Recently, Andrej Karpathy, a well-known figure in the AI field, shared his profound insights on the current development of artificial intelligence (AI) during a conversation with podcast host Dwarkesh Patel. His views have sparked widespread attention, especially regarding the maturity of agents and AGI (Artificial General Intelligence).
Karpathy stated that current agents such as Claude and Codex, although impressive, still need ten years to reach a level where they can truly be "on the job." He pointed out that current agents are more like inexperienced interns, lacking multimodal capabilities, continuous learning abilities, and a complete cognitive structure. He emphasized that the current technological bottleneck is not about computing power, but rather the incomplete cognitive components of agents, which prevent them from achieving real continuous learning and reasoning.
When discussing the learning mechanisms of AI, Karpathy criticized reinforcement learning, arguing that this "trial and error" learning approach does not reflect the real learning process of humans. He believes that human learning is complex and nonlinear, while reinforcement learning tends to view all attempts as successful paths, ignoring the errors and accumulated experiences in between.
Karpathy also pointed out that future AI research should focus on making agents "learn more like humans." He mentioned that future agents need to have self-growth capabilities and more complex cognitive structures, only then can they truly make a leap from simple tools to intelligent companions. He suggested incorporating structural long-term memory systems into agent design to simulate the way humans learn and remember.
Karpathy's views have filled us with anticipation for future AI agents, while also reminding us not to rush too much; we still need to continue exploring and innovating.
