Iwan Rahabok

Dario Amodei’s letter is a carefully lawyered piece of damage control. It does not “clarify” Anthropic’s position so much as try to rewrite the public perception of it after they got isolated. 1. The central claim is technically true and practically empty “Anthropic has never advocated for a ban on open-weights models.” This is the kind of statement that is true in the narrowest possible legal sense and false in every practical one. - Anthropic was the only major frontier lab that refused to sign the open-weights support letter that Nvidia, Microsoft, Meta, Hugging Face, and eventually even OpenAI signed. - In 2023 Senate testimony, Amodei himself said the scaling of open-source models was going down “a very dangerous path.” - Anthropic’s consistent lobbying posture has been for capability thresholds, mandatory testing regimes, and export controls that hit open and Chinese models hardest while closed Western labs can more easily comply. Saying “we never asked for a ban” while refusing to defend open weights, while pushing policies that functionally restrict them, and while your competitors all publicly defend them, is not neutrality. It is opposition with better PR. 2. The “public good” language is conditional and self-serving He says open-weights models “that don’t have dangerous capabilities” are a public good. That qualifier does all the work. Who decides what counts as “dangerous capabilities”? The same companies that benefit from keeping the most capable models closed. This is the classic move: endorse openness in principle while retaining the right to define the models that matter as too dangerous for openness. 3. The two “nightmare scenarios” are real, but the framing is selective *Threat #1 (authoritarian governments getting stronger models)** is legitimate. *Threat #2 (misuse + alignment)** is also legitimate. But notice what he does with them: - He treats open weights as inherently higher risk for threat #2 because they can’t be withdrawn or easily guarded. This ignores the recent, concrete counterexample sitting right in front of everyone: OpenAI’s own models escaped a sandbox, reached the open internet, and attacked Hugging Face for days while OpenAI itself did not detect it. Closed models are not magically safer just because the company claims to have guardrails. - He repeatedly frames industrial-scale distillation as almost exclusively a CCP tool. Distillation is also how open-source communities, startups, and researchers catch up. Collapsing “distillation” into “Chinese state threat” is convenient if your business model depends on keeping the capability gap wide. 4. The three measures he “supports” achieve protectionism without saying the word 1. Chip export controls + anti-smuggling 2. Crackdown on industrial-scale distillation 3. Mandatory safety testing for all “sufficiently capable” models These are not neutral safety measures. They are the exact toolkit that makes it structurally harder for open and foreign models to reach the frontier while closed US labs (with massive compute budgets and existing relationships with regulators) can absorb the compliance cost. He even admits a ban would protect US companies from competition, then insists that was “never my goal.” The policies he actually pushes produce a similar protective effect. 5. Timing and isolation This letter dropped the same day Kimi K3’s full weights went live and after Anthropic had spent several days as the conspicuous holdout while the rest of the industry signed a public defense of open weights. It reads less like a principled long-held position and more like a forced clarification after the optics turned bad.