Inside the Mind of Anthropic CEO Dario Amodei | The Circuit | Extended Interview

Anthropic recorded more than 3x quarterly revenue growth in Q1 2026, translating to 80x annualized growth. This unexpected rapid growth made the planned 10x annual compute increase insufficient.

Dario Amodei estimates the probability of civilizational collapse at 10-25%, a risk arising from the unpredictability of AI technology and the development race among numerous nations and companies.

Anthropic's Mythos model found 271 new vulnerabilities in Firefox, but its public release was postponed due to the risk of cyberattacks. Early access companies reacted with comments like 'superweapon' and 'needs a gun license'.

Regarding his departure from OpenAI, he stated that not only technical safety disagreements but also a lack of trust and misalignment of values were major factors, saying, 'If you can't trust someone and your values don't align, it's better to go your separate ways than to argue'.

Anthropic adopted a strategy focusing on the enterprise market because consumer models rely on advertising revenue and promote addiction or 'slop', whereas enterprise aligns with long-term trust and the company's values.

Anthropic voluntarily blocked model access to China, incurring hundreds of millions of dollars in losses. This accounted for a significant portion of their revenue at the time, but they believed it was crucial for the US and democratic nations to lead China in AI capabilities.

AI has the potential to compress 100 years of scientific and medical progress, significantly improving the quality of human life through disease treatment, energy efficiency, and educational enhancement.

Regarding government involvement in AI companies, he argues that a system of mutual 'checks and balances' is necessary due to risks from both private and government control. Anthropic has introduced public governance elements, giving the Long-Term Benefit Trust the power to appoint and remove board members.

Anthropic's corporate culture and unity are key to fast product development, additionally accelerating model and product development by using Claude itself. As the company grows rapidly, with more hires from big tech, Dario and Daniela Amodei spend half their time maintaining corporate culture.

Current cyber defense measures (classifiers) are jailbreakable and insufficient to defend against powerful models like Mythos, he assessed. The policy is to wait for public release until defense measures, including those of other companies, achieve sufficient reliability.