Dario Amodei wants to hit the brakes. In an open letter published yesterday, Anthropic’s CEO called for the AI industry to “pace the frontier” and slow down development. Sam Altman agreed. So did Elon Musk. Even Demis Hassabis at Google offered tentative support.

Meanwhile, Donald Trump and House Speaker Mike Johnson think they’re all being ridiculous. Their concern? China will eat our lunch if we pause.

Here’s what’s wild about this debate: both sides are probably wrong.

The Case for Pausing (and Why It Won’t Work)

Let’s start with Amodei’s argument. The head of Anthropic, a company that positions itself as the “safety-focused” AI lab, says we need to slow down because the risks are getting too serious to ignore. On the surface, this sounds reasonable. We’re building systems we don’t fully understand, racing toward capabilities that could be genuinely dangerous, and the competitive pressure means nobody wants to be the one who blinks first.

But here’s the problem: a voluntary pause only works if everyone pauses. And they won’t.

When Altman and Musk publicly support Amodei’s call, what they’re really doing is signaling virtue while changing nothing. OpenAI isn’t going to stop training GPT-5 or whatever they’re calling it this week. Musk isn’t shutting down xAI. They’re just saying the right things while their GPU clusters keep humming.

Even if the major US labs genuinely slowed down, which they won’t, it doesn’t matter if China doesn’t join in. Or the EU. Or the dozens of well-funded startups that would love nothing more than for OpenAI and Anthropic to take a nap while they catch up.

A pause without enforcement is just theater.

The Case for Racing Ahead (and Why That’s Insane)

Trump and Johnson’s position is somehow even less coherent. Their argument boils down to: we can’t slow down because China won’t slow down, and whoever gets to AGI first wins some undefined geopolitical prize.

This is the same logic that gave us the nuclear arms race, except worse, because at least with nukes we eventually figured out treaties and verification. With AI development, we don’t even have a shared definition of what we’re racing toward, let alone any mechanism to verify who’s doing what.

The “China will beat us” argument also conveniently ignores that Chinese AI labs are dealing with the same fundamental challenges everyone else faces. They don’t have magic solutions to alignment or interpretability. They’re running the same architectures, hitting the same walls, and dealing with the same uncertainty about what happens when these systems get more capable.

Racing to build something potentially dangerous faster than your adversary can build something potentially dangerous is not a strategy. It’s a coin flip with civilizational stakes.

What Nobody’s Talking About

Here’s what frustrates me about this whole debate: it’s premised on the idea that the speed of development is the main variable that matters. Faster or slower, pause or accelerate, those are the only options on the table.

But speed isn’t the problem. The problem is that we’re building systems we can’t fully control or understand, and neither slowing down nor speeding up changes that fundamental fact.

A five-year pause doesn’t give us five years of safety research. It gives us five years of the same research we’re doing now, just with less data from frontier systems. The hard problems in AI safety, things like mechanistic interpretability, scalable oversight, and robust alignment, don’t get solved faster just because you stop training large models. If anything, you need access to the most capable systems to figure out how to make them safe.

On the flip side, racing ahead without solving these problems doesn’t somehow make them go away. You just get to catastrophic capabilities faster, with even less understanding of how to handle them.

A Better Question

Instead of arguing about pace, we should be arguing about disclosure and access.

If Anthropic is genuinely worried about existential risk, they should be publishing more of their safety research, not less. Same for OpenAI, Google, and everyone else. The biggest labs have the most data about how these systems actually behave at the frontier. That knowledge shouldn’t be proprietary.

We should be demanding that frontier labs give serious external researchers real access to their models for safety testing. Not filtered API access, actual ability to probe these systems and understand what’s happening under the hood.

We should be building institutions that can actually evaluate AI systems before deployment, with teeth to enforce standards. Not a voluntary pause that nobody will follow, but actual regulatory infrastructure.

And yes, we should be talking to China and other major players about coordination. Not a naive “let’s all just stop” agreement, but serious technical dialogue about shared risks and potential cooperation on safety research.

The Real Risk

The AI pause debate is seductive because it makes the problem seem simple. Too fast? Slow down. Too slow? Speed up. Problem solved.

But the real risk isn’t that we’re moving too fast or too slow. It’s that we’re moving in the dark, and we’re all pretending we can see.

Amodei’s call for a pause at least acknowledges uncertainty, even if the proposed solution won’t work. Trump and Johnson’s position doesn’t even do that. It just assumes everything will work out fine as long as we get there first.

Neither approach grapples with the actual hard question: how do we build increasingly capable AI systems in a way that we can actually verify is safe, regardless of the pace of development?

Until someone has a good answer to that, arguing about speed is just rearranging deck chairs.