In a new conversation with Dwarkesh Patel, Andrej Karpathy outlines why he believes AGI is roughly a decade away. He frames the 2020s as the “decade of agents,” arguing that today’s systems remain cognitively limited — lacking durable memory, richer multimodality, and reliable computer-use — but are on a steady path of improvement. Karpathy emphasizes that useful agents should resemble capable digital interns that can navigate real products and workflows, not just pass benchmarks. Full transcript: Andrej Karpathy — AGI is still a decade away.

Tracing the field’s shifts, he contrasts early per-task neural networks, a premature wave of game-centric reinforcement learning, and the present focus on LLMs as representation engines that enable higher-level behaviors. In his view, attempts to “get the full agent” too early (e.g., game environments) missed the prerequisite of strong, general representations learned via large-scale pretraining. With those foundations now in place, he expects agents to gain competence through improved memory, computer interaction, and continual learning.

Karpathy is skeptical that classic RL alone will unlock intelligence, suggesting humans rely far less on RL for abstract reasoning than for motor skills. Instead, he sees progress coming from better training data, tooling, and evaluation — plus pragmatic product loops where agents do real work under supervision. He also expects computer-use (UI control) to be a major catalyst, letting models click, type, and navigate apps to complete end-to-end tasks.

Why it matters: the interview provides a grounded roadmap for practitioners — build agents that operate real interfaces, invest in memory and multimodality, and measure success on useful work rather than leaderboard snippets. For founders and teams, the takeaway is to design systems that compound: tighter feedback from deployment → better data → more reliable agents over time. Read the full discussion here: Dwarkesh x Karpathy.