Zvi Mowshowitz's post about Anthropic opens with a frustration that will be familiar to anyone who has tried to evaluate the company: "One can both be the best given the alternatives, and insufficient to meet the moment."
That's the right frame, and it's rarer than it should be. Most commentary picks one half. Either Anthropic is the responsible lab, so criticism is ingratitude, or it's a hype machine with a safety brand, so its good acts are theater. Zvi's position is that both are true at different levels of analysis, and that "what you think about Anthropic depends on what parts you care about most, and what you think would be sufficient."
The occasion is a leaked draft bill — a variation on the Frontier Act, stronger in some ways, weaker in others, and not clearly better than passing nothing given how much state law it would have preempted. Dario Amodei responded at length to Nathan Calvin's criticism, and Zvi's summary of that response is the most useful part of the post.
The list
Amodei's defense is a list of costs Anthropic has paid: SB 53, aggressive bio filtering, delaying a model, the "Policy on the AI Exponential" essay, a clash with the Department of War, support for export controls (which, Zvi notes, OpenAI stopped supporting), not selling to PRC-based companies, extensive reports and model cards, and a 5% stock contribution to a public benefit fund.
Zvi grants that these "came at substantial cost." That's the honest move, and it's why the piece is worth reading: he isn't arguing the list is fake. He's arguing about what it proves.
The identity question
Calvin's core question, as Zvi frames it, is: what even is Anthropic?
That's not a policy question. It's a question about whether an institution has a stable character that predicts its behavior when the costs rise. A company can be genuinely safety-motivated and still be organized around a business that makes certain safety decisions impossible, and the two facts don't cancel.
The leaked bill matters here because legislation is where a lab's preferences become everyone's constraints. A draft that preempts state law while being weaker than the status quo is not a neutral act; it's a choice about which level of government gets to decide. Zvi's objection isn't that Anthropic wrote a bill — it's that the bill's content revealed a preference for preemption over strength.
The closing argument
The last paragraph is the one to remember:
"I think that the admirable sacrifices Anthropic has made in a few areas it really believes in (bio, China) do not reassure me it'll do the extremely hard work to get things right in a few other areas that it seems to have a lot more internal or leadership disagreement about (alignment, takeover, disempowerment)."
That's a precise claim about revealed preference. Willingness to pay costs in domains where leadership agrees tells you about conviction. It tells you much less about capacity to act in domains where the leadership disagrees with itself — which are precisely the domains where the hard work is political as well as technical.
My take
Three things generalize beyond this company.
Comparative and absolute evaluations are different questions. "Best available" is a ranking. "Sufficient" is a threshold. A vendor can win the ranking and miss the threshold, and the marketing will always emphasize the ranking. Ask which question you're answering.
Look for the disagreement domains. Every organization has areas where it moves decisively and areas where it stalls. The stalls are usually where the real risk lives, because that's where the internal coalition is split. A public list of accomplishments is a map of the areas that aren't split.
The artifacts settle it. A leaked bill and an S-1 filing are more informative than a blog post or a tweet thread, because they're the places where intentions become constraints. Zvi says he's reading the S-1 next, which is the right instinct: the document that has to satisfy the SEC is harder to fudge than the essay that has to satisfy the discourse.
The part I'd push back on is the fatalism of the ending — "I don't expect things to go well." It may be correct, but it's not testable, and a critic who wants the company to succeed (Zvi clearly does) is more useful when the prediction is specific enough to be wrong.
The bill and the S-1 are specific enough. That's where I'd look.