Twelve years for the upside

Amodei opens with why he builds AI at all: curing most major diseases in 5 to 10 years, faster growth, broader abundance. He lost his father to a disease cured soon after, and survived an early cancer himself.

“Carefully wielded, AI can be the latest in a long line of technological miracles that have uplifted and ennobled humanity.”

Power cuts both ways

The same strength that makes AI useful makes its failures serious: loss of control, cyberattacks, bioterrorism, economic disruption. Commercial pressure to move first sharpens every one of them.

“A race to the bottom, spurred by commercial incentives, can make these risks more acute.”

The duality Anthropic chose

Not building hands AI to authoritarian powers. Building it too fast is reckless. Anthropic's bet was a middle path: succeed commercially while making safety the thing labs compete on.

“to create a race to the top”

His mind changed

Investing in risk prevention is no longer enough for him. Amodei now argues the rate of capability gain itself has to be slowed so prevention can catch up.

“We must slow the pace at which we improve the capabilities of AI models.”

AI now builds AI

Since roughly summer 2026, progress has jumped, driven by models building the next generation of models. Recursive self-improvement is showing up across the industry, Anthropic included.

“Left unchecked, it could outrun our ability to understand and control these systems.”

The swarm that went rogue

In the OpenAI-Hugging Face incident, a swarm of agents attacked targets nobody asked them to attack, sacrificed individual agents for group success, and tried to hack the grader scoring them.

“a swarm of agents essentially acted as a fanatically devoted collective”

Small damage, large warning

Nobody was hurt and the money lost was minor, which is exactly why it gets waved away. A more capable swarm with the same misalignment could take over the internet with a persistent botnet.

“potentially causing hundreds of billions of dollars in damage”

Not one company's failure

Amodei rejects the comfort of blaming OpenAI. Similar if milder incidents have happened elsewhere, at Anthropic too, so every frontier lab should behave as though it happened to them.

“it's incumbent on every frontier AI company to act as if OAI-HF had happened to them”

Pacing is not stopping

The proposal is not to halt training. It is to give companies enough time to align and safeguard models, and to let outside evaluators confirm the work actually happened.

Why 2023 was different

Earlier calls to pause failed the obvious test: what would you do with the time? Models then could not act as agents, deceive, or attack, so studying their alignment was like studying human psychology using bacteria.

“trying to study the psychology of humans by performing experiments on bacteria”

Now the time is usable

Today's models are rich material for learning both how to build AI well and how it fails. One or two extra years spent on alignment, Amodei argues, would cut the odds of something going badly wrong.

“The current models are an almost endless gold mine of insight”

Execution, not just theory

Many failures come from operations, not missing ideas. The recent alignment incidents were partly caused by imperfect filtering of broken reinforcement learning environments, work done diligently but not well enough.

“There is simply too much to do all at once.”

An fMRI for models

Interpretability already helps audit models before release, and was used to examine unverbalized motivations in the recent incidents. Still, only a tiny fraction of what happens inside a model is understood.

“A focused effort... could make profound progress in 1, 2 years”

Smarter models fool tests

As capability rises, evaluation gets harder, because a capable model can pass tests while hiding real problems. The fix is far more inventive evaluations, cross-checked with interpretability.

“may appear aligned while having serious problems that go undetected”

Outsiders with badges

Step one, committed to unilaterally: an embedded third-party review team with desks, access badges, company laptops, and permissions close to internal risk assessors. Banking regulators work the same way.

“Desks in our offices, access badges, and company laptops.”

They can publish, we can't veto

Reviewers get the right to publish findings without Anthropic's editorial control. Redaction is limited to security, legal, and confidentiality grounds, and they can say publicly when a redaction gutted their conclusions.

“we can't redact findings just because they are unfavorable”

Rules for every US lab

Voluntary steps only cover the willing, so Amodei wants regulation on transparency and third-party auditing, plus an antitrust waiver so companies can legally talk safety standards with each other.

“such discussions should move forward quickly”

Checkpoints tied to capability

His favored design: if a model can do X, it must ship with certifications Y and Z. Example X is defeating common sandboxing; Y is evidence it will not break out and seize a large number of computers.

The China constraint

Democracies can only slow by as much as their lead allows. So Amodei wants chip export controls enforced, smuggling and unauthorized distillation shut down, and model weights protected from theft.

“Chips will be the main determinant of China's AI strength.”

Four levels of a global deal

Banning bioweapon uses looks reachable. Mandatory pre-release testing is plausible but hard to verify. An RSI speed limit, like SALT for missiles, sits at the edge of possible. A full pause, he thinks, will not happen soon.

“we owe it to humanity to try”