With the continuous improvement of cutting-edge model capabilities, OpenAI CEO Sam Altman stated that AI companies should accept a slower development pace and move more security checks to the training phase earlier on, rather than waiting until just before the product is ready to be released to handle them all at once.
Identify two main types of risks
Altman posted on the X platform that the industry needs to avoid two outcomes: one is that humans lose control over AI, and the other is that extremely powerful capabilities are concentrated in the hands of a single company, individual, or country, and are used to impose their will on the outside world.
He stated that AI must always serve humanity; therefore, the advancement of security and alignment technologies must at least keep up with the speed of model capability improvements. In his view, the industry needs to maintain a balance between continuing to advance technology and preventing excessive concentration of power.
Move security assessment to the training phase in advance
Altman says that so-called "slowing down the pace" does not mean stopping research and development, but rather acknowledging that security incidents, monitoring, and evaluation come with significant costs. Therefore, the development speed should be lower than the theoretically fastest level.
He revealed that OpenAI had established clear security protocols before the commencement of training in advanced reinforcement learning, which is expected to significantly enhance capabilities. This indicates that security efforts are no longer limited to the model release phase but have been extended to the training stage as well.
Support federal rules, but practice self-discipline first.
Altman welcomes the security regulations and independent audits at the federal level, but at the same time emphasizes that companies should not wait until the legislation is completed before taking action. He stated that the industry should also advance necessary security cooperation without waiting for anti-monopoly exemptions.
Prior to this statement, there were reports that OpenAI had inquired with the U.S. Congress whether competitors could legally coordinate to slow down their development pace. Altman believes that the more important role of the government is to promote international coordination, while companies should first complete the part that they themselves can accomplish.
Recent test events have increased attention.
This statement also came after the controversy over frontier model testing intensified. Reports mentioned that Anthropic previously discovered during testing that after Claude came into contact with the real system, it tightened isolation and monitoring measures.
On the OpenAI front, a model was isolated, and the originally largest-scale training program was suspended. The reason was that the relevant agents exhibited collaborative cheating during testing and attacked the behavior of Hugging Face.
