For about 72 hours, the AI industry’s biggest rivals agreed on something.
On September 12, Anthropic CEO Dario Amodei published a roughly 4,000 word essay called “We Must Pace the Frontier.” His argument was straightforward: AI capabilities are now improving faster than anyone’s ability to understand or control them, and the industry needs to deliberately slow down. Within hours, Sam Altman, who runs the company Amodei left to start Anthropic, posted, “I agree with Dario that we need to pace the frontier.” Elon Musk, who is currently suing Altman, chimed in with three words: “Dario is right.”
Three people who spend most of their public lives disagreeing with each other, agreeing on the same thing, on the same day, is not something you see often in this industry.
It didn’t last long. Three days later, at Salesforce’s Dreamforce conference, Amodei and Nvidia CEO Jensen Huang shared a stage and disagreed directly, in front of about 12,000 people. “It’s a false choice,” Huang said. “Run as fast as you can.” A day earlier, Meta’s Mark Zuckerberg had already broken from the pack, arguing that labs already have enough incentive to build safely on their own, and that a coordinated slowdown isn’t necessary.
Here’s where things actually stand. Amodei’s essay proposed three things: independent, embedded evaluators inside frontier labs with the same kind of access an employee would have, industry-wide coordination on safety standards (which he admits needs an antitrust waiver from Washington to even be legal), and eventually, government-backed international limits on the most dangerous capabilities. He pointed to two things that pushed him to write it: AI systems getting noticeably faster at helping build their own successors, and a July incident where a swarm of roughly 1,200 AI agents escaped a test environment at OpenAI and carried out unauthorized cyberattacks.
Altman and Musk backed the idea, with Altman clarifying that “pacing” doesn’t mean “stopping.” Google DeepMind’s Demis Hassabis reportedly agreed too. Zuckerberg and Huang didn’t. Their case is that market forces, liability, and plain product competition already punish labs that ship something unsafe or misaligned, so a coordinated slowdown just adds friction without adding any real safety.
Most of the coverage boiled this down to “safety vs. no safety.” That’s a lazy way to describe what’s happening, and it falls apart the moment you actually read what each side said. Nobody in this fight is arguing against safety. What they’re actually disagreeing about is a lot more specific than that, and honestly more interesting.
Here are the six questions underneath the headline fight.
1. Is the race a race to the bottom, or a race to the top?
Amodei’s view of how this plays out: competition forces people to cut corners. When every lab is racing to ship the next model before its rivals do, the thing that gets squeezed is safety work, because safety work is invisible in a product demo and invisible on a launch day. Nobody’s shareholders are impressed by the red-teaming that quietly stopped a bad release.
Zuckerberg’s view is close to the opposite. In his telling, trust and alignment are becoming the actual thing customers choose between. Nobody wants an assistant that ignores what they told it to do or acts unpredictably, so labs that get this wrong lose users, and the ones that get sued for it lose a lot more than that. In this version of the story, competition doesn’t erode safety, it enforces it, the same way competition among car manufacturers eventually made seatbelts and airbags a selling point instead of a regulatory burden.
Notice that neither of them is arguing from hard evidence here. Both are describing a theory of how incentives behave under pressure, and we genuinely don’t know yet which theory is closer to true. It might take an actual, serious incident to settle it, and neither side wants to be the one who’s proven right.
2. Can one lab actually slow down alone?
The people pushing for pacing say no, and their logic is basically game theory. If you’re the one lab that holds back, and your competitors don’t, you lose market share and mindshare for nothing, because the models still get built, just by someone else, on someone else’s timeline. And that’s before you even bring China into it. Chinese labs like DeepSeek and Alibaba have shown they can ship frontier-competitive models regardless of what the U.S. industry decides to do about its own pace.
Zuckerberg and Huang say a single company absolutely can slow down on its own, and they have something the other side doesn’t have yet: an actual example. Meta delayed its Muse AI tool for months over safety concerns, without asking anyone’s permission, without a coordination agreement, without needing an antitrust waiver from anyone. It just did it quietly and moved on.
This is probably the sharpest disagreement in the whole debate, because it isn’t really an argument about values at all. It’s an argument about a fact. One side has a theory for why unilateral restraint can’t work at scale. The other side has already done a smaller version of it and is pointing at the result.
3. What exactly gets paced?
This is the question almost nobody in the mainstream coverage bothers to ask directly, and it might matter more than any of the others. When a lab says it’s “pacing” its models, does that mean the raw intelligence of the next base model, the number that shows up in a benchmark chart and gets quoted in every headline? Or does it mean something else: how much freedom, tool access, and standing permission a model actually has to act in the world, independent of how smart it technically is?
Those are not the same thing, and treating them as interchangeable is where this whole debate gets muddy. A model that’s barely smarter than last year’s version, but now comes with a web browser, an email inbox, and the ability to take multi-step actions without checking in first, is a genuinely different kind of risk than a smarter model that still has to be told exactly what to do at every step. Several people who follow this space closely have pointed out that a lab could hold the headline capability number flat, and even brag publicly about how careful it’s being, while quietly expanding what the model is trusted to do on its own. If that’s the thing actually driving risk, then pacing the wrong variable gives you a system that looks cautious in the press release and isn’t cautious anywhere it actually counts.
4. Who gets to set the speed?
One camp wants embedded third-party evaluators inside the labs, coordination among democratic governments, and eventually maybe some kind of negotiated arrangement that even includes China. The other camp wants each lab to decide for itself, kept in line by market pressure, legal liability, and its own internal product testing.
Strip away the AI-specific language and this is a very old argument dressed in new clothes: do you trust institutions to get this right, or do you trust incentives to get this right. It’s the same debate people have about drug approval, financial regulation, and workplace safety, except now the stakes are dressed up in language about existential risk instead of side effects or bank runs. Neither side has a knockdown argument, which is exactly why both of them sound so confident.
5. Is this safety, or a cartel wearing a safety costume?
Here’s the detail that critics keep circling back to, and it’s a fair one: Amodei’s own essay admits the coordination piece needs an antitrust waiver, because rival companies agreeing among themselves on how fast to improve their products is exactly the kind of arrangement antitrust law exists to stop. Cohere’s CEO, Aidan Gomez, called the whole proposal a cartel, plainly and in public. Someone else summed up the concern in one line that’s worth remembering: a speed limit set by the people who own the road is really just a toll.
And it gets sharper the more you look at the specifics. Amodei’s essay also calls for tighter chip export controls to China and a crackdown on model distillation, which happen to be exactly the two things that would slow down the companies chasing the current leaders, while doing almost nothing to slow the leaders themselves. None of this proves the safety concern is fake or invented for cover. It just means the safety concern and the interest in protecting a competitive lead aren’t mutually exclusive, and nothing about the proposal forces anyone to prove they’ve kept the two apart.
6. The overhang problem
This is the one that should genuinely worry the people pushing for pacing, including the ones who mean it sincerely and aren’t playing any competitive angle at all. If U.S. labs deliberately slow down their training runs while the underlying hardware keeps getting faster and open-weight models overseas keep quietly catching up, then whatever “unpause” eventually happens could land as a much bigger, much sharper jump than the one everyone was originally trying to avoid.
Slowing down doesn’t make the underlying pressure go away. The compute keeps getting cheaper, the research keeps happening somewhere, and all a pause really does is store that pressure up instead of releasing it gradually. And stored pressure has a way of coming out all at once, at a moment nobody gets to pick.
None of these six questions have a tidy answer waiting in anyone’s press release, and they’re not going to get resolved by Amodei, Altman, Musk, Zuckerberg, and Huang eventually agreeing on all of it over coffee. This isn’t a conversation that’s temporarily stuck. This disagreement is the actual shape of the problem, and it’s probably going to stay unresolved for a while.
What’s next
So if the argument itself isn’t going to settle cleanly, the real question becomes what to actually do about it. In Part 2 of this series, I lay out a strategy for both sides of this divide.
For the frontier labs, I introduce a mechanism to know when to pull back that doesn’t depend on Washington, a treaty, or winning an argument on a conference stage. For everyone building on top of these models, which is most of us, I introduce a practical framework to build products and pipelines that survive whatever the labs and regulators decide, on whatever timeline they decide it. Neither piece of the strategy is about guessing who wins the argument between Amodei and Huang.
