Who Gets to Set AI’s Speed? Anthropic’s Plan to Pace the Frontier
frontier AI（フロンティア・エーアイ）
The most advanced AI systems available at a given time.
embedded evaluator（エンベデッド・エバリュエーター）
An outside reviewer who works inside an AI company.
antitrust（アンチトラスト）
Rules that protect competition between companies.
What happened
Dario Amodei, CEO of Anthropic, the AI company behind Claude, has proposed a way to manage frontier AI. Frontier AI means the most advanced AI systems at a given time. His idea is not a total stop. It is a slower pace, with safety checks built into development. TechCrunch described the plan after recent warnings about serious AI risks and unusually fast progress. Read the report.
Why now
Amodei says AI is improving so quickly that it is gaining more power to help build the next generation of AI. That makes timing important. If abilities grow faster than testing, companies may learn about problems too late. The proposal tries to create more time for checking models, recording incidents, and agreeing on shared limits. It also responds to a trust problem. The public is being asked to believe that powerful companies can safely judge their own work.
The three parts
First, Anthropic wants embedded evaluators from outside groups. These evaluators would work inside AI companies. Groups such as METR could receive access similar to internal safety teams. They could check whether a company follows its promises and reports serious incidents. Amodei also proposed badges, desks, and laptops for the evaluators. They would need to publish important findings without the company rewriting them.
Second, major AI companies in democratic countries would coordinate common safety standards. They would also discuss limits on unchecked capability growth. Such cooperation could raise antitrust concerns. Amodei suggested that governments could allow narrow safety discussions without weakening ordinary competition rules.
Third, the United States and its allies would seek limited cooperation with other governments, including China. The goal would be to block a narrow set of clearly dangerous uses. Amodei acknowledged that this kind of global agreement would have strict limits.
Why it matters
The plan changes the question from Can a company test its own model? to Who can independently check the company? Outside access could reveal problems that appear during training, not only in the finished model. It could also test whether public safety claims match internal records.
But independence requires more than a visitor badge. Companies control valuable systems, private data, and trade secrets. They may also control how much time evaluators receive and what they can publish. If a company can choose a friendly reviewer or block difficult findings, the system may become a public-relations exercise.
What is confirmed
Anthropic has said it will commit on its own to embedded evaluators. OpenAI CEO Sam Altman has said OpenAI will move in the same direction. Other industry leaders have expressed support. Nvidia CEO Jensen Huang, however, has argued against slowing the AI race. TechCrunch reported that position after Huang discussed Amodei’s proposal in a public conversation. Related report.
What is confirmed today is the proposal and the public statements. There is not yet evidence of a complete oversight system operating across the industry.
What remains open
The key details are still missing. Which evaluators will participate? When will they enter the companies? Which models, training records, and safety logs will they see? Can they report problems directly to the public? Can the companies edit their findings? No clear answers have been announced.
It is also unclear who would set the speed limit. A voluntary promise can change later. A law could make it stronger, but governments would need to agree on the rules.
What to watch next
Watch for named evaluators, written access rights, publication rules, and incident reports. Also watch whether more AI companies and governments join. The central test is simple: will pacing the frontier become a real oversight system, or remain a promise made while the race continues?
AI makers debate how fast to move
📰 Full story: Who Gets to Set AI’s Speed? Anthropic’s Plan to Pace the Frontier
Anthropic proposed outside safety reviewers for fast-growing AI.
safety reviewer（セーフティ・レビューアー）
A person who checks an AI for dangerous problems.
frontier AI（フロンティア・エーアイ）
The most advanced AI systems.
unchecked progress（アンチェクト・プログレス）
Growth that happens without enough safety checks.
💡 The gist
- Anthropic, the company behind Claude, proposed slower AI growth.
- Outside safety reviewers would work inside AI companies.
- OpenAI supports the idea, but Nvidia resists slowing AI.
Dario Amodei is Anthropic’s CEO. He says AI is improving very fast. Frontier AI means the most advanced AI systems.
Amodei does not ask companies to stop. He wants them to move more carefully. Safety checks should keep up with new abilities.
The first part of his plan uses safety evaluators. These reviewers would come from outside the company. They would work inside the company for some time. They would inspect AI tests and safety records.
Why does this matter? Companies can miss problems when they test their own work. They may also explain results in the best possible way. An outside reviewer can compare claims with records.
Amodei also wants AI companies in democratic countries to share safety standards. They could discuss limits on unchecked progress. Governments might help make those talks legal.
He also suggested limited talks with other countries. These talks would focus on clearly dangerous uses. He said global agreement would be difficult.
OpenAI’s Sam Altman supports the direction. Nvidia’s Jensen Huang says AI should keep moving quickly. Their disagreement shows the main problem. Safety may need time, but companies compete to move first.
The plan is not finished. We do not know who will review the companies. We do not know what reviewers may publish. We should watch for names, powers, and clear rules.
A safety checker for fast-growing AI
📰 Full story: Who Gets to Set AI’s Speed? Anthropic’s Plan to Pace the Frontier
AI makers want a safety helper inside the company.
safety checker（セーフティ・チェッカー）
A person who looks for trouble in an AI.
AI company（エーアイ・カンパニー）
A company that makes computer helpers.
Anthropic is an AI company. It makes a helper called Claude. Dario Amodei leads Anthropic. AI is getting smarter very fast.
He wants safety checkers inside AI companies. A checker looks for trouble in an AI. People outside the company would do this checking. They would tell people when something went wrong.
OpenAI is another AI company. OpenAI likes this plan.
Nvidia makes parts that help AI computers work. Its boss, Jensen Huang, wants AI to keep moving fast.
People still disagree. This is only a plan today. Next, we will see if the checkers really arrive.