The Zhitong Finance App learned that OpenAI chief scientist Jakub Pachocki recently issued a stern warning: artificial intelligence (AI) is rapidly evolving, and humans are becoming increasingly unable to understand and control it. He anticipates that AI laboratories will voluntarily slow down the development process due to safety considerations.
In a long article published on Sunday, Pachocki pointed out that in a relatively short period of time, AI models can control computers, collaborate with humans and other AI systems, and carry out research projects. He further stated that soon, AI will have the ability to self-evolve without human intervention, a process known as “recursive self-improvement.” According to reports, recursive self-improvement is seen as a critical path to achieving general artificial intelligence (AGI), and its self-acceleration characteristics may trigger an “intelligence explosion.”
“This is a time of extreme caution,” Pachocki wrote, “I'm worried that no one is really prepared for the consequences of the continued rapid rise in machine intelligence.”
He proposed that developers either guide AI more closely to align with human interests, or “slow down future development if necessary.” He wrote, “I hope and expect that voluntary slowdown will become the norm in the industry until common safety standards are reached.”
Shouting to stop while running wild: the AI industry's security anxiety and expansion instincts
Pachocki's statement is consistent with the pessimistic remarks of many AI technology executives. Anthropic CEO Dario Amodei has warned that AI may replace half of junior jobs. OpenAI CEO Sam Altman made similar predictions, but he confessed in an interview last week that the AI industry is not doing enough to communicate the potential benefits of technology to the public.
Just as Pachocki warned, the path of recursive self-improvement was no longer limited to theory. Anthropic published an article in June this year, detailing Claude's deep participation in AI research and development processes such as code writing, review, and research, and clearly stating that the path of recursive self-improvement has become visible, and calls on the community to consider slowing or suspending cutting-edge AI development so that aligned research can keep up with technological progress.
In early June, SoftBank Group CEO Sun Zhengyi revised his previous predictions, believing that recursive self-improvement could lead to the arrival of superintelligence within two years. He revealed that after communicating with OpenAI, he learned that “models are designing future models,” and that when models can generate next-generation models and human engineers cannot follow, that means the arrival of superintelligence.
More alarming, OpenAI's own research also confirms this concern. A research paper published by OpenAI in June 2026 indicated that there are “early signs” of recursive self-improvement in the current system, and warns that society and government are unprepared.
However, tech companies are still racing to launch more capable models. Nvidia CEO Hwang In-hoon wrote on the X platform that “the era of general artificial intelligence (AGI) has arrived” in response to the release of OpenAI's latest model.
The rapid development of AI technology occurred at the same time as breakthroughs in the field of cybersecurity, a series of scientific discoveries, and the overcoming of several long-standing mathematical problems. OpenAI said its AI models had coordinated with each other through message boards to attack the research platform Hugging Face undetected. The company chose to limit the release of its latest GPT-6 Astra model, citing its advanced cybersecurity capabilities. For similar reasons, Anthropic has yet to release its cutting-edge Mythos model to the public.
More and more startups are fully pursuing the direction Pachocki warns of — self-improving AI. Investors are flocking to the recursive self-improvement circuit, betting that this is the only way to truly human-like and even surpass human intelligence. Inherent, founded by former DeepMind researchers, raised $50 million earlier this year. Recursive Superintelligence, an AI lab founded by renowned scientist Richard Socher, is also advancing related research and has received US$650 million in financing.
Meanwhile, Pachocki's employer, OpenAI itself, has clearly proposed to create a “true automated AI researcher” within two years.