OpenAI Co-founder Warns It May Take a Decade for AI Agents to Evolve

OpenAI co-founder Andrej Karpathy has voiced skepticism about the current capabilities of AI agents, asserting that functional AI agents may be a decade away. Karpathy shared his insights during an appearance on the Dwarkesh Podcast last week, where he highlighted the significant limitations that still affect these technologies.

According to Karpathy, the existing AI agents lack sufficient intelligence and versatility. He stated, “They just don’t work. They don’t have enough intelligence, they’re not multimodal enough, they can’t do computer use and all this stuff.” He emphasized that these agents are unable to learn continuously or retain information, describing them as “cognitively lacking.” This stark assessment underscores the challenges ahead in the development of AI technology.

Karpathy’s comments come at a time when many in the tech industry have high expectations for AI agents, with some investors predicting that 2025 will be “the year of the agent.” These agents are envisioned as virtual assistants capable of autonomously completing tasks, breaking down complex problems, and executing plans without direct user input. Yet, Karpathy’s insights reveal a different reality.

In a follow-up post on social media platform X, Karpathy reiterated his concerns about the industry’s trajectory. “My critique of the industry is more in overshooting the tooling w.r.t. present capability,” he noted. He criticized the prevailing narrative that suggests fully autonomous AI entities could soon take over coding tasks, arguing that such a future would render human programmers obsolete.

Karpathy envisions a more collaborative relationship between humans and AI, where AI tools can assist programmers by providing relevant documentation and engaging in dialogue when uncertainties arise. He expressed a desire for AI to facilitate learning rather than simply delivering extensive code without context. “I want to learn along the way and become better as a programmer,” he explained, highlighting a more constructive approach to AI development.

The concerns raised by Karpathy are echoed by others in the field. For instance, Quintin Au, a growth lead at ScaleAI, pointed out in a LinkedIn post that AI agents currently face a significant error rate. He noted that there is roughly a 20% chance of error for each action an AI takes, meaning that if an agent is tasked with completing five actions, the likelihood of achieving perfect accuracy drops to just 32%.

Despite his reservations about AI agents, Karpathy clarifies that he is not a skeptic of AI overall. He describes his timelines for AI progress as “5-10X pessimistic” compared to the more optimistic views often expressed at industry events. Nonetheless, he remains cautious, particularly in light of a growing number of individuals who question the trajectory of AI advancements.

As discussions around AI agents continue, Karpathy’s insights serve as a reminder of the complexities and challenges that lie ahead in the pursuit of truly functional AI systems. The journey to develop reliable AI agents is likely to be long and fraught with obstacles, calling for patience and a realistic understanding of the technology’s current limitations.