Anthropic, OpenAI and Elon Musk suddenly agree on slowing AI development. Here's what "pace the frontier" means and why it happened now.
What does “pace the frontier” actually mean? #
“Pace the frontier” is the phrase Anthropic CEO Dario Amodei used in a roughly 9,000-word essay arguing that AI labs should slow the rate at which they push capability forward, without stopping entirely. Amodei’s concern centers on recursive self-improvement: AI systems now helping build and refine the next generation of AI systems faster than researchers can track or understand what’s changing. The essay laid out steps Anthropic said it would take regardless of what competitors did, and within hours it had public backing from Sam Altman at OpenAI and, more surprisingly, Elon Musk.
TL;DR #
- Dario Amodei’s essay “We Must Pace the Frontier” argues that AI development is outrunning the industry’s ability to understand what the models it builds are actually doing.
- The essay centers on recursive self-improvement , the idea that AI is now helping build the next generation of itself faster than predicted, and that this feedback loop needs deliberate slowing rather than a halt.
- The public alignment happened fast: Sam Altman agreed within the hour , saying pacing had already been an internal OpenAI conversation for weeks, and Elon Musk posted his agreement shortly after.
- This unusual unity followed a researcher resignation from Anthropic pre-training staffer Jacob Kaczon, whose four-sentence resignation post was read by roughly 170 million people in days.
- A July letter signed by 1,386 employees across OpenAI, Anthropic, Google DeepMind and Meta, including chief scientists, had already stated that the world may need the option to slow AI down.
- Not everyone welcomed the moment: a prominent figure from China’s open-source AI scene publicly called the pacing talk a pretext, arguing Western labs should simply say they don’t want competition.
- The same week the safety alliance formed, Chinese labs kept shipping competitive open-weight models, narrowing the gap with US frontier systems despite export controls on advanced chips.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
How did this start? #
The immediate trigger was a resignation post from Jacob Kaczon, a researcher who had spent three years doing pre-training work at both OpenAI and Anthropic. His post was short, four sentences, stating plainly that neither company was acting responsibly. It spread fast, reportedly reaching around 170 million reads within days.
Four days later, Amodei published his essay. He didn’t walk back Anthropic’s mission (he has spent over a decade arguing AI could help cure diseases, including the one that killed his own father) but he sharpened the warning considerably. The core claim: AI systems are already accelerating AI research itself, and that acceleration could move faster than anyone’s ability to monitor or steer it. He proposed pacing as a middle path between full-speed development and a moratorium, and said Anthropic would begin taking the first step immediately, independent of what any other lab decided.
What made the moment notable wasn’t the essay alone. It was the reaction. Sam Altman replied within the hour saying he agreed completely and that pacing had already been discussed internally at OpenAI for weeks. Then Elon Musk, who has spent years publicly feuding with both Altman and Amodei, posted three words under the same essay: “Dario is right.”
Why is this alignment unusual? #
Because it has essentially never happened before. Go back to 2023 and these same three men couldn’t agree to sign the same safety letter. Musk signed one open letter in March of that year. OpenAI declined to sign it. Altman and Amodei signed a separate safety statement in May, without Musk attached. Two competing letters in the same year, and never once did all three names appear together.
The 2025 moment reverses that pattern almost entirely. Within hours, competitors who normally fight over the same researchers, the same chips and the same headlines landed on the same sentence in public.
It’s also worth noting the pacing conversation didn’t start with Kaczon’s resignation, even though media coverage framed it that way. Back in July, roughly 1,386 employees across the frontier labs, including chief scientists at OpenAI, Anthropic, Google DeepMind and Meta, had already signed a statement arguing the world may need the option to slow AI development down, and that no single company can create that option alone. Kaczon’s resignation and Amodei’s essay didn’t invent the idea. They amplified a conversation that insiders were already having quietly across competing labs.
Is the pacing consensus really about safety? #
That’s the open question, and it’s the one that generates the most pushback. Not everyone reading Amodei’s essay interpreted it as a good-faith safety measure. A prominent figure associated with China’s open-source AI scene, known for building a widely downloaded open model family, responded publicly and sharply: paraphrased, the argument was that Western labs accelerate when it benefits them and ask everyone else to slow down when it doesn’t, and that nobody needs their permission to keep building.
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
The timing gives that critique some weight. In the same window that Anthropic, OpenAI and Musk were converging on a safety message, Chinese labs including DeepSeek and Alibaba’s Qwen team, along with groups like GLM and MiniMax, kept releasing models on a near-continuous cycle. Export controls on advanced chips were designed to slow this down at the hardware level. They don’t appear to have stopped the release cadence, and the performance gap between these open-weight Chinese models and the closed American frontier has been narrowing on a range of benchmarks month over month.
So the uncomfortable read is this: a coordinated call to pace development lands very differently depending on whether you’re the lab currently ahead or the lab currently catching up. Pacing is a far easier thing to propose once your lead starts to look less secure than it used to.
Why would researchers walk away from labs building this technology? #
Setting aside the competitive-strategy reading, there’s a simpler explanation for why someone like Kaczon would resign, and it doesn’t require assuming bad faith from anyone. People doing pre-training research are closer to the raw capability curve than almost anyone else on earth. They watch model behavior shift week to week in ways the public never sees until a product ships months later.
That proximity cuts both ways. It gives researchers the clearest possible view of what’s actually happening inside these systems, and it also means they’re the first to feel unsettled by what they see, whether or not that feeling translates into an accurate read on the actual risk. A pre-training researcher isn’t a policy analyst or an outside activist with an agenda. They’re someone who watched the numbers move faster than their own mental model of the technology could keep up with.
That doesn’t resolve whether pacing is primarily a safety response or a competitive one. It probably isn’t a clean either/or. Both dynamics can be true in the same essay, from the same person, in the same week.
Frequently Asked Questions #
What did Dario Amodei actually propose in his essay?
Amodei outlined steps meant to slow the rate of AI capability gains without halting development outright, framed around the risk of recursive self-improvement, where AI systems accelerate the building of future AI systems faster than researchers can track or understand. He said Anthropic would begin the first step regardless of what other labs chose to do.
Who is Jacob Kaczon and why did his resignation matter?
Jacob Kaczon was a pre-training researcher who had worked at both OpenAI and Anthropic. His short resignation post, stating that neither company was acting responsibly, spread to an estimated 170 million reads within days and is widely seen as the spark that preceded Amodei’s essay and the wider public conversation.
Did Sam Altman and Elon Musk actually agree with Anthropic?
Yes, according to their public replies. Altman said OpenAI would match the pacing commitment and that similar discussions had already been happening internally for weeks. Musk, despite years of public conflict with both Altman and Amodei, posted his agreement shortly afterward.
Are Chinese AI labs slowing down too?
There’s no indication of that in the available reporting. Labs including DeepSeek, Qwen, GLM and MiniMax continued releasing models at a rapid pace through the same period, and export controls on advanced chips do not appear to have stopped that cadence or the narrowing gap with Western frontier models.
Is the pacing movement primarily about safety or competition?
Both explanations have support. The July letter signed by over a thousand employees across major labs points to genuine, widely shared internal concern about capability racing ahead of understanding. At the same time, the timing, arriving just as the competitive gap with Chinese labs was closing, has led critics to argue that pacing is easier to call for once a lead feels less secure.