Anthropic CEO published a lengthy response to safety concerns, stating that it is necessary to stay at the forefront of technology
Anthropic recently submitted a draft AI safety bill to U.S. lawmakers, which has been questioned by U.S. AI safety advocates. The draft sets a high threshold for the government to urgently halt dangerous models, requiring risks to involve scenarios such as biological weapons, cyberattacks, or AI going out of control, and potentially causing over 50 deaths or serious injuries, or losses exceeding $1 billion; some provisions may also limit state regulatory authority.
Nathan Calvin, the strategic head of AI safety organization Encode, subsequently compiled Anthropic's decisions over the past year. In February, the company modified its previously strict safety policy: it had previously committed to pausing training or releases if safety measures did not keep pace with model capabilities, but the new policy no longer requires the company to independently halt R&D at higher risk stages, citing that competitors might take advantage of the opportunity. During the same period, the company invested heavily in computing power, formed a chip team, and assigned Claude more AI R&D tasks.
On October 21, Anthropic CEO Dario Amodei posted a lengthy rebuttal on X. He acknowledged that the company had made some mistakes but insisted that competing for cutting-edge AI capabilities did not violate its safety mission, stating that the company adheres to a race to the top, and only by staying at the forefront of technology can it drive the industry to improve safety standards. He listed nine safety decisions that had cost the company commercially, including restricting access to Mythos-level models for several months and strictly limiting biology-related requests. The company's updated safety roadmap on October 19 states that existing policies mainly target the deployment of dangerous models, and in the future, it may need to directly limit the speed of R&D, planning to propose relevant policy options by January 15, 2027, or explain why none were proposed.