Did one of the founders of OpenAI burst the AI ​​bubble? “Forms do not exist”

In wide participation interview With podcasts Dwarkesh PatelKarpathy, a YouTuber with more than 1 million followers, said he believes the race to build artificial general intelligence is going much slower than the hype suggests.

Despite rapid progress in large language models (LLMs) over the past three years, he said that AGI is still at least a decade away, and warned that many companies are overestimating the powerful capabilities of AI in a way that could harm the field.

“In general, the models aren’t there,” Karpathy said on the podcast. “I feel like the industry is taking too big a leap and trying to pretend that this is amazing, and it’s not. “It’s dirty.”

The interview sparked an immediate reaction across the tech community, as expectations for AGI rose alongside capital investment and competition.

“If this Karpathi interview doesn’t burst the AI ​​bubble, nothing will,” Prithveer Jhaveri, CEO of prediction markets aggregator TradeFox, said. books on X.

John Coogan, host of the TBPN technology podcast, male Karpathy’s interview came just weeks after the AI ​​pioneer Richard Sutton described LLMs as a “dead end.”

“The general tech community is in shock right now,” Kogan wrote on X.

Karpathy, who previously served as senior director of AI at Tesla and helped lead OpenAI in its early years, described his AI timeline as “five to 10 times pessimistic” compared to many public forecasts. But he rejected the gloomy idea that it would take a decade to achieve artificial general intelligence. He said: “Ten years.” books On X after the interview, “It should otherwise be a very bullish timeline for AGI.”

As for Silicon Valley, It’s a slow drop. Sam Altman, Karpathy’s colleague and OpenAI co-founder and current CEO, He predicts Artificial intelligence will surpass the intelligence of any human in any discipline by 2030. Elon Musk predicted that artificial general intelligence will come either this year or next year.

Karpathy argued that much of the confusion stems from metrics that give an inflated sense of ability. Public demos, benchmark competitions, chatbot conversations, and code generation tests tend to reflect narrow improvements, rather than tackling the toughest unsolved problems in AI, he said. These include long-term planning, structured thinking, and ultimately, safe system design.

Karpathy made his strongest criticism of “AI agents,” a concept that has spread across the industry in recent months.

Built on top of LLMs, these systems are presented as independent digital workers who can write and run code, search the Internet, run programs, and perform work tasks with minimal supervision. The idea is promising, Karpathy said, but implementation, at least as it stands today, is far from reliable.

“We are in this intermediate stage,” Karpathy said. “The models are amazing. They still need a lot of work.”

Many other AI leaders are more optimistic. For example, Nvidia CEO Jensen Huang And he called 2025 is “the year of the AI ​​customer,” Anthropic CEO Dario Amodei recently described He said That by 2026 or 2027, AI systems will be “better than almost all humans at almost everything.”

But Karpathy cautioned that most current AI systems produce fragile and unpredictable results and lack basic reliability. He said they do not have sufficient ability to think, have limited perceptions of software environments, and struggle to use tools correctly.

“If this is not done well, we could end up accumulating mountains of errors across software, an increase in vulnerabilities (and) security breaches,” Karpathy said.

However, he insisted that artificial intelligence is still on a Long but solvable road. He said that the technical challenges ahead are difficult, but can be controlled with time, research and better safety practices.

He added: “I feel that the problems can be overcome.” “But it’s still tough.”