Tuning In
AI researcher Andrej Karpathy, formerly of OpenAI and Tesla, thinks the hype around “agentic AI” is getting ahead of itself.
In an October 2025 Decoder interview, he told podcaster Dwarkesh Patel we’re not entering the year of agents but the decade of agents. The reason? Today’s models still lack memory, multimodal reasoning, and reliable task execution. “They just don’t work,” he said flatly.
Signal Strength
Karpathy’s view cuts through the optimism driving today’s “agent economy.” He sees value in narrow, autocomplete-style use cases but not yet in autonomous problem-solving. The training data itself, he argues, is “total garbage”; a web of fragments and noise that rewards memorization over understanding.
His proposed fix is evolutionary, not revolutionary: better curation, better data, better architectures, better hardware. No single breakthrough—just many small, coordinated ones. “I feel like the problems are tractable,” he says, “but if I average it out, it just feels like a decade to me.”
Return Signal
It’s a timely reminder that the next leap won’t come from another model drop. Instead, it’ll come from re-examining what and how we teach these systems to learn. Progress may be steady, not spectacular. As Karpathy puts it, the real work ahead isn’t making AIs act like us—it’s teaching them to understand what they’re doing.
Logging the Frequency
“I feel like the problems are tractable, they’re surmountable, but they’re still difficult. If I just average it out, it just feels like a decade to me.” — Andrej Karpathy
Source: The Decoder: “Andrej Karpathy says agentic AI is years away from matching industry hype”
