Anthropic CEO Dario Amodei says the company has sacrificed business and restricted access to powerful AI capabilities to reduce safety risks. Nathan Calvin, chief strategy officer at AI-safety nonprofit Encode AI, argues that its compute spending, internal development practices, and policy proposals point toward a different priority: winning the race to superintelligence.
Their public exchange, covered in Business Insider’s October 21 report, centers on how Anthropic reconciles increasingly powerful AI development with its positioning as a safety-focused company.
In a roughly 2,000-word response, Amodei defended what he calls a “race to the top,” arguing that staying at the technological frontier gives Anthropic greater influence over industry standards. He offered nine decisions he characterized as commercially costly but beneficial for safety. He also acknowledged that several matters Calvin identified could have been handled better, without accepting Calvin’s broader accusation.
The disagreement concerns both the strength of Anthropic’s safety commitments and who should have authority to restrict dangerous AI development or deployment.
Compute Spending Is Part of the Strategy, Amodei Says
Calvin’s critique linked Anthropic’s pursuit of more computing capacity with its growing use of AI to develop subsequent models. Business Insider reported that he cited compute deals exceeding $500 billion and increasing reliance on Claude in AI development.
That figure is part of Calvin’s argument as reported by the publication. It should not be read as an independently verified amount Anthropic has already spent, or as the cost of training a single model.
According to explainx’s account of Calvin’s post, his other examples included an in-house custom-chip team and substantial compute devoted to internal AI research and development. He presented these activities as a pattern in which reaching superhuman capabilities takes precedence over safety.
Amodei explicitly defended Calvin’s first four examples as ordinary consequences of competing at the frontier, consistent with Anthropic’s mission.
“Your items 1-4 are all simply examples of what it means to compete at the frontier, and we stand by them.”
Where Calvin sees accelerating development as evidence of the company’s priorities, Amodei sees technological competitiveness as necessary to its safety strategy. In his view, demonstrating responsible development carries more weight when the company doing it is building leading systems.
Amodei acknowledged a limit to that reasoning. He said important questions arise when AI-driven acceleration becomes fast enough that safety struggles to keep up, adding that Anthropic is “nearing that point.” He connected that concern to the company’s thinking about pacing development, while maintaining that the activities Calvin listed occurred before that threshold.
He defended the investments without claiming that acceleration can continue safely under all circumstances.
Restricted Access Is Amodei’s Main Evidence of Sacrifice
Amodei’s strongest counterexamples involved decisions that limited who could use Anthropic’s models or what those models would help users do.
On biological safeguards, he said concerning results in capability evaluations and attempted misuse justified restrictive controls. He acknowledged that reducing dangerous requests that slip through can also increase false positives, blocking legitimate work and frustrating researchers.
Those restrictions have commercial consequences, Amodei argued, because competitors can promote more permissive products. This is his account of the tradeoff, not an independently measured estimate of lost revenue. The specific decision is still open to scrutiny: Anthropic says it is accepting friction for legitimate biology users to reduce the chance of enabling dangerous misuse.
In cybersecurity, Amodei said Anthropic had Mythos-class models in February but waited until June to make them generally available, after improving cyber safeguards. Before that broader release, he said, access was limited to partners focused on cyberdefense.
He characterized the delay as sacrificing part of a several-month competitive lead. Neither the size of that lead nor its commercial value is independently established by the exchange. What he described was a delay in broad access while selected defenders used the technology.
Restrictions on sales provided another example. Amodei pointed to limits on sales to companies he described as controlled by the People’s Republic of China, including those operating through third countries. He argued that these restrictions forfeited short-term business to avoid supporting military and intelligence applications. Separately, he cited advocacy for chip and chipmaking-equipment export controls despite friction with industry partners.
These examples address the allegation that Anthropic always chooses the commercially easiest path. They do not, by themselves, settle whether its internal development process has adequate safeguards. Restricting customer access and controlling the pace of research are different safety interventions.
The costs can also extend beyond Anthropic’s balance sheet. Amodei acknowledged that restrictions can delay beneficial biological research, defending those tradeoffs as justified despite their costs.
The Draft Federal Bill Tests Who Can Block a Model
The most direct policy dispute concerns a draft federal AI safety bill Anthropic circulated to Congress.
Business Insider reported that critics said the proposal would make it harder for the federal government to block dangerous AI models. Explainx’s account described Calvin’s objections more specifically: he argued that the draft would preempt significant state law and impose a high bar for executive-branch intervention.
These are attributed criticisms of a draft, not established findings about an enacted law. Assessing its precise legal effects requires examining the provisions themselves.
Support for AI regulation does not resolve the dispute. A company can favor government oversight while advocating rules that narrow which institutions can intervene, under what circumstances, and through which procedures.
Amodei cited his earlier public advocacy for giving government the power to block model releases. He also said Anthropic had played a key role in California’s SB 53 and had argued for a stronger version than the one that passed. Calvin’s objection, as reported, concerns whether the newer federal proposal would weaken meaningful intervention despite that record.
The draft bill was among three matters Amodei agreed Anthropic could have handled better. The others were a later-walked-back letter to Congress about Claude hacking incidents and a statement that “you can’t do safety from second place,” according to explainx’s summary.
He described problems involving rhetoric, insufficiently careful policy communication, or changing circumstances that did not travel between teams quickly enough.
“I am of course the CEO, and I take responsibility.”
The acknowledgment had limits. Amodei conceded shortcomings in handling and communication. He did not clearly accept Calvin’s interpretation of the draft bill’s legal consequences or the broader claim that Anthropic had subordinated safety to winning.
Safety Commitments Remain the Harder Test
Beyond the bill, Calvin challenged changes to Anthropic’s Responsible Scaling Policy, its framework for managing risks as models become more capable.
Sources
- Business Insider’s October 21 reportbusinessinsider.com
- explainx’s account of Calvin’s postexplainx.ai
- https://x.com/DarioAmodei/status/2112732230124818891x.com
Editorial standards · Chief editor: Jim Clyde Monge · Report a correction
