"
  • The "curse of knowledge" and effective teaching methods for complex topics
  • Superintelligence and the future of human-AI collaboration
  • Self-driving cars as a case study for AI development challenges
  • The economic and societal impact of AGI, including job displacement and re-education

MAIN TAKEAWAYS

  • Andrej Karpathy firmly believes that AGI is still about a decade away, contrasting with more optimistic "year of agents" predictions due to significant unresolved challenges in AI capabilities.
  • Current AI models, while impressive, lack crucial elements of human intelligence such as robust common sense, true multimodality, and continuous learning, which are vital for AGI.
  • The distinction between "ghosts" (current AI models) and "animals" (human-like intelligence) highlights that AI is not evolving through biological processes but through data imitation, leading to a different form of intelligence.
  • In-context learning is a powerful current capability, but it relies on external data rather than deep, internalized representations, making it akin to "working memory" in humans.
  • Human-like intelligence involves complex cognitive functions beyond what current AI can emulate, including emotional understanding, instinct, and a true cognitive core stripped of superfluous knowledge.
  • The development of AI requires a multidisciplinary approach, with advancements needed across data, hardware, software, and algorithms, all progressing synergistically.
  • AI's application in coding workflows is currently most effective for boilerplate tasks and auto-completion rather than genuinely creative or architecturally complex coding, suggesting limitations in independent innovation.
  • Education needs to evolve significantly to prepare humans for an AI-driven future, focusing on accessible, adaptable learning pathways and potentially leveraging AI as a personalized tutor, rather than just an information source.
  • The long-term vision for AI is not necessarily about replacing humans but about creating more autonomous and capable entities that collaborate with humans, necessitating new interfaces and management strategies.

NOTABLE QUOTES

"This will be the decade of agents, not the year of agents."
"I feel like the problems are tractable. They're surmountable. But they're still difficult."
"I feel like there's some overpredictions going on in the industry."
"I feel like there's some miraculous compression going on... obviously the weights of the neural net are not stored in ATCGs."
"Pre-AGI, education is useful. Post-AGI, education is fun."
"Learning feels good. And I think it's technical problem to get there."
"You're presenting the pain before you present the solution."

Summarized with DriftNote — AI-powered podcast summaries

Try it free