132 Companies Signed the Open Weights Letter. Anthropic Asked for Distillation Enforcement Instead

AI 요약

• What happened: The July 24 open letter Open Weights and American AI Leadership launched with 25 signatories. As of July 29 the count is 132.
  • Amazon, Google, Meta, Microsoft, NVIDIA, OpenAI, IBM, Hugging Face, Mistral, and the Linux Foundation are all on the list.
  • Among major US frontier labs, only Anthropic is absent.
• Anthropic's answer: In a July 27 position statement, Dario Amodei wrote that Anthropic "has never advocated for a ban on open-weight models."
  • His alternatives: chip export controls, enforcement against industrial-scale distillation, and mandatory pre-release safety testing for open and closed models alike.
• Where they actually split: Both documents oppose banning open weights. The disagreement is whether distillation is a legitimate technique or something to police.
• Developer impact: The US House already sent inquiries to Airbnb and Cursor in April 2026 about Chinese model usage, and widened the scope in July. Now is a good time to inventory which weights your workloads run on.

The first post Jensen Huang ever made on his X account was not a product announcement. It was a link to Open Weights and American AI Leadership, published July 24, which asks Washington not to rush regulation onto open-weight models. That post passed 11 million views.

The more interesting part of the letter is who is missing from the signature list. It opened with 25 companies, doubled to 50 within a day, and the letter page now carries 132 logos as of July 29. Amazon, Google, and OpenAI all joined late. Anthropic never did. On July 27, Anthropic published a separate statement laying out what it opposes and what it wants instead.

What 132 companies asked for

The letter makes four requests: expand compute access for startups and researchers, invest in shared training assets (datasets, tooling, evaluation frameworks), avoid premature regulation of open models, and use targeted legal frameworks instead of blanket restrictions on techniques like distillation.

Its argument leans on 1980s open-source software as precedent, and it directly rejects the premise that a closed-only ecosystem is the safer one.

Relying exclusively on closed models is not inherently safe. Closed models can also be compromised or misused, and can fail in ways that are not externally detectable.

The letter does concede the risk. Once weights are released they leave the original developer's control, and modified versions are hard to track or recall. It just argues that a ban is the wrong response to that risk.

This fight surfaced now because there is a live policy question in Washington.

What turned the signature list into pressure was David Sacks, the White House's outside AI adviser. Sacks wrote on X that leading closed labs "want the government to eliminate their open-source competitors," then added that "they've put their cards on the table." He named no company, but by then the set of frontier labs that had not signed was small.

The three things Amodei asked for instead of a ban

Anthropic's position statement opens with one sentence: Anthropic has never advocated for a ban on open-weight models. Amodei adds that open-weight models without dangerous capabilities are a public good.

His argument against a ban is that it would not work.

Banning US companies from using these models does nothing to address this risk, because malicious actors are unlikely to be legitimate US companies.

The measures Anthropic says it does support are three:

1. Chip export controls. Powerful chips and chip-making equipment should not be sold to China, paired with enforcement against smuggling. Anthropic's case is that this is the most direct lever on the national-security risk.
2. Enforcement against industrial-scale distillation. The reasoning is that distillation lets China build better models without matching compute. Anthropic says it already enforces this through "our own practices, including identifying and banning accounts."
3. Mandatory safety testing. Every sufficiently powerful model, open or closed, should be tested for cyber, biological, and alignment risk before release.

The third is the heaviest for anyone shipping open weights. Today a team can run its own evaluation and publish. If that requirement became rule, publishing acquires a gate it has to clear first.

Distillation is the real fault line

Put the two documents side by side and neither one wants open weights banned. The head-on collision is over distillation.

Here is how the letter frames it:

Policymakers should be careful not to conflate legitimate model development techniques with improper acquisition. Distillation, the practice of using one model's outputs to train or improve another, is a widely used technique for model improvement, evaluation, and validation.

One side calls distillation a standard technique for evaluation and validation. The other calls it extraction that routes around compute. That definitional gap lands directly on engineering teams. A team tuning a small model on frontier-model outputs is either following normal development practice or operating a terms-of-service violation, depending on which definition wins.

Anthropic is already past the proposal stage here. CNBC reported on July 6, 2026 that Alibaba banned employees from using Anthropic AI after Anthropic named the company as the source of what it called the largest known distillation attack. Account bans are not a policy suggestion; they are terms enforcement running today.

What to check now

No rule has changed yet. US companies are still not barred from using Chinese open models, and only some federal departments have imposed their own restrictions. The investigations, however, are live. The US House sent inquiries to Airbnb, Cursor, and others in April 2026 about Chinese AI model usage, and in July widened the scope to ties with Moonshot AI, DeepSeek, MiniMax, Alibaba, Zhipu AI, ByteDance, Tencent, and Baidu.

Three things worth doing:

• Inventory model provenance. Write down which workload runs on whose weights. With Chinese open models reaching 46% of token usage at US enterprises, the odds that some sit in your stack are not low.
• Read the distillation clauses. If any pipeline trains or tunes your own model on frontier API outputs, check that provider's terms and its account-ban record.
• Keep local copies of weights. For models published openly, such as Moonshot's Kimi K3, a downloaded copy buffers you against later policy shifts. Regulation can close a distribution channel; it cannot easily recall files already on disk.

That last item is also the clearest read on how enforceable any of this is. Multiple AI policy researchers have called a full open-weight ban ultimately unworkable, and legal scholars have raised First Amendment objections to restricting published weights. Kimi K3's weights went out at midnight UTC on July 27 and topped Hacker News that day. Long before any rule is finalized, the files are already copied onto disks worldwide.