Andrej Karpathy — AGI is still a decade away - Dwarkesh Podcast Recap

Podcast: Dwarkesh Podcast

Published: 2025-10-17

Duration: 2 hr 25 min

Summary

Andrej Karpathy argues that we are still a decade away from achieving true artificial general intelligence (AGI), emphasizing that the development of advanced agents requires significant work and innovation. He reflects on the historical context of AI advancements and the challenges that remain.

What Happened

In this episode, Andre Karpathy discusses why he believes this decade is pivotal for the development of AI agents, contrasting it with the common expectation that this would be the year of agents. He notes that while there are already impressive early agents like Cloud and Codex, they still lack essential capabilities such as continual learning and cognitive depth. Karpathy emphasizes that these issues are not trivial and will require a concerted effort over the next decade to resolve.

Karpathy reflects on his extensive experience in the AI field, mentioning that he has observed various trends and predictions over the past 15 years. He points out that the problems surrounding the development of sophisticated agents are indeed tractable but still complex. His intuition, shaped by years of engagement in both research and industry, leads him to estimate that it will take about a decade to overcome the existing limitations. He also notes the potential for seismic shifts in AI development and recounts some of the pivotal moments in the field, such as the introduction of deep learning through AlexNet and the early attempts at reinforcement learning.

The conversation dives into the historical missteps in AI research, particularly the focus on agents without the necessary foundational work in representations and language models. Karpathy illustrates this by discussing his own experiences at OpenAI, where he aimed to develop agents that could interact with web pages but faced challenges due to the complexity of the tasks and the inadequacy of existing technologies. He concludes by highlighting that the current advancements in language models provide a more promising foundation for developing capable agents, but stresses that there are still crucial components missing in the overall stack needed for AGI.

Key Insights

• Karpathy believes we are a decade away from achieving true AGI due to the complexity of developing capable agents.
• Current AI agents lack essential features like continual learning and cognitive depth, which are critical for their functionality.
• Historical trends in AI show that researchers often underestimated the foundational work needed before tackling complex agent development.
• Advancements in large language models are crucial for creating competent agents, but more foundational work is still necessary.