Level 1 — Absolute Beginner

Dario Amodei is the boss of an AI company. The company is called Anthropic.

He wrote a long letter. In the letter he says AI is getting better too fast.

He does not want to stop AI. He wants it to go more slowly. Then people can learn how to keep it safe.

Other AI bosses read the letter. Sam Altman and Elon Musk said yes, we agree. After that, some computer chip companies lost money.

AI

Short for artificial intelligence. Computer systems that can learn and answer.

company

A business that makes or sells things.

boss

The person who leads a company or a group of workers.

letter

A piece of writing sent or shown to other people.

fast

Moving or happening in a short time.

slow

Not fast; taking more time.

safe

Not dangerous; free from harm.

agree

To have the same opinion as someone else.

Level 2 — Elementary

On Saturday 12 September, Dario Amodei published an essay of about 3,800 words on his personal website. Amodei runs Anthropic, one of the small number of laboratories building the most powerful artificial intelligence systems in the world. The title of the essay was We Must Pace the Frontier.

His argument is not that AI should be stopped. It is that the speed of improvement should be reduced, so that safety research and rules for governments can catch up with what the machines can already do. He warns about systems that improve themselves and about AI agents that act on their own without a person checking each step.

Amodei asks for three things. He wants outside experts to test frontier AI companies regularly. He wants the laboratories and democratic governments to work together more closely. And he wants countries to cooperate internationally on the risks.

The response was unusual. Sam Altman of OpenAI agreed that the industry needs to slow down and do more on safety. Elon Musk, who owns xAI, said the same within a day. On 14 September, shares in chip companies fell, because slower AI development could mean fewer orders for the expensive processors these systems need.

essay

A piece of writing that presents an argument on one subject.

laboratory

A place where research is done. Here it means a research company.

pace

To control the speed of something.

frontier

The furthest edge of what is currently known or possible.

governance

The system of rules by which something is controlled.

agent

A program that can take actions on its own to reach a goal.

cooperate

To work together toward the same purpose.

processor

The chip inside a computer that does the calculations.

Level 3 — Intermediate

It is rare for the person building something to argue publicly that it is being built too quickly. On Saturday 12 September, Dario Amodei, chief executive of Anthropic, posted a 3,800-word essay titled We Must Pace the Frontier on his personal site. The central claim is that the rate at which frontier model capabilities are improving now outruns the rate at which anyone, including the laboratories themselves, is learning to control them.

Amodei is careful about what he is not proposing. This is not a call to halt development or to freeze research. It is a call to slow the gradient of capability gains long enough for safety work and governance structures to close the distance. The specific hazards he names are recursive self-improvement, in which a system contributes to designing its successor, and increasingly autonomous agents that execute long chains of actions without a human reviewing each one.

Three concrete measures follow. The first is continuous third-party evaluation of frontier companies rather than occasional voluntary disclosure. The second is tighter coordination among laboratories and democratic governments. The third is broader international cooperation, on the reasoning that a risk which crosses borders cannot be managed inside one jurisdiction.

What gave the essay weight was the response. Sam Altman of OpenAI publicly accepted the premise within hours, and Elon Musk of xAI followed inside a day, producing a moment of agreement between three rivals who disagree about almost everything else. Markets read it as a forecast rather than a sentiment: on 14 September, semiconductor shares fell, since a slower capability race implies a smaller appetite for the accelerators that train these models. By 18 September the pacing debate had moved from technical circles into mainstream business coverage.

chief executive

The most senior manager responsible for running a company.

capability

The range of things a system is able to do.

halt

To bring something to a complete stop.

recursive

Repeating a process by applying it to its own result.

autonomous

Able to act independently without direct control.

disclosure

The act of making information known to others.

jurisdiction

The area over which a particular authority has legal power.

accelerator

Specialised hardware built to speed up demanding computation.

Level 4 — Advanced

Arguments for restraint usually come from outside an industry, which is what made Dario Amodei's intervention notable. On 12 September the Anthropic chief executive published a roughly 3,800-word essay, We Must Pace the Frontier, on his personal site, contending that the rate of capability improvement at the frontier has decoupled from the rate at which anyone is acquiring the means to govern it. The argument is not that progress is undesirable. It is that a derivative, the speed of gain, has become the relevant variable, and that it is currently set by competitive dynamics rather than by any deliberate judgement.

The hazards Amodei identifies are specific rather than atmospheric. Recursive self-improvement, in which a model materially contributes to the design of its successor, compresses the interval available for external review with each iteration. Increasingly autonomous agents, executing extended action sequences without per-step human adjudication, shift the locus of error from a single wrong output to a chain of consequential decisions. Neither concern requires assuming malice or consciousness; both follow from ordinary engineering pressure applied at pace.

His prescriptions are institutional rather than technical. Continuous third-party evaluation would replace episodic voluntary disclosure, on the principle that self-reported safety is structurally weak where commercial incentives run the other way. Closer coordination between laboratories and democratic governments addresses the collective action problem directly: no single firm can unilaterally slow down without ceding position, which is precisely why the constraint must be shared to be credible. Broader international cooperation acknowledges that capability diffuses across borders faster than regulation does.

The reception was the more instructive development. Sam Altman endorsed the premise almost immediately, and Elon Musk concurred within a day, producing a rare alignment among three principals whose competitive and personal disputes are extensively documented. Whether that convergence survives contact with quarterly incentives is an open question, and the market supplied an early answer of sorts: semiconductor equities declined on 14 September, the straightforward inference being that a moderated training race reduces demand for accelerators. By 18 September the pacing question had migrated from research commentary into mainstream financial press, which is usually the point at which a technical debate acquires a constituency that can enforce an outcome.

intervention

An act of becoming involved in a situation to influence it.

decouple

To separate two things that previously moved together.

derivative

In mathematics, the rate at which a quantity changes.

adjudication

The act of judging or deciding on a matter.

locus

The place where something is situated or occurs.