Karpathy assesses AI agents' realism

In an interview with Dwarkesh Patel, Andrej Karpathy, a former researcher at OpenAI and Tesla, downplayed the excitement around AI agents, estimating that it will take ten years for autonomous AI systems to fulfill their current promises. The specifics: Karpathy says the models "aren't there yet" and that industry messaging is overhyping the current agentic coding capabilities that produce "slop." He claimed that because of basic flaws including low intelligence, multimodal constraints, and a lack of ongoing learning, agents "just don't work." Additionally, Karpathy referred to reinforcement learning as "dreadful" and "noisy," but it appears to be a good thing because "everything we had before it is much worse." Karpathy responded that he would rather to work with the model than compete against it when Elon Musk challenged him to compete against Grok 5 on X. Karpathy is one of the most renowned AI experts, thus his remarks are quite credible. They also serve as a crucial technical reality check for the hype surrounding the "Year of the AI Agent." However, it's also feasible that systems that don't impress a top brain are still incredibly productive for the other 99 percent of users, even in spite of the severe criticisms.