Part 2/15:

Karpathy clarifies that the industry’s hype often predicts an imminent "year of agents," but he advocates for a more measured view: this is the decade of agents. While early agents like Claude or Codex demonstrate impressive capabilities, they are still nascent. The work ahead involves refining these systems over ten years, with steady improvements in multimodality, continual learning, and understanding.

He argues that current agents are akin to interns or employees: they lack the comprehensive intelligence and capacity for autonomous action yet. For example, today’s models can't yet handle complex computer use, remember past interactions, or learn continually—issues that should take roughly a decade to resolve through research and development.