Anthropic CEO Dario Amodei Responds to Safety Organization's Concerns
AI safety organization Encode's legal advisor Nathan Calvin published a lengthy post on X, questioning whether Anthropic is a safety organization or a competitor striving to create superintelligence. CEO Dario Amodei subsequently responded personally.
The issue arose after Politico reported that Anthropic privately circulated a draft federal AI bill to Congress. The draft stipulates that the government must prove a model poses an imminent risk of causing over 50 deaths or serious injuries, or over $1 billion in losses, before it can be halted. Calvin believes this threshold is too high and preempts certain state laws.
Calvin pointed out that Anthropic signed approximately $500 billion in computing power agreements over the past year, and the proportion of R&D led by Claude increased from nearly zero at the beginning of the year to 26% in August, while the safety rules were revised to remove the commitment to "proactively pause if risks cannot be controlled." He questioned why mistakes always seem to benefit the company.
Dario acknowledged some mistakes but disagreed with the overall conclusion. He stated that the company's strategy has always been to "compete upwards": first establishing responsible models to raise industry standards, which requires the company to be at the forefront, thus expanding computing power and using AI to accelerate R&D is a normal action.
Dario listed the commercial costs the company has incurred for safety, including strict biosafety filtering that has been mocked by biologists, delays of several months in releasing the strongest Mythos-level models, not providing services to Chinese-controlled companies, and this week establishing a public welfare fund with 5% of the company's shares.
Dario admitted to three communication errors: a letter to Congress, a statement that "safety cannot be done in second place," and the draft bill itself. He believes these were issues of wording and team synchronization, and as CEO, he takes responsibility. Regarding the removal of the pause commitment, he argued that the old rules were too rigid, and the modification was reasonable.
Source: Public Information
ABAB AI Insight
Anthropic, founded by Dario Amodei and others after splitting from OpenAI, has long emphasized responsible scaling and safety research. The company supports transparency bills like California's SB53 while promoting large-scale computing power procurement and model iteration.
On the capital front, the company is expanding its capabilities through computing power agreements and internal AI R&D acceleration. Safety-related restrictions (such as model release delays and export limitations) incur commercial costs, while the company attempts to influence the federal regulatory framework with the draft bill, concentrating intervention rights on specific thresholds and judicial processes.
This is similar to the paths taken by other leading labs between safety commitments and scaling competition, where early safety advocates diverge from later commercialization stages. The current phase is characterized by models crossing critical thresholds while the regulatory framework is still taking shape.
Essentially, this is a regulatory change: the company seeks to replace more decentralized state-level intervention rights with a high-threshold federal framework. Mechanically, leaders shape the competitive environment by defining the standards for "intervenable risks"; when internal rules are relaxed in tandem with external legislation, the relative weight of safety constraints and capability expansion shifts.
ABAB News · Cognitive Laws
Leaders define the rules, while laggards bear the costs of those rules.
Commercial costs can be demonstrated, but decision-making power remains within the company.
Admitting mistakes does not equate to changing direction; the route is determined by strategy rather than individual cases.