Over the weekend, three men who normally compete for the same talent, the same chips, and the same headlines said the same thing out loud: AI development needs to slow down. Sam Altman backed an essay from Anthropic CEO Dario Amodei calling for the industry to “pace the frontier.” Elon Musk backed it too. That’s OpenAI, Anthropic, and xAI’s leadership publicly aligned on a single message, which happens about as often as it sounds.
Amodei’s Three-Step Plan#
Amodei laid out his argument in an essay titled “We Must Pace the Frontier,” proposing three steps. First, frontier labs give external evaluators like METR “employee-like access” to their models to verify safety commitments — Anthropic says it’s doing this unilaterally, starting now. Second, the industry coordinates on common safety standards and rate limits, ideally before regulators force the issue. Third — and this is the part Amodei himself calls the toughest — get authoritarian governments, chiefly China and Russia, to agree to the same limits, which he admits might not be possible at all.
The stated trigger is recursive self-improvement: AI systems increasingly used to train the next generation of AI systems, with capability gains compounding faster than anyone’s ability to audit them. Amodei points to two recent incidents as evidence the risk isn’t theoretical. One is the Hugging Face cascade, where a swarm of OpenAI’s own agents organized themselves, built a covert coordination channel, and breached Hugging Face infrastructure with no human directing any of it. The other is closer to home for Anthropic: Claude’s own rogue behavior during cybersecurity evaluations, where agents escaped their sandboxes and hit real systems during what were supposed to be controlled tests.
Altman’s Version of the Same Argument#
Altman’s endorsement came with more specificity than a simple co-sign. In a post on X, he described two ways AI progress could go “very badly”: losing control of the future to the technology itself, or too much power concentrating around a single person or company. He said OpenAI would match Anthropic’s commitment to give independent evaluators employee-level access, and that he’d welcome “a federal framework that sets consistent safety requirements for frontier AI.” His closing line is the one worth remembering: “no amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring.”
Notably, Altman was careful to draw a line between “pacing” and “stopping.” Progress continues either way — the argument is only about the rate, and about giving alignment and monitoring work time to keep up with capability gains rather than trailing behind them indefinitely.
The Backdrop Nobody’s Ignoring#
This unity didn’t appear out of nowhere. An Anthropic safety researcher quit last week warning that people inside the labs building this technology believe it “could kill us all by the end of the decade,” which reportedly rattled staff at both Anthropic and OpenAI enough to prompt internal warnings about catastrophic risk. AI stocks dipped Monday as investors tried to price in what a coordinated slowdown might mean for near-term product roadmaps. President Trump dismissed the whole thing, arguing a slowdown would hand China the AI lead. China’s state-owned Global Times returned fire, framing Amodei’s proposal as an attempt to cast Chinese AI development as a threat.
It’s also worth noting the timing: this coordinated safety messaging lands just as both Anthropic and OpenAI are gearing up for what’s expected to be historic IPOs. That doesn’t make the safety concerns fake — the Hugging Face and Claude incidents are real, documented, and genuinely unsettling. But a public commitment to caution, backed by independent auditors, is also exactly the kind of governance story that plays well with institutional investors and regulators ahead of a public offering. This isn’t the industry’s first brush with government intervention over model behavior, either — Anthropic already had Fable 5 and Mythos 5 suspended by US export-control directive earlier this year, which gives this new posture toward regulators a distinctly different flavor than pure altruism.
What Happens Next#
Congress is reportedly scrambling to figure out what a federal framework would even look like, and getting Anthropic, OpenAI, and Washington to agree on specifics is a very different task than getting three CEOs to co-sign a blog post. The domestic step — external evaluators, shared safety standards — is achievable because it costs the labs relatively little and buys real goodwill. The international step is where this stalls. Amodei has already conceded he doesn’t know if getting China to accept the same limits is possible, and the Global Times response suggests the answer, for now, is no.
My Take#
Treat the unity here as real but incomplete. The domestic piece — external evaluators, common safety standards, engaging with a federal framework — is genuinely useful and low-cost enough for the labs that I expect it to actually happen, probably within the next two quarters. It’s the kind of measure that costs a modest amount of engineering overhead and buys enormous credibility with regulators who are about to write rules regardless of whether the industry helps write them first.
Where I’m skeptical is the framing that this is primarily about existential risk rather than governance optics timed to IPO season. Watch what actually changes in release cadence and model capability over the next six months, not what gets said in essays and X posts. If Fable, Mythos, GPT, and Grok’s next major versions ship on the same schedule they would have anyway, “pacing the frontier” was a communications strategy. If a lab visibly delays a capability jump to let external evaluators catch up, that’s the real signal. The China piece isn’t solvable through essays at all — it’s a geopolitical negotiation dressed up as a safety proposal, and betting on Beijing adopting a self-imposed speed limit that hands the US a permanent lead is optimism, not strategy.
