OpenAI cofounder Andrej Karpathy says it will take a decade before AI agents actually work

In a recent episode of the Dwarkesh Podcast, Andrej Karpathy, co-founder of OpenAI, expressed a measured outlook on the imminent development of functional AI agents. Despite the rapid advancements in artificial intelligence, Karpathy believes that significant challenges remain, stating, "They just don't work. They lack sufficient intelligence and multimodal capabilities, and they struggle with tasks like computer usage." He highlighted the absence of continual learning in current AI, meaning that these systems cannot retain information or adapt from previous interactions. Karpathy anticipates that it might take around ten years to address these shortcomings effectively. AI agents, which many predict will become mainstream around 2025, are designed to act autonomously, tackling problems and executing plans without constant user input. However, Karpathy's critique of the industry suggests that there is an overestimation of current capabilities, as he noted, "The industry imagines a future where fully autonomous entities can code without human assistance, but I don't want to live in that world." Instead, he envisions a collaborative relationship between humans and AI, where AI tools assist in coding while allowing for human input and learning. He expressed a desire for AI to provide documentation and verify its actions, emphasizing the importance of collaboration in enhancing programming skills rather than simply generating code. Karpathy's concerns about the efficacy of AI agents resonate with others in the industry. For instance, Quintin Au from ScaleAI pointed out that the errors made by AI agents compound with each action they take, creating reliability issues. He noted that even a modest error rate could lead to significant inaccuracies when multiple tasks are involved. While Karpathy is critical of the current state of AI agents, he is not entirely pessimistic about the future of artificial intelligence. He describes his projections as being more conservative compared to the optimistic views often depicted in tech circles, yet he remains hopeful amidst a growing number of skeptics regarding AI advancements.