Dario Amodei's 'We Must Pace the Frontier' essay ignites the biggest AI safety debate of the year

Dario Amodei's "We Must Pace the Frontier" Ignites the Biggest AI Safety Debate of the Year

Dario Amodei released a major essay on Saturday morning (September 12) calling for coordinated slowdowns in frontier AI development. Within hours, Elon Musk, Sam Altman, and Demis Hassabis publicly endorsed the proposal. By Sunday, Trump rejected it outright. What began as a safety-focused policy argument has become a flashpoint between tech industry leaders, regulators, and the White House over whether AI should be slowed down, and if so, by whom.

The rollout was deliberate. Amodei posted the essay at 7 AM Pacific, Musk endorsed by 8:30 AM, and Altman followed at 4 PM the same day. This was not accidental coordination—OpenAI had earlier told Axios it helped shape language in a July open letter on pacing, and Altman had discussed pacing with White House officials. By mid-August, OpenAI published its own pacing proposal, citing slowed scaling and a paused reinforcement learning run. Anthropic's official post followed, calling for coordinated mechanisms.

The Three-Part Proposal

Amodei's essay puts forward three main mechanisms:

Independent third-party evaluators. Anthropic wants external auditors with direct access to AI companies' systems to verify safety practices. The proposal floats MIRI as a possible candidate, which immediately drew criticism—several people have moved between MIRI, OpenAI, and Anthropic, making true independence questionable. Some view this as a feature (safety-focused evaluators who understand the stakes) rather than a bug, but conflicts of interest are real.

Government-coordinated safety standards. Frontier labs and democratic governments should establish formal AI safety standards together. This framing dodges the core legal problem: companies agreeing in concert to restrict output or capabilities may violate the Sherman Antitrust Act. Amodei explicitly calls for an exception to antitrust law so labs can coordinate without fear of prosecution. The analogy is crude but instructive—if American and Delta agreed to fly fewer planes "for safety," that would drive margins up and consumer prices up, which is exactly why cartels are illegal.

International coordination with authoritarian regimes. The US and democratic countries should negotiate AI pacing with China. This proposal is strategically murky. China is still catching up to frontier capabilities. If both countries agree to slow down equally, does the leading power benefit? Amodei doesn't address the equilibrium problem—in nuclear nonproliferation, every state wanted at least some capability, creating a multipolar world. AI could follow the same pattern.

The Immediate Pushback

David Sacks, White House advisor, delivered the sharpest rebuttal on Saturday evening. His post cuts through the proposal's framing:

Anthropic and OpenAI control frontier capabilities by any reasonable metric. Sacks argues they should simply slow down unilaterally if models are truly dangerous, rather than demand government permission or Sherman Act waivers. He flags what looks like regulatory capture—requesting antitrust exemptions, embedding friendly evaluators with ties to Anthropic, creating review processes that would stifle smaller competitors like Instinct (which cannot afford to spend a year negotiating with evaluators and government agencies).

Gavin Baker, CIO of Atreides Management, called the third-party evaluator pledge "the only tangible new fact" from the weekend. He noted that liability protections matter—companies with Section 230–style shields survived; others went bankrupt. Showing duty of care will matter in future litigation, making embedded evaluators strategically valuable. But he questioned whether the broader proposals change the investment thesis materially.

The Trump Response

Trump rejected the entire premise. On Sunday morning, he posted that the only guardrail AI needs is "a strong and smart high IQ president," and the US already has one. His follow-up posts treated pacing as a conspiracy against American AI dominance and lumped it in with his standard grievance list (Russia hoaxes, impeachment, etc.). He called AI "bigger than diamonds" and the Internet and said China is doing nothing to slow down, which he framed as a threat to US leadership.

Thirty minutes later, he posted again, asking rhetorically: when have industry leaders ever called for regulation that would drive them into bankruptcy? He dismissed AI safety concerns as a hoax, compared it to his usual targets, and ended with: "It will not be stopped by brilliantly run destructive forces during the term of president, Donald Trump."

The Real Tension

The commercial math is genuinely ambiguous. Pacing could benefit or harm frontier labs depending on how it plays out.

One reading: slower frontier development lets competitors catch up (Meta, DeepMind, Grok), fragmenting the current duopoly of OpenAI and Anthropic into a four- or five-way market with lower margins. This would be bad for business.

The other reading: pacing buys time to establish regulatory moats and lock in third-party evaluation frameworks that smaller labs and open-source competitors cannot easily replicate. Instinct, for instance, cannot afford a year-long regulatory dance. This would entrench the leaders.

Pricing pressure already exists, but not from Chinese models. OpenAI and Anthropic are discounting against each other (price competition within the duopoly), and customers are using models more judiciously—avoiding expensive frontier models for lightweight tasks. If pacing leads to more competitors at the frontier, that pressure compounds and margins suffer. Sam Altman has told investors OpenAI is structured as a public benefit corporation and will prioritize safety over shareholder returns if needed, and investors signed up knowing that. Dario has similar cover.

No one is giving them credit for that structural choice. Critics assume the proposal is cynical 4D chess to 10x the market cap by establishing favorable regulation. But the simpler explanation—that slowing down frontier AI is actually bad for their business, and they're willing to accept that cost—gets almost no airtime.

What's Unstated

The transcript flags but doesn't resolve the electricity constraint. If pacing means frontier labs train fewer chips and consume fewer watts, that capital and human talent could redirect to energy infrastructure—nuclear, solar, grid reliability. Everyone claims to want cheaper, abundant energy. No one is arguing for slower power plant construction. Yet energy supply is arguably the harder constraint on AI scaling than safety evaluations.

Trump's framing—that pacing is anti-American competition—may resonate more broadly than the safety case, especially if Chinese labs do accelerate in response. That's the real geopolitical risk Amodei's proposal doesn't solve.