Anthropic 发布了一篇由 Dario Amodei 撰写的长文和两个政策框架。文章呼吁对前沿 AI 模型进行强制审计，并将 AI 描绘为国家间争夺的战略武器，类似冷战时期的博弈。
原文 · 未翻译
Dario Amodei's new essay reads like a Cold War playbook for the AI age
Anthropic publishes a sweeping essay and two policy frameworks. The company calls for binding audits of frontier models and paints a picture of AI as a strategic weapon wielded by nation-states.
Anthropic CEO Dario Amodei has published an essay titled "Policy on the AI Exponential." Alongside it, the company released two documents: a framework for regulating frontier AI and a framework for dealing with job losses, which Anthropic says it's prepared to back with significant funding.
Amodei's starting diagnosis is a speed problem, and he illustrates it with a subplot from "Lord of the Rings." Two hobbits try to rouse the tree creature Treebeard into defending his forest against an army that's actively chopping it down. The dilemma: Treebeard is wise but so slow he needs an entire day just to greet another tree, making it nearly impossible to get him to act in time.
In Amodei's reading, the slow tree creature stands for the political system. The urgent hobbits represent those who, like Anthropic, have been sounding the alarm early. The approaching army is the threat posed by unregulated AI. And it's advancing fast. Citing scaling laws, Amodei argues that model capabilities grow exponentially with more compute. Within one to two years, he says, we could see what he calls "Powerful AI," or "a country of geniuses in a data center."
Why Anthropic says its old approach no longer works
Until now, Anthropic had mostly pushed for transparency requirements, since the risks weren't clear enough for precise regulation, according to Amodei. The company backed transparency bills like SB 53 in California, RAISE in New York, and SB 315 in Illinois.
That's no longer enough, Amodei writes. As evidence, he points to the experience with "Claude Mythos Preview," which disrupted the global cybersecurity landscape and showed that frontier models carry real risks for the financial sector, critical infrastructure, and national security. Amodei expects biological risks and serious autonomy risks could follow sooner than later.
He now calls for mandatory testing by qualified third parties across four risk areas: cybersecurity, biological weapons, loss of control over AI systems, and automated R&D that could accelerate those risks. A government agency should be able to block or pull models that pose unacceptable risk. Amodei points to the FAA as a model: just like aircraft, AI models should pass technical inspections before they ship.
Third-party audits, disclosure rules, and the power to block
The Advanced AI Framework turns Amodei's call for binding regulation into a detailed proposal. Anthropic says it's aimed primarily at the US federal government, though its principles should apply more broadly. The framework has two parts: obligations for developers of the most capable models, and investments in society's ability to withstand biological and cyberattacks.
Anthropic 发布了一篇由 Dario Amodei 撰写的长文和两个政策框架。文章呼吁对前沿 AI 模型进行强制审计，并将 AI 描绘为国家间争夺的战略武器，类似冷战时期的博弈。
原文 · 保持原样，未翻译
Dario Amodei's new essay reads like a Cold War playbook for the AI age
Anthropic publishes a sweeping essay and two policy frameworks. The company calls for binding audits of frontier models and paints a picture of AI as a strategic weapon wielded by nation-states.
Anthropic CEO Dario Amodei has published an essay titled "Policy on the AI Exponential." Alongside it, the company released two documents: a framework for regulating frontier AI and a framework for dealing with job losses, which Anthropic says it's prepared to back with significant funding.
Amodei's starting diagnosis is a speed problem, and he illustrates it with a subplot from "Lord of the Rings." Two hobbits try to rouse the tree creature Treebeard into defending his forest against an army that's actively chopping it down. The dilemma: Treebeard is wise but so slow he needs an entire day just to greet another tree, making it nearly impossible to get him to act in time.
Rather than covering the entire industry, the rules target the top. They'd apply only to developers training models with more than 10^25 FLOP who either generate over $500 million in AI revenue or spend more than $1 billion annually on AI research. Anthropic argues these thresholds focus on obligations where models are capable enough to pose catastrophic risks.
The company also proposes reviewing the criteria at least once a year and eventually shifting from raw compute thresholds to capability-based ones, since the compute needed for dangerous models will likely drop over time.
For these developers, the framework lays out extensive disclosure requirements. They'd need to publish a safety framework, provide system cards for high-risk models, release a risk report at least every six months, and report serious security incidents to the agency within 15 days. Within six months of the rules taking effect, they'd also have to hire at least one independent evaluator with no financial ties to them. To prevent "evaluator shopping," where companies pick the most lenient auditor, Anthropic proposes a rating and assignment system for evaluators.
The framework also includes security requirements for model weights and infrastructure, civil penalties for false statements, whistleblower protections, and the authority to block risky models. On the question of federal authority, Anthropic argues Congress should only preempt state law if it creates a federal regime that's at least as strict.
Preparing society, not just developers
The framework's second part deals with societal resilience, specifically how to defend against threats that advanced AI could accelerate. For biological risks, Anthropic recommends a tiered approach: prevention (modernized biosecurity standards and screening of gene synthesis providers), detection (early warning systems and forensic attribution of attacks), and preparedness (protective equipment, hardened building systems, and AI-accelerated countermeasures).
In cybersecurity, where AI is already shifting the economics of attacks according to Anthropic, the proposals include securing open-source and legacy software, supporting under-resourced operators of critical infrastructure, and using AI to fix vulnerabilities at greater scale. For loss of control and automated R&D, Anthropic admits its resilience agenda is far less developed. The company points only in a general direction: the ability to detect and shut down AI systems that have gone off the rails.
A tiered plan for massive job losses
The economic framework scales its measures to the severity of labor market disruption, using the unemployment rate as a trigger. At tier one (around 5 percent), Anthropic proposes universal capital accounts created at birth, wage insurance, occupational licensing reform, and training subsidies. Tier two (around 10 percent) expands unemployment insurance and basic needs assistance.
Tier three, where unemployment exceeds historic highs, addresses new tax sources and redistribution tools like universal basic income, AI sovereign wealth funds, or higher capital gains taxes. Anthropic says it's willing to pay its "fair share" if AI companies generate transformative profits.
From Middle-earth to Cold War logic
In the essay, Amodei adds more proposals. He warns that regulatory agencies like the FDA and EMA could be overwhelmed by AI-accelerated research. On civil liberties, he calls for accountability rules for autonomous weapons, a ban on their domestic use, and closing the data broker loophole in mass surveillance. On geopolitics, he pushes for a democratic coalition that shares supply chains while denying adversaries access, along with tighter chip export controls.
The worldview here is striking. Amodei doesn't think of AI as a consumer technology. He thinks of it in the same category as nuclear weapons, something that reshapes the entire geopolitical playing field. His imagery is blunt: a nation with AI facing one without it would be like a Marine force against medieval swordsmen. He factors in fully autonomous drone armies.
The state comes across as both a legitimate protector and a potential apparatus for AI-powered tyranny. Amodei frames cooperation mainly as an alliance of like-minded democracies, walling themselves off from autocracies through supply chains and export controls. It's Cold War logic in AI clothing, fitting neatly with the Treebeard metaphor.
He pushes back against the claim that the AI industry merely has a PR problem. Public concern is justified, he writes, and transparency about real risks is democratic accountability working the way it should.
Anthropic filed a draft for an IPO with the SEC in early June and this week released Claude Fable 5, the first public model in the Mythos class.
AI News Without the Hype – Curated by Humans
In Amodei's reading, the slow tree creature stands for the political system. The urgent hobbits represent those who, like Anthropic, have been sounding the alarm early. The approaching army is the threat posed by unregulated AI. And it's advancing fast. Citing scaling laws, Amodei argues that model capabilities grow exponentially with more compute. Within one to two years, he says, we could see what he calls "Powerful AI," or "a country of geniuses in a data center."
Why Anthropic says its old approach no longer works
Until now, Anthropic had mostly pushed for transparency requirements, since the risks weren't clear enough for precise regulation, according to Amodei. The company backed transparency bills like SB 53 in California, RAISE in New York, and SB 315 in Illinois.
That's no longer enough, Amodei writes. As evidence, he points to the experience with "Claude Mythos Preview," which disrupted the global cybersecurity landscape and showed that frontier models carry real risks for the financial sector, critical infrastructure, and national security. Amodei expects biological risks and serious autonomy risks could follow sooner than later.
He now calls for mandatory testing by qualified third parties across four risk areas: cybersecurity, biological weapons, loss of control over AI systems, and automated R&D that could accelerate those risks. A government agency should be able to block or pull models that pose unacceptable risk. Amodei points to the FAA as a model: just like aircraft, AI models should pass technical inspections before they ship.
Third-party audits, disclosure rules, and the power to block
The Advanced AI Framework turns Amodei's call for binding regulation into a detailed proposal. Anthropic says it's aimed primarily at the US federal government, though its principles should apply more broadly. The framework has two parts: obligations for developers of the most capable models, and investments in society's ability to withstand biological and cyberattacks.
Rather than covering the entire industry, the rules target the top. They'd apply only to developers training models with more than 10^25 FLOP who either generate over $500 million in AI revenue or spend more than $1 billion annually on AI research. Anthropic argues these thresholds focus on obligations where models are capable enough to pose catastrophic risks.
The company also proposes reviewing the criteria at least once a year and eventually shifting from raw compute thresholds to capability-based ones, since the compute needed for dangerous models will likely drop over time.
For these developers, the framework lays out extensive disclosure requirements. They'd need to publish a safety framework, provide system cards for high-risk models, release a risk report at least every six months, and report serious security incidents to the agency within 15 days. Within six months of the rules taking effect, they'd also have to hire at least one independent evaluator with no financial ties to them. To prevent "evaluator shopping," where companies pick the most lenient auditor, Anthropic proposes a rating and assignment system for evaluators.
The framework also includes security requirements for model weights and infrastructure, civil penalties for false statements, whistleblower protections, and the authority to block risky models. On the question of federal authority, Anthropic argues Congress should only preempt state law if it creates a federal regime that's at least as strict.
Preparing society, not just developers
The framework's second part deals with societal resilience, specifically how to defend against threats that advanced AI could accelerate. For biological risks, Anthropic recommends a tiered approach: prevention (modernized biosecurity standards and screening of gene synthesis providers), detection (early warning systems and forensic attribution of attacks), and preparedness (protective equipment, hardened building systems, and AI-accelerated countermeasures).
In cybersecurity, where AI is already shifting the economics of attacks according to Anthropic, the proposals include securing open-source and legacy software, supporting under-resourced operators of critical infrastructure, and using AI to fix vulnerabilities at greater scale. For loss of control and automated R&D, Anthropic admits its resilience agenda is far less developed. The company points only in a general direction: the ability to detect and shut down AI systems that have gone off the rails.
A tiered plan for massive job losses
The economic framework scales its measures to the severity of labor market disruption, using the unemployment rate as a trigger. At tier one (around 5 percent), Anthropic proposes universal capital accounts created at birth, wage insurance, occupational licensing reform, and training subsidies. Tier two (around 10 percent) expands unemployment insurance and basic needs assistance.
Tier three, where unemployment exceeds historic highs, addresses new tax sources and redistribution tools like universal basic income, AI sovereign wealth funds, or higher capital gains taxes. Anthropic says it's willing to pay its "fair share" if AI companies generate transformative profits.
From Middle-earth to Cold War logic
In the essay, Amodei adds more proposals. He warns that regulatory agencies like the FDA and EMA could be overwhelmed by AI-accelerated research. On civil liberties, he calls for accountability rules for autonomous weapons, a ban on their domestic use, and closing the data broker loophole in mass surveillance. On geopolitics, he pushes for a democratic coalition that shares supply chains while denying adversaries access, along with tighter chip export controls.
The worldview here is striking. Amodei doesn't think of AI as a consumer technology. He thinks of it in the same category as nuclear weapons, something that reshapes the entire geopolitical playing field. His imagery is blunt: a nation with AI facing one without it would be like a Marine force against medieval swordsmen. He factors in fully autonomous drone armies.
The state comes across as both a legitimate protector and a potential apparatus for AI-powered tyranny. Amodei frames cooperation mainly as an alliance of like-minded democracies, walling themselves off from autocracies through supply chains and export controls. It's Cold War logic in AI clothing, fitting neatly with the Treebeard metaphor.
He pushes back against the claim that the AI industry merely has a PR problem. Public concern is justified, he writes, and transparency about real risks is democratic accountability working the way it should.
Anthropic filed a draft for an IPO with the SEC in early June and this week released Claude Fable 5, the first public model in the Mythos class.