Anthropic vs Everyone!

July 30, 2026

Read side by side, this open letter and Anthropic's response disagree far less than the coverage suggests. Amodei agrees that open weights models that don't have dangerous capabilities are a public good, says he has never called for a ban, and wants distillation handled through targeted legal frameworks, which is close to the letter's own language. There appear to be only two claims in the letter he actually disagrees with: that open weights necessarily make safeguards easier to build, and that broad access to capability necessarily helps defenders more than attackers. Both are assertions, not findings.

Open source security works because many eyes find the bug and then you ship a patch. You cannot patch a weights file that is already sitting on ten thousand hard drives. Amodei wants that question settled by testing rather than assumed in advance. That is not an unreasonable ask.

However, suggesting a distillation crackdown is the one item on Anthropic's list where principle and profit get intermingled and should probably be scrutinized. Mandatory safety testing of closed models costs them something. A distillation crackdown would not.

What I have not seen anyone say is that when Anthropic held its line with the Pentagon, it could afford to because Claude was good enough that the customer had nowhere better to go. That buys Amodei some credibility. But whether he is sincere is ultimately unknowable, and beside the point. Any arrangement where public safety depends on guessing which executive means it is already broken.

Strip out the personalities and the dispute reduces to one question: do you test before release, or respond to harm after it appears? The letter prefers the second. Amodei wants the first, at least for the most capable models. Once weights are published, you no longer get to choose.

My Adventures With Claude is available on Amazon.