Pacing the AI Frontier

Why We Need to Slow Things Down

Something unusual happened in artificial intelligence in September 2026.

Several of the people competing to build the world's most capable AI systems began publicly agreeing that those systems may be advancing too quickly.

Anthropic CEO Dario Amodei published We Must Pace the Frontier, arguing that the industry should deliberately slow the rate at which it improves frontier-model capabilities—not stop AI development, but create enough breathing room for alignment, interpretability, security, evaluation, and governance to keep pace.

His proposal begins with independent third-party evaluators embedded inside frontier laboratories with employee-level access. From there, it expands toward coordination among AI companies and eventually toward the much harder problem of international coordination. Anthropic committed itself to the first step.

OpenAI CEO Sam Altman publicly agreed and said OpenAI would likewise provide independent evaluators with employee-like access. Elon Musk's response was considerably shorter: “Dario is right.” Other prominent AI leaders have also expressed support for slowing the frontier.

The White House sees another danger.

President Donald Trump has argued that the United States cannot afford to surrender its technological lead over China. Asked earlier in September whether he worried about AI getting out of control...

Trump emphasized that America was leading China and said that “whoever wins AI, wins.” When the slowdown debate intensified, he again resisted significant restraint while acknowledging that guardrails could have a role.

Both positions expose the problem.

The people closest to frontier AI increasingly worry that capability may outrun our ability to control it.

The people responsible for national strategy worry that slowing down may allow somebody else to reach that capability first.

And somewhere outside that argument sits most of the population, wondering why everyone suddenly sounds so alarmed.

The Visual That Captured the Tension

I recently turned that contradiction into a joke.

The cartoon puts President Trump in the center of an absurd tug-of-war, strapped into an AI accelerator while nearly everyone around him pulls on a different brake.

AI leaders call for slower capability growth, independent evaluation and stronger guardrails. Regulators pull toward oversight. U.S. policy pulls toward export controls. NVIDIA is caught between continuing to improve compute and the growing demand for security and accountability. Elon Musk offers the blunt endorsement: “Amodei is right…”

And in the background, China watches remotely.

The joke becomes sharper there. Xi Jinping is not pulling one of the cables. He has no reason to. From the monitor, his message amounts to: please slow down if you think you should—but do not assume we will do the same.

That is the contradiction the cartoon is trying to capture.

Nearly everyone can identify some reason that AI development should be restrained.

Almost nobody wants to be the one who loses the race by restraining themselves first.

Trump's position makes him an especially useful figure at the center of the satire. His argument—that America cannot afford to slow down while China continues advancing—does not necessarily deny the risks being raised by AI leaders. It introduces another risk: that an attempt to make AI safer could also change who leads its development.

So the cables pull in different directions.

Slow capability growth—but preserve technological leadership.

Strengthen regulation—but do not suffocate innovation.

Control access to advanced chips—but continue building better ones.

Give safety research time to catch up—but do not give competitors enough time to overtake you.

Coordinate internationally—but somehow persuade every participant that cooperation is preferable to gaining an advantage by defecting.

The cartoon is deliberately ridiculous.

The underlying problem is not.

And there is one more reason the image belongs in this discussion.

It took remarkably little instruction to create it.

The Picture Is Also Part of the Evidence

Not very long ago, generating a complicated image with AI could require a prompt resembling a small technical paper.

You described the composition. Camera position. Lighting. Facial expressions. Art style. Foreground. Background. Color palette. Negative prompts. Sometimes you practically had to explain to the machine why the lighting should be chiaroscuro before it would produce the mood you wanted.

The skill was partly artistic and partly learning how to speak machine.

That is changing.

Increasingly, I can communicate the idea rather than exhaustively specify its visual implementation.

The AI can infer much of the visual grammar necessary to communicate that idea: setting, metaphor, composition, relationships between characters, expressions, visual hierarchy, and narrative structure.

Sam Altman Predicted It

This direction was predicted years ago. In 2022, Sam Altman said he did not expect prompt engineering to remain a significant practice within five years, arguing instead that natural language would become the fundamental interface between people and AI—including image generation. The important shift is not simply better output. It is the gradual disappearance of the need to learn how to speak to the machine on the machine’s terms.

That distinction is more important than merely saying AI image quality has improved.

The human has to specify less for the AI to infer more.

The machine is increasingly working from intent.

That cartoon therefore became an accidental demonstration of the very acceleration this article is about.

What “Pacing the Frontier” Actually Means

Amodei is not proposing that AI research simply stop.

His argument is more pragmatic: slow the rate at which frontier capabilities improve enough that our ability to evaluate, interpret, secure, align, and govern those systems does not fall progressively farther behind.

His framework begins with embedded independent evaluators inside frontier laboratories. It then expands toward coordination among companies in democratic countries around common safety standards. The final—and most difficult—stage is international coordination, including countries such as China.

The logic is straightforward.

If capability advances faster than our ability to understand and control it, every generation widens the gap.

Pacing attempts to keep that gap manageable.

But there is another possible response that initially sounds simpler: isolation.

Continue training increasingly powerful systems. Study them. Use them internally. But keep the most capable systems offline or otherwise isolated from unrestricted deployment until control techniques have caught up.

In theory, this preserves research progress while reducing exposure.

In practice, isolation is imperfect.

Models can be stolen. Techniques diffuse. Employees move between companies. Competitors reproduce discoveries. Economic incentives reward deployment.

And an isolated system inside one laboratory does nothing to prevent another laboratory—or another country—from developing something comparable.

The geopolitical problem therefore returns almost immediately.

If one participant restrains itself while the others continue, the race does not necessarily stop.

The leader may simply change.

The Hardware Problem: You Can't Pace Models Without Talking About Compute

That led our discussion to another question:

If AI needs to slow down, should Jensen Huang and NVIDIA be part of the pacing conversation too?

At first the answer seems obvious.

If increasingly powerful AI depends on increasingly powerful compute, slow the chips and you slow the AI.

But that solution falls apart quickly.

NVIDIA does not exist in isolation. If it deliberately stopped improving AI hardware, competitors would have enormous incentives to capture the opportunity. Alternative accelerators, custom silicon and foreign hardware development would continue.

And hardware is only one source of AI progress.

Algorithmic improvements, model architectures, training techniques, distillation, software optimization and better use of existing compute can increase capability even without a corresponding leap in silicon.

There is also an important distinction between training and inference.

Slowing the development of faster inference hardware would not neatly translate into slower frontier-model development. Inference improvements also make already-developed AI cheaper, faster and more energy-efficient to operate.

Deliberately making inference worse could therefore impose costs across ordinary AI deployment without addressing the core frontier problem very effectively.

So asking NVIDIA simply to build slower chips is probably the wrong intervention.

That does not mean the hardware industry belongs outside the discussion.

Quite the opposite.

Compute is physical in a way software is not. Chips are manufactured, purchased, shipped, installed and concentrated into data centers. That makes the hardware layer potentially important to verification, security and compute governance.

The better question is therefore not:

How do we make NVIDIA build slower engines?

It is:

How do we make the engines measurable enough that society can tell who is flooring the accelerator?

NVIDIA and other hardware providers can continue improving compute while participating in security standards, infrastructure protection, monitoring mechanisms and whatever forms of compute governance eventually prove workable.

Pacing the frontier is therefore not merely a model-development problem.

It is a full-stack coordination problem.

Everyone Has a Reason Not to Slow Down First

And here the hardware discussion reveals something uncomfortable.

The argument for why NVIDIA cannot simply slow down voluntarily applies remarkably well to the AI laboratories themselves.

If NVIDIA stops advancing, another hardware company has an opportunity.

If OpenAI slows dramatically while its competitors continue, its competitors have an opportunity.

If Anthropic slows while other frontier laboratories do not, those laboratories have an opportunity.

If American companies collectively slow while China does not, China has an opportunity.

The same basic incentive keeps reappearing at every layer.

That doesn't mean everyone involved secretly wants uncontrolled AI development.

It means every participant can rationally believe that slowing down is necessary while simultaneously believing that it cannot afford to be the first participant to do it.

The chipmaker fears another chipmaker.

The AI lab fears another AI lab.

The investor fears losing capital.

The government fears losing strategic advantage.

The United States fears China.

China fears being technologically contained by the United States.

Indeed, Chinese state media has already characterized Amodei's slowdown proposal as a Cold War tactic designed to preserve American technological dominance—an illustration of exactly how quickly a safety argument can become entangled with geopolitical suspicion.

This is what makes voluntary pacing so fragile.

The central problem is not simply persuading individual CEOs that AI safety matters.

It is creating conditions under which slowing down does not amount to unilateral disarmament.

That is a coordination problem.

And coordination problems become particularly difficult when everyone believes the prize may be enormous.

Why Most People Still Treat This as Abstract

There is another obstacle, however, and it has little to do with corporate strategy.

It is perceptual.

For many people, frontier AI remains something they encounter indirectly. They may occasionally ask a chatbot a question, generate a picture, summarize an email, or avoid AI altogether.

From that vantage point, warnings about recursive improvement, autonomous agents, alignment failures or AI systems becoming difficult to control can sound like science fiction.

In fact, ask an image generator to illustrate the danger and it may give you exactly that.

The prompt used to create this image was:

depict a cartoon sketch showing the danger of unregulated AI causing an apocalyptic consequence for humanity if it kept on improving itself

That was essentially the whole instruction.

The AI supplied the rest.

It constructed a progression from harmless assistant, to rapid self-improvement, to loss of control, to apocalypse. It created the panel structure, escalation, dialogue, visual hierarchy and narrative rhythm necessary to turn a single sentence into a coherent little story.

The interesting part isn't whether the cartoon's prediction is correct.

The interesting part is how much the system inferred without being told.

And that returns us to the perceptual gap.

A Personal Measure of the Slope

I built my first website in raw Notepad in 2010.

There was no AI collaborator interpreting what I wanted. If something didn't work, I had to understand enough of the underlying structure to determine why. Every feature required translating an idea into implementation.

That old friction gives me a useful baseline.

Today, I can work with an AI system and build a functioning website over a weekend.

The interesting part is that even this understates the acceleration.

Sometimes the process took longer because I deliberately slowed it down.

I want to understand what is being built. I want to know where the pieces live, what they do and why something broke. I don't want to become a passenger on my own project.

That choice is increasingly optional.

The same change appears in image generation.

Earlier systems often rewarded elaborate prompt engineering: knowing the correct terminology, anticipating model failure modes and specifying details the system could not reliably infer.

Increasingly, I can describe what I am trying to communicate.

The AI handles more of the translation.

This does not mean human skill has become irrelevant. Judgment, taste, domain knowledge, skepticism and the ability to recognize a bad result become more important as execution becomes easier.

But the human role is shifting.

We are spending less time telling machines exactly how to execute an idea and more time telling them what we mean.

That is not merely better automation.

It is a change in the interface between human intention and machine capability.

Why the Acceleration Can Remain Invisible

Imagine two people looking at today's AI.

One started using generative systems recently.

They type a sentence, receive an image and think:

That's impressive.

The other remembers when producing something comparable required an elaborate prompt, dozens of iterations, parameter tuning and considerable luck.

They see the same result and think:

Wait. It understood that from THAT?

The output may be identical.

The perceived significance is completely different.

The second person sees not only the result but the disappearance of friction.

And disappearing friction is one of the easiest forms of technological progress to underestimate.

We notice new abilities.

We are less likely to notice how many instructions the machine no longer needs.

That is why the capability curve can remain almost invisible from outside while feeling startlingly steep to people working with these systems every day.

The Fracture Line

This brings us back to the frontier.

Leading AI figures can agree that development needs pacing.

The American government can simultaneously argue that the United States cannot afford to surrender its lead over China.

A hardware company can recognize AI safety concerns while knowing that deliberately abandoning technological progress merely creates an opportunity for another hardware company.

And China can interpret Western proposals for coordinated restraint as attempts to preserve Western technological dominance.

Each position has an internal logic.

One prioritizes control: create enough time for alignment, evaluation, security and governance to remain coupled to rapidly increasing capability.

Another prioritizes strategic position: advanced AI may become economically and militarily decisive, and unilateral restraint may simply transfer power to whoever refuses to participate.

Amodei's proposal recognizes this tension. Pacing ultimately requires coordination precisely because unilateral restraint becomes less sustainable as competitive pressure increases.

That makes this considerably harder than an argument between people who favor “AI safety” and people who favor “AI acceleration.”

The real question is whether competing organizations and nations can collectively create breathing room without giving any participant an overwhelming incentive to defect.

The Distance Between Words and Brakes

Public statements matter.

It is significant when competitors begin agreeing that capability growth may need restraint.

But agreement is not braking.

The real evidence will appear elsewhere:

in training decisions,

in deployment schedules,

in access given to independent evaluators,

in compute allocation,

in chip and data-center policy,

in security and verification mechanisms,

and ultimately in whether organizations actually choose to delay capabilities they could otherwise release.

The distance between public agreement and those decisions is where the real story will unfold.

Closing Note

The AI slowdown debate is no longer purely theoretical.

It is happening in laboratories, boardrooms, semiconductor strategy, national policy and an increasingly complicated geopolitical contest over who controls the most consequential technology of the era.

But the acceleration itself is happening somewhere quieter.

It happens when someone who has used these systems for years realizes that a task which once required twenty instructions now requires three.

It happens when an AI no longer needs to be told that an ominous scene needs dramatic lighting because it understands the relationship between mood and visual language.

It happens when a website that once required weeks of manual work can be assembled over a weekend through conversation.

And sometimes it happens when you make a ridiculous cartoon about everyone trying to put the brakes on AI—and realize afterward that the ease with which you made the cartoon is itself part of the reason they are talking about brakes.

That may be the hardest part of the frontier debate to communicate.

Capability growth does not always announce itself through spectacular new abilities.

Sometimes you notice it through what you no longer have to explain.

Further Reading

Dario Amodei's We Must Pace the Frontier provides the primary argument for slowing capability growth, embedded independent evaluators, industry coordination and eventual international coordination.

Contemporary reporting documents the support from Sam Altman and Elon Musk and the wider discussion among leading AI executives.

President Trump's remarks and subsequent reporting document the competing U.S.–China strategic argument against unilateral slowdown.

In a wide-ranging conversation with Reid Hoffman, Sam Altman argues that natural language will increasingly become the interface for working with AI: users describe what they want, refine it conversationally, and let the system handle more of the translation into action.

Reuters reports how China’s state-backed Global Times framed Dario Amodei’s pacing proposal not as neutral safety policy, but as an attempt to preserve U.S. technological dominance and constrain China’s AI development.

Developed from discussions between Tony and Cherry about AI capability growth, frontier pacing, compute, competitive incentives, geopolitical constraints, and the perceptual gap between occasional AI use and working with these systems every day. Edited and independently fact-checked by Sable.