In “We Must Pace the Frontier”, Dario Amodei, CEO of Anthropic, makes an argument that at first sounds almost paradoxical coming from the head of one of the world’s leading AI companies: we should slow down.
More precisely, he argues that we should slow the rate at which the capabilities of frontier AI models increase.
That distinction matters.
As Amodei puts it:
“Pacing does not mean halting model training or technical progress.”
His proposal is not a call to stop developing artificial intelligence. It is a call to put some distance between two things that have increasingly been treated as synonymous: progress in AI and progress at the frontier of general-purpose model capability.
They are not the same thing.
We have been treating one axis as the whole map
For the past several years, much of the AI industry has been organized around a remarkably simple objective: build increasingly capable general-purpose models.
More compute. Larger training runs. Better reasoning. More autonomy. Longer context. Better coding. More capable agents.
This has produced extraordinary results.
But it has also created a tendency to think of AI development as a single-dimensional race. A model is ahead if it can do more things, with greater autonomy, across more domains.
The frontier becomes synonymous with progress itself.
Amodei’s proposal implicitly challenges that assumption.
He argues that the growth in capability may soon be moving faster than our capacity to understand, evaluate and control the systems we are building. His answer is not to abandon AI, but to create enough friction at the frontier for safety research, interpretability, evaluation and operational discipline to catch up.
I think there is another possible consequence that deserves attention. What follows is my extrapolation from his argument.
If the frontier slows, AI development does not have to slow with it.
It may simply spread sideways.
Innovation can move horizontally
There is an enormous space between “train the next most powerful model in the world” and “stop working on AI.”
We can build smaller models.
We can build highly specialized models.
We can build systems designed to run locally.
We can improve inference efficiency.
We can develop better memory architectures, interfaces, tools, orchestration systems and agent workflows.
We can train models for medicine, engineering, logistics, education, law, manufacturing or scientific research without requiring every one of them to become a universally capable artificial intellect.
We can also spend far more effort learning how to use the extraordinary models we already have.
That last point is easy to underestimate.
The capabilities currently available are advancing faster than most organizations — and most individuals — can absorb them. There is already an enormous implementation gap between what existing AI systems can theoretically do and how effectively they are actually deployed.
A slower frontier could therefore coexist with an accelerating application layer.
Intelligence does not necessarily scale by becoming more general
Biology offers an interesting analogy.
The human brain is not composed of billions of identical general-purpose processors waiting for increasingly powerful central intelligence to tell them what to do.
It relies heavily on specialization.
Different neural structures process vision, language, movement, memory, spatial relationships and countless other functions. Intelligence emerges partly from the interaction of specialized systems.
The analogy does not determine how artificial intelligence must evolve, but it suggests another possible direction.
Perhaps the future is not dominated by a succession of increasingly enormous monolithic models.
Perhaps powerful general models will increasingly help us create large populations of smaller, narrower and highly adapted models, each optimized for a particular environment or task.
In that world, frontier models may become less like the final product and more like infrastructure: systems used to design, train, coordinate and improve other systems.
The most powerful AI might help us create less powerful AI — in enormous quantities.
A speed limit can change the direction of traffic
Amodei raises checkpoints as one possible approach to pacing frontier development: as systems acquire more consequential capabilities, developers would need stronger evidence of alignment, interpretability and safety before moving further. He also raises the possibility of limits on inputs such as compute or the use of AI to accelerate AI research itself.
The obvious interpretation is that these mechanisms would slow technological progress.
But that assumes technological progress has only one direction.
Constraints often redirect innovation rather than suppress it.
If increasing raw model capability becomes more expensive, slower or more regulated, engineering effort could migrate toward other dimensions: efficiency, specialization, robustness, usability and deployment.
The result might not be less AI.
It might be more kinds of AI.
The real question is what we mean by progress
There is a recurring tendency in technology to confuse the dominant trajectory with the inevitable one.
Faster processors once seemed to define progress in computing. Eventually, energy efficiency, parallelism, mobile computing and specialized accelerators became just as important.
Telecommunications was once largely about increasing the capacity of centralized networks. The Internet changed the architecture of the problem.
Software itself moved repeatedly between centralized and distributed models.
AI may be approaching a similar bifurcation.
One path continues the current race toward increasingly capable general systems, with every generation attempting to push the frontier as far and as quickly as possible.
Another path still advances the frontier — but more deliberately — while an increasingly diverse ecosystem expands beneath it.
The second path is not technological stagnation.
It may actually produce a richer technological landscape.
Amodei ends his essay by arguing that AI can still deliver extraordinary benefits even if we proceed more deliberately at the frontier.
I think the implication goes further.
Slowing the frontier does not necessarily mean slowing AI.
It may simply force us to remember that progress has more than one direction.