Elon Musk backs Anthropic's call to slow down frontier AI

AI leaders who normally trade jabs are suddenly finishing each other’s sentences.

Anthropic CEO Dario Amodei published an essay calling for frontier AI labs to pace capability gains so safety research can catch up. Hours later, Elon Musk reposted the piece on X with a three-word endorsement: Dario is right.

Musk is not alone. OpenAI CEO Sam Altman pledged to welcome outside evaluators, according to The Guardian. Google DeepMind chief Demis Hassabis told Axios he backs the general direction while reserving judgment on the specifics. Seeing the heads of the most competitive labs in the world nod in unison is anything but routine.

What is spooking the frontier

At the heart of Amodei’s essay is a self-reinforcing loop: advanced AI is already helping researchers build even more powerful models. He warned that, under certain conditions, a swarm of capable agents acting against human interests could potentially take over the internet within six to 12 months, though the timeline remains uncertain.

The anxiety is grounded in internal lab benchmarks. In an announcement dated September 8 and updated September 10, OpenAI revealed that an unreleased model—described as substantially more capable than GPT-6 Astra—produced a proposed solution to the Navier–Stokes existence and smoothness problem. The experiment ran roughly 10,000 agents in parallel, with researchers continually redirecting resources and pooling intermediate findings.

While mathematicians have yet to independently verify the result, the scale of internal progress has clearly unsettled researchers. AP reported that AI researcher Jacob Coxon personally estimates a 10% chance of human extinction caused by AI within the next decade, though consensus on catastrophe odds remains nonexistent across the field.

Amodei’s response is a three-part framework. Anthropic is unilaterally adopting the first step: granting third-party evaluators permanent, employee-level access to internal infrastructure. These auditors can inspect alignment during training and report safety incidents. Amodei also proposed common standards across democratic nations and eventual coordination with authoritarian governments, allowing labs to continue training while capping the speed of capability leaps.

The pushback against artificial brakes

Slowing down, however, immediately triggers resistance from developers and politicians.

Open-source advocates worry the plan would close the door behind the incumbents. Hugging Face CEO Clement Delangue urged broader participation and transparency, according to The Guardian, cautioning that safety pacts written exclusively by frontier labs could concentrate control over who gets to build advanced intelligence.

Venture capitalist David Sacks went further. Axios reported that Sacks challenged labs to slow down voluntarily while alleging regulatory capture, arguing that self-styled safety rules conveniently favor the multi-billion-dollar labs helping write them.

Then there is geopolitical reality. On September 13, Donald Trump stressed the need for the U.S. to stay ahead of China, remarking that whoever wins with AI wins, according to AP. While Trump acknowledged potential guardrails, national security imperatives make voluntary restraint almost impossible to police across borders.

Why it matters

When rival founders agree to tap the brakes, it usually means unreleased frontier systems are behaving in ways that startle their creators. While users will not see immediate product changes, this shift hints that agentic swarms and multi-thousand-model orchestrations are progressing far faster behind closed doors than public chat interfaces suggest.

Anthropic’s move to give outsiders employee-level access during active training is a genuine step toward accountability. It replaces vague marketing promises with observable audits. Yet we think the plan faces an obvious paradox: without enforceable international treaties, voluntary lab agreements crumble the moment one player suspects a competitor is quietly pulling ahead.

There is also a real danger of collateral damage for independent builders. If safety pacing hardens into mandatory compliance regimes, the massive compute overhead of ongoing audits could box out open-source developers while cementing the dominance of today's incumbents. The industry needs safety guardrails, but it must ensure those guardrails do not simply become a moat.