Anthropic The Chief Executive Officer Dario Amodei releases the latest blog, suggesting that the pace of advancing cutting-edge technologies should be slowed down. AI There is a need to increase the speed of model capability improvement, and three directions for advancement are proposed. Recently, AI security debates have heated up again, with the accelerated improvement of model capabilities also leading to a significant increase in discussions about risk control within the industry.
Anthropic first commits to introducing on-site evaluations
The first measure proposed by Amodei is to have third-party organizations send "on-site evaluators" to AI company to verify whether the enterprise complies with existing safety commitments and to supervise whether safety incidents are reported in a timely manner.
He gave an example, stating that such arrangements could be undertaken by external organizations similar to METR. The relevant personnel would be provided with company ID cards, desks, and computers, and their scope of access would be roughly similar to that of the internal risk assessment team of the enterprise, although there would be exceptions under legal or contractual restrictions. Amodei indicated that Anthropic would unilaterally take the lead in implementing this approach, while also calling on the government to require other leading AI companies to follow suit.
Call on leading companies to coordinate safety standards
The second direction is for leading 'democratic countries' to coordinate the establishment of common security standards and to set limits on the uncontrolled rate of capability enhancement. Amodei believes that such discussions would be best coordinated by the U.S. government, or at least limited exemptions should be provided to reduce concerns among companies about being reluctant to communicate due to anti-monopoly issues.
This statement also responds to the realities of the industry. Currently, there is fierce competition among leading model companies, and relations between them are tense. Direct discussions about slowing down research and development simultaneously may expose them to regulatory scrutiny. According to Amodei, the government may not necessarily need to be directly involved, but it should create conditions for communication on specific security issues.
Incorporate Chinese factors into the plan.
In response to the common objection that slowing down research and development will allow China to catch up, Amodei also made a comment. He believes that if the U.S. government and tech companies continue to restrict the flow of high-performance chips and semiconductor manufacturing equipment to China, and crack down on practices such as model distillation, the United States has the opportunity to further expand its lead in the next 3 to 5 years.
On this basis, he also proposed a third direction, which is to promote limited coordination at a global level. According to him, the United States and its allies should engage with authoritarian governments as much as possible, including cooperation with China. The focus is not on full coordination, but rather on seeking consensus on a few high-risk areas first, such as restricting the use of AI for the manufacture of biological weapons.
Security debates continue to intensify.
Before the release of this statement, Anthropic had already attracted attention due to the public criticism of the risks associated with AI by internal researchers. Amodei did not directly respond to individual resignations in the statement, but mentioned two changes that made him more inclined to proceed with caution: one is the security incidents related to OpenAI and Hugging Face, and the other is that in recent months, AI has significantly accelerated its efforts in developing the next generation of AI.
However, this plan also faces two types of skepticism. Some who support accelerated development argue that Amodei overemphasizes long-term disaster risks, which fuels a sense of backlash within the industry; on the other hand, critics believe that such arguments may allow leading companies to strengthen their positions under the guise of security, while ignoring the real problems that AI has already caused.
Amodei stated in the text that he still believes that AI can significantly improve the quality of human life, but on the premise of advancing development in a more cautious manner and truly using the gained time for security construction.
