Amodei Urges AI Speed Limits, Warning Recursive Self-Improvement Could Hit the Internet in a Year

Dario Amodei, chief executive of Anthropic, is pushing for a deliberate slowdown in AI development, arguing that recursive self-improvement could put the entire internet at risk within six to twelve months. His warning lands days before what may become the largest initial public offering in history, a timing that gives the proposal unusual weight. Amodei is not asking labs to stop building; he is asking them to accept limits before the technology outruns the people supervising it.

The core of his argument is that once models can meaningfully improve their own successors, the pace of change stops being set by human review cycles. That compression matters because safety work, red-teaming, and deployment decisions all assume a window of months between capability jumps. If that window shrinks to weeks, the institutions meant to catch failures are working on stale assumptions. Amodei has framed the six-to-twelve-month figure as a planning horizon, not a prediction, but it is far shorter than most public roadmaps assume.

His proposed remedies are concrete. He wants auditors embedded inside AI companies rather than reviewing from outside, so they can see training runs and evaluation results as they happen. He also calls for shared safety standards across labs, so no single company gains an advantage by cutting corners, and for international agreements modeled on the SALT disarmament treaties that constrained US and Soviet nuclear arsenals. That last comparison is deliberate: SALT worked because rivals accepted mutual limits and verification instead of waiting for a crisis.

The IPO context sharpens the tension. A public listing would put enormous pressure on Anthropic to show revenue growth and ship products faster, the opposite of the slowdown Amodei is describing. Critics of the industry's safety rhetoric will note that calls for restraint often arrive alongside competitive positioning. Supporters counter that a CEO warning about his own product category, weeks before a market event that could enrich him, is not a cheap move.

What remains unresolved is enforcement. Embedded auditors require companies to grant access they currently control, and shared standards require rivals to agree on thresholds nobody has defined. Treaty-style limits would need verification mechanisms that do not yet exist for software. Amodei has set out the shape of a regime without explaining who signs it first, and that gap is where the next phase of this debate will play out.