Three open letters in five days exposed the fault lines running through AI's safety debate

Three letters landed in the last week of July 2026. Each one claims the word "safety." They agree on almost nothing else, and Washington is being asked to pick among them.

On July 24, Microsoft published "Open Weights and American AI Leadership," a letter signed by 235 companies including NVIDIA, Amazon, and Y Combinator 1. Three days later, Anthropic published its own position paper, with CEO Dario Amodei warning that open-weight models enable misuse 1. Then on July 28, 1,324 employees of frontier AI companies published "Pacing the Frontier," asking the U.S. government to build international tools to deliberately slow AI development 2.

Same vocabulary. Three irreconcilable answers.

What each letter actually demands

The three positions answer the same question differently: who should control advanced AI, and how fast should it advance?

- Microsoft's coalition (235 companies): Keep open-weight models unrestricted. The letter argues that open weights expand access, strengthen competition, and make AI safer by letting independent researchers inspect models for vulnerabilities. It explicitly defends distillation, the practice of using one model's outputs to train another, framing it as a standard technique rooted in the open-source software tradition 3.

- Anthropic: Restrict open-weight model releases. Amodei warned that open models could be misused for cyberattacks or biological attacks and called for a crack down on industrial-scale distillation operations, while noting that Anthropic has not advocated for a blanket ban on open-weight models 1.

- Pacing the Frontier (1,324 employees): Slow the frontier itself. The statement asks the U.S. government to "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development," citing the risk that AI systems automating their own research could accelerate beyond human oversight 2.

The first two letters fight over who controls the models. The third asks whether they should be advancing this fast at all.

Who profits from each definition of safety

Follow the commercial logic, and the philosophical disagreement narrows into something more concrete. Microsoft's Azure cloud infrastructure earns more when companies build applications on open-weight models that run on Microsoft servers. Anthropic's revenue depends on customers paying for Claude through Anthropic's own closed API, where the model's weights never leave Anthropic's control. The open-weight fight is, underneath the rhetoric, a fight over whether AI value flows through a platform toll or directly to the model creator.

The distillation debate maps onto the same fault line. Distillation lets a company build a competitive model by training on the outputs of a frontier model it did not build. Microsoft's letter frames this as legitimate innovation 3. Anthropic frames large-scale distillation as a practice worth restricting 1. Neither side uses the word theft. The disagreement is about whether a practice that benefits one business model should be regulated to protect another.

The Pacing the Frontier signatories cut across this divide. The list includes chief scientists and executives from OpenAI, Anthropic, Google DeepMind, and Meta 2. Dario Amodei's name appears on both Anthropic's restriction argument and the slowdown letter. The CEO of a company racing to build frontier models is publicly asking for tools to slow that race. Ilya Sutskever, CEO of Safe Superintelligence Inc, wrote that managing the risks of what he called a coming intelligence explosion requires building pacing tools before they are needed 2. When the people closest to building these systems are asking for a brake, the safety arguments from their employers start to look like they are answering a different question than the one the public is being asked.

What Washington actually has to decide

All three letters address the same audience: the U.S. government. Microsoft wants regulators to leave open models alone. Anthropic wants regulators to restrict the practices that undermine closed models. Pacing the Frontier wants regulators to build coordination tools that neither open nor closed advocates have proposed.

Each faction has defined safety to match its commercial or institutional interest. Microsoft says openness is a path to security 3. Anthropic says safety requires control. More than a thousand industry employees say safety requires slowing down. The practical question for regulators is not whether AI should be safe but which of these three definitions becomes the default that vendors must comply with, and what enforcement mechanism matches it. That decision will determine who builds AI, who profits from it, and who controls it.