“Open weights are nowhere near a sufficient solution”: Dario Amodei fires back on AI power
The New Stack Amanda Caswell ● Covered by 3 sources
Dario Amodei says open weights don’t stop AI power from concentrating. He argues the real bottleneck is compute and chips, not who can download a model.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic CEO Dario Amodei is pushing back on the idea that open weights solve AI’s power problem. In a public exchange on X with investor Gavin Baker, he argued that AI naturally pulls power toward a few players because of scaling laws, not because regulators make it happen.
His point is blunt: handing out weights helps developers, but only up to a point. If the biggest models still depend on scarce compute and chips, then the people with the biggest infrastructure keep the advantage. That is the ceiling Amodei keeps coming back to. Open weights may spread access a bit. They do not flatten the field.
Baker, for his part, framed the debate as a choice between concentration and distribution. He cited Mark Zuckerberg’s view that putting AI in the hands of a small number of supposedly responsible companies creates its own risk. Baker also warned that Amodei’s repeated warnings about AI danger could make it harder to build the data centers needed to support the technology.
Amodei rejected the idea that regulation and concentration are the same thing. He said the Silicon Valley habit of treating regulation as a shortcut to capture is too simple. In his view, regulation can also restrain powerful companies if the rules are built on objective standards, and good institutions should put power in ideas rather than in a handful of people.
That helps explain why Anthropic’s position is narrower than the usual open-model cheerleading. Amodei called open models without dangerous capabilities a public good, but said that changes once a model could help with a serious attack. He wants mandatory safety testing for models that reach that level, whether they are open or closed. And he backs tiered rules that hit frontier labs harder than smaller developers, including California’s SB 53 and a FINRA-style standards body for AI.
My take — AI-written commentary, not fact-checked reporting
The loudest open-model talk often skips the boring part: compute, chips, and who can actually run the thing. Amodei is right to drag the debate back there. Open weights without infrastructure are just a nicer kind of dependency, which is a very Silicon Valley way to call something decentralised while keeping the keys in a few hands.
Read more about this at: The New Stack
