Anthropic CEO Dario Amodei is drawing a line in the sand on open-weight AI models. Amodei disputes the increasingly popular claim that releasing model weights to the public inherently makes it easier to build safety guardrails. His position is essentially the opposite: once you release the weights of a frontier AI model, you lose the ability to revoke access, push patches, or even monitor how the thing is being used.
The core disagreement
Amodei isn’t saying all open models are dangerous. He’s drawn a clear distinction between smaller open models, which he’s acknowledged provide real benefits for research and competition, and the most advanced frontier systems. His concern is specifically about the latter category, where capabilities might enable serious harm and where, once released, there’s no “undo” button.
During Senate testimony back in 2023, he characterized the open release of advanced AI models as a “dangerous path.”
The China factor and US policy response
