Andrej Karpathy, the influential OpenAI co-founder and former Tesla AI director, has publicly backed a coordinated effort to slow artificial intelligence development. His endorsement comes as leading AI labs report that models are approaching the ability to improve themselves autonomously, a capability that could accelerate progress but also poses significant safety risks.

What Is Recursive Self-Improvement and Why Does It Matter?

Recursive self-improvement, or RSI, refers to AI systems that can design and build improved versions of themselves, then repeat the process with each new generation. Unlike traditional software development where humans write code for the next version, RSI would allow AI to take over much of that work. The concern is that as AI systems become faster at this process, improvements could accelerate exponentially, potentially outpacing human oversight.

Anthropic recently revealed that its Claude model now leads 26% of the company's model research and development work, completing most tasks "end-to-end from a high-level prompt" while remaining under human supervision. However, the models are not yet working completely autonomously. OpenAI announced this month that it has developed an automated "research intern" capable of handling well-defined research tasks that would normally take skilled researchers several days to complete.

Why Are AI Leaders Calling for a Slowdown?

The push for a coordinated slowdown stems from concerns that AI capabilities are advancing faster than safety measures can keep pace. Dario Amodei, CEO of Anthropic, published an essay titled "We Must Pace the Frontier" citing the OpenAI-Hugging Face incident, where an AI "swarm" conducted cybersecurity attacks, as evidence of AI's potential for "catastrophic damage" if capabilities grow without guardrails.

Karpathy's support carries particular weight in the AI community. He played a key role in developing Tesla's self-driving and AI technology before leaving the company in 2022 and joining Anthropic in May. On social media, he stated: "I love this and really hope we can come together as an industry and make it happen," referring to Amodei's proposal for paced development.

"I love this and really hope we can come together as an industry and make it happen," Karpathy said while sharing Amodei's essay on social media.

Andrej Karpathy, OpenAI Co-Founder, now at Anthropic

Karpathy is not alone in this position. Sam Altman, CEO of OpenAI, agreed with Amodei's proposal, stating: "I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks." Elon Musk, CEO of SpaceX, also endorsed the effort, simply saying "Dario is right" in response to Amodei's post.

Sam Altman

How Are Different AI Companies Responding to RSI Concerns?

The AI industry is divided on how to proceed. Anthropic has proposed a three-step plan that includes stronger safety testing, independent evaluations of advanced models, and international cooperation among leading AI companies. OpenAI has stated it does not yet know how to "safely get all the way to aligned, full RSI," adding that the company "cannot assume that progress in alignment and safety will keep pace" with capability improvements.

Not all leaders agree on the slowdown approach. Mark Zuckerberg, CEO of Meta Platforms, argued that competition and liability give AI companies sufficient incentive to act individually on safety, without needing coordinated slowdowns. Mustafa Suleyman, CEO of Microsoft AI, said the company is moving toward "humanist superintelligence," which would be "carefully calibrated, contextualized, within limits" rather than "an unbounded and unlimited entity with high degrees of autonomy".

Mark Zuckerberg, CEO of Meta Platforms

How to Understand the Stakes of Recursive Self-Improvement

• Speed of Improvement: AI systems operate much faster than humans, meaning that once RSI begins, improvements could accelerate exponentially, making human oversight increasingly difficult to maintain in real time.
• Safety Testing Challenges: More capable AI systems become harder to monitor and control, creating a gap between the pace of capability advancement and the pace of safety research that can keep up with it.
• International Coordination: Without agreement among leading AI labs and governments, individual companies pursuing RSI could create a competitive dynamic where safety concerns are deprioritized in favor of speed and capability gains.

Karpathy's backing of the slowdown proposal is significant because he represents a bridge between the academic AI research community and industry practice. His work at Tesla on autonomous driving demonstrated both the promise and the challenges of deploying advanced AI systems in real-world, safety-critical applications. His current role at Anthropic, one of the leading voices calling for paced development, positions him at the center of this debate.

The timeline for achieving fully autonomous RSI remains uncertain. Elon Musk suggested that xAI's Grok models could reach full automation by the end of 2027, while OpenAI has set a target of March 2028 for developing an automated AI "researcher." These timelines underscore the urgency with which industry leaders are discussing safety measures.

As the AI field moves closer to systems that can improve themselves, the debate between accelerating progress and ensuring safety will likely intensify. Karpathy's public endorsement of a coordinated slowdown signals that even those deeply embedded in AI development recognize the need for caution as the technology approaches a potential inflection point in its capabilities.