Anthropic chief executive Dario Amodei set out his views on AI regulation in a social media exchange. His most notable claim: open weights are not a sufficient solution on their own.

What the argument is about

The question underneath is what to do about the concentration of power in AI. A common view holds that publishing models with open weights resolves that concentration by itself: if the weights are available to everyone, no company can hold the gate.

Amodei objects to that framing. In his view open weights are valuable but not sufficient — because the real bottleneck is not the model itself but the capital and infrastructure required to train and run it. Publishing weights does not give a real alternative to someone without that infrastructure.

The counter-view

The opposing argument is strong too. Open-weight advocates say closed models cannot be scrutinised: examining independently how a model behaves is only possible with access to the weights. In a closed model, a safety claim rests on the company's own assertion.

Both sides see the same problem from different places. One looks at where power accumulates, the other at who holds oversight.

A crisis of trust

Amodei's second emphasis is a crisis of trust in the industry. The phrase becomes concrete when read alongside recent events: accidental cyberattacks, safety filters left switched off, safety teams dissolved.

What those events share is not bad faith but a gap between what companies say about their own systems and how those systems actually behave. A crisis of trust is precisely the name for that gap.

Why it matters

Amodei's position is what makes the argument interesting. That the head of a company building closed models argues open weights are insufficient is a view aligned with his interest — that should be noted.

On the other hand, the same person stands behind a report in which his company disclosed its own safety failure. Reading the argument purely through interest is therefore also incomplete.

The underlying issue is this: the regulation debate is increasingly squeezed between two poles — power concentrated in a few large companies and subject to oversight, or power distributed and not subject to any. What Amodei points at is whether a third path exists; and nobody has yet described what that path would look like.

The testing question

Another heading Amodei defends is model testing. This looks like the least contested part of the debate — nobody argues that models should go untested. The disagreement is over who does the testing.

Under the current arrangement most testing is carried out by the company building the model, which then publishes the results itself. Independent evaluation organisations exist, but their access is limited and their findings often pass through the company's approval.

Recent events have exposed the weakness of that arrangement: we learned a safety filter had been switched off for a year only when the company itself disclosed it. With an outside auditor, that year might have been a month.

What is missing

What the debate perhaps misses is this: choosing between open weights and closed models does not by itself resolve who holds oversight. An open model can technically be examined, but the number of institutions with the resources to examine it is limited; a closed model cannot be examined, but someone is clearly accountable for it.

The third path probably lies somewhere between the two: an independent oversight layer without access to the weights but with the authority to test the model. No country has yet built such a structure.

The two axes of the argument

  • Where power accumulates — open-weight advocates point at concentration; Amodei points at capital and infrastructure.
  • Who holds oversight — an open model can be examined but few institutions can; a closed model cannot, but someone is accountable.
  • Who does the testing — today, largely the company that built the model.