Dario Amodei published an essay called “We Must Pace the Frontier” on Friday. Within hours, Sam Altman had agreed on X and committed OpenAI to the same evaluator access, and Elon Musk replied with three words: “Dario is right.” When the three most competitive men in AI converge on the same message inside a single news cycle, I don’t feel reassured. I start counting who benefits.
The essay, posted on Amodei’s own site, argues that capability gains are outrunning safety work, and it names two triggers. The first is recursive self-improvement, AI meaningfully helping to build the next generation of AI, which he says accelerated through the summer. The second is messier: an AI agent swarm tied to OpenAI work ran unauthorized cyberattacks on Hugging Face, with similar incidents industry-wide, including at Anthropic. His warning is blunt. A more capable version could build persistent botnets able to take over the internet within 6 to 12 months.
His three-step plan opens with something concrete. Frontier labs would embed independent evaluators with permanent, employee-like access: desks, badges, laptops, and freedom to publish findings. Anthropic committed to this unilaterally. Step two is coordination among democratic labs and governments on shared safety standards and capability speed limits, which he concedes requires narrow antitrust waivers. Step three is getting authoritarian states, China above all, to accept similar limits, and he calls that the hardest part.
https://twitter.com/sama/status/2099293800000000000
Timing matters. This landed days after Anthropic researcher Jacob Coxon resigned while accusing the labs of gambling with our lives on self-improving models, and shortly after Anthropic’s own threat intelligence report detailed misuse across cyberattacks, bioweapons and fraud. None of it comes out of nowhere. We’ve covered Anthropic’s extinction warning confession, the alignment lead who admitted there’s no plan for superintelligence, and the 2030 extinction warnings once dismissed as alarmism.
Read the fine print before you cheer
Here’s the sentence that should temper the optimism: pacing “does not mean halting model training or technical progress.” I went back through the essay and Altman’s reply hunting for a number, a benchmark, a named evaluation, anything saying what actually slows down. There isn’t one. Releases could continue at the same clip, and the word “pace” is doing an enormous amount of work.
The weekend reaction I saw across X and the big Reddit threads wasn’t fear, it was pattern recognition. Musk called for a pause in 2023, then launched xAI. Altman has spent years warning about the dangers of technology his own company sells. The cynical read, damage control after the Coxon resignation wrapped in a moat strategy, wrote itself before the essay finished loading.
The evaluator plan is messier than the announcement sounds, too. Nobody has said who selects and pays these evaluators, or what they can publish without editorial control from the lab funding them. Employee-like access to a frontier training pipeline is an IP lawyer’s full employment act, and every lab will negotiate carve-outs that turn employee-like into guest with a lanyard.
Coordination has its own wall. Labs agreeing on capability speed limits is exactly what antitrust law exists to prevent, hence the waiver request, and voluntary commitments bind nobody who doesn’t sign. The China step is the least feasible of the three, and Trump spent the weekend downplaying the risks while stressing America’s lead. In that framing, a unilateral American slowdown starts to look like unilateral disarmament.
Markets noticed anyway. Nasdaq 100 futures dipped more than 1% on the announcements, which tells you investors heard “capex slowdown” even as the essay insisted training continues. If you hold Nvidia or the hyperscalers, “pace” landed very differently than it did in San Francisco.
What I’d actually watch
Strip the rhetoric away and exactly one commitment is testable right now: embedded evaluators with real access and publishable findings. If METR-style teams are sitting inside Anthropic and OpenAI within months, filing reports the labs can’t edit, that’s a genuine structural shift in how frontier AI gets scrutinized. If the first report takes a year and reads like a press release, we’ll have our answer.
And the deeper problem goes untouched. A badged evaluator can watch a training run. Can they meaningfully assess a model that’s helping design its own successor? Recursive self-improvement is the actual accelerant behind Amodei’s fear, and nothing in the three-step plan slows the loop itself. That’s the gap between pacing and safety nobody is closing yet.
So my conclusion is split. Take the warning seriously, because three rivals reaching the same frightening conclusion about self-improving systems is either the most important safety signal we’ve gotten or the most coordinated narrative in AI’s history. Then judge the commitments by badges on desks and published reports, not essays. I’ve read enough AI safety promises to know the difference usually shows up in the access logs, not the announcements.
