Amodei: Well-Designed AI Rules Can Decentralize Power, Not Just Entrench Incumbents

Anthropic CEO Dario Amodei pushes back on the Silicon Valley reflex that treats regulation as synonymous with regulatory capture and the concentration of power in a few firms and politicians. He calls that a false choice, arguing that fair, objective institutional processes can actually distribute power by vesting it in ideas rather than in individuals — comparing a formal court system, however stuffy, to the mob justice it replaces. Rules, in his framing, are not inherently pro-incumbent; it depends entirely on how they are written.

To back this up, he points to Anthropic’s policy track record of proposals designed to slow frontier labs while sparing or advantaging smaller competitors. California’s SB53 (which Anthropic supported) and the more contested SB1047 (which it was ambivalent on) both exempt companies below revenue or training-cost thresholds — $500M in the case of SB53. Testing regimes Anthropic has advocated at CAISI and the White House impose stricter scrutiny on frontier models than on off-frontier ones, and the “Pacing the Frontier” concept would throttle only the leading edge, leaving challengers, including open-weight developers, room to catch up.

Amodei’s deeper claim is that AI concentrates power structurally because of scaling-law economics, not because of regulation, and that open weights only partly offset this by shifting concentration toward whoever controls the most compute and chips. He argues the right rules can simultaneously address cyber, bio, and alignment risks, constrain the frontier labs, and preserve space for open models. He endorses the Trump administration’s reported approach of pre-deployment testing for frontier models — plus testing of open-weight models as they near the frontier — and Demis Hassabis’ idea of a FINRA-like self-regulatory body, framing this as progress from six months earlier when the industry was still lobbying to preempt state regulation with no federal plan in sight.