Amodei's Three-Step Plan to Pace AI Faces Deep Verification Challenges

Read "We Must Pace the Frontier" as someone with a frontier model sitting in a production dependency, and you are looking for two things: a capability threshold that would stop a release, and a date. The first of Dario Amodei's three steps is neither. It is an access commitment, and Anthropic says it is making that one on its own ^[2]. Sam Altman committed OpenAI to matching it within hours of publication ^[3]. Two labs now sit under the same pledge ^[17].

The comparison SiliconAngle's column reaches for is SALT I. Nixon and Brezhnev signed in May 1972 after years of negotiating over weapons capable of destroying both countries, and neither government would take the other's word ^[10]. Inspectors inside the Soviet Union were politically unacceptable. So the agreement leaned on what it called "national technical means of verification": each side watching the other through its own technology, including reconnaissance satellites ^[11]. The column puts the lesson as "Agreement was desirable. Verification was indispensable." ^[12]

For a buyer, the split runs between step one and the other two. Verification asks a lab to expose more of what it is doing; pacing asks it to do less. The column argues the first can coexist with a market that rewards speed, and that collective restraint asks that market to work against its own incentives ^[13]. It then asks who determines the correct pace, listing the United States, China, Europe, the frontier labs and an international organization as candidates. It asks who determines when someone has crossed the line. And it asks what happens when a country decides the cost of slowing down outweighs the risk of breaking the agreement ^[14].

By the column's reading, Amodei's own essay puts a genuine global slowdown as the least likely of his proposed outcomes ^[15]. The reaction fits. Of the seven people the column names, four supported the call and three rejected or questioned it ^[18]. Mark Zuckerberg argued that competition, legal liability and independent evaluation already give labs enough reason to build safely ^[7]. David Sacks said companies that believe slower progress is necessary can slow themselves down ^[8]. Michael Burry called the slowdown talk self-serving and argued it could benefit incumbent labs facing growing competition ^[9]. Among the four supporters, one attached a commitment on behalf of a company. Elon Musk endorsed the broader call ^[4]. Demis Hassabis said the essay points toward the right path forward ^[5]. Ursula von der Leyen backed the call to "pace the frontier" and said she would convene the major labs [c6, c19].

Two tests sort any safety promise a vendor makes to you. First, whether keeping it requires the vendor to disclose more or to build less. Second, whether it survives a competitor not joining. The evaluator pledge is a disclosure one lab can make alone. A rival could match it the same day ^[13]. Shared standards and an agreement with China fail both tests. The diagnosis they rest on is that capability is advancing faster than our ability to understand and control it, particularly as AI becomes capable of building the next generation of AI ^[20].

The column describes the evaluator access only as employee-level and aimed at verifying safety practices, and does not say what data or systems an evaluator would see ^[21]. For a buyer, that scope is the part of this plan that could reach a contract.