Nhà khoa học trưởng của OpenAI cảnh báo về sự tiến bộ nhanh chóng của AI hướng tới khả năng tự cải tiến lặp lại

The Huoxing Finance report states that OpenAI's Chief Scientist, Jakub Pachocki, published an article titled "An Alien Mind" on September 6, 2026. The article reviews how, in mid-2023, the RLSlow research project first produced results demonstrating scalable training of reasoning models, noting that three years later, reasoning language models have become a key driver of rapid economic growth, beginning to push the boundaries of science by operating computers and graphical interfaces, collaborating with humans and other AIs on research projects, while simultaneously transforming the landscape of computer security and introducing new dangers. Based on internal results, the author anticipates that the current pace of progress may continue toward recursive self-improvement. If AI development continues along its current trajectory, systems in the coming years may exhibit equal or even greater leaps in capability, increasingly driving their own evolution. The author calls for extreme caution, arguing that no one is prepared for the consequences of the sustained and rapid rise of machine intelligence. OpenAI will continue seeking technical solutions for alignment and monitoring, building defensive systems, and unilaterally halting further scaling if necessary, but believes broader intervention is required. The article states that machine intelligence is primarily driven by increased computational power, with AI evolving more through growth than design. In alignment, it distinguishes between goal alignment and value alignment, with the core challenge being generalization. GPT-6 Astra shows significantly better alignment than GPT-5.6 Sol, yet further progress is still needed. Regarding monitoring, the primary focus is on chain-of-thought monitoring; evaluations indicate that reliance on capability is gradually weakening. In terms of scalable defense, models are becoming superhuman in their ability to breach computer systems, placing us in a narrow window where using the best available models can significantly enhance the security of critical systems.