Open letters about AI development

Over the past few weeks, three open letters have roiled the AI world. Though seemingly independent, they collectively trace a deep fault line over the technology’s future: open vs. closed, safety vs. innovation, and blanket regulation vs. layered governance. This is more than a technical spat—it’s about commercial stakes, geopolitics, and the fate of the developer ecosystem.

Microsoft’s coalition: open weights are safer, distillation is legitimate. On July 24, Microsoft shepherded “Open Weights and American AI Leadership,” signed by 235 companies ranging from NVIDIA and Amazon to Y Combinator and the Linux Foundation—OpenAI later added its name. The letter was a direct response to murmurs that the U.S. government might restrict open-weight models on safety grounds, a fear fueled by the earlier suspension of Claude Fable 5. Its core argument is sharp: closed models are not inherently safer—they can be breached and misused with no outside oversight—while open models allow communities to inspect, patch, and improve them. Even more striking, the letter openly endorsed distillation. Often derided as “freeloading” or “theft,” distillation uses outputs from one model to train another. Microsoft rebranded it as a time-honored technique and warned policymakers against conflating it with misappropriation. Beneath this lies raw ecosystem calculus: Microsoft profits from Azure and open models like Phi, so restricting open weights would hit its bottom line. OpenAI’s co-signature signals a possible softening of its IP stance, perhaps teasing future distillation collaborations.

Anthropic’s riposte: open weights are dangerous, distillation is theft. Three days later, Anthropic—absent from Microsoft’s letter—published “Our position on open-weights models.” CEO Dario Amodei doubled down on risks: authoritarian regimes could use open models to outpace the U.S., and malicious actors could mount cyber or bio attacks. He called for a crackdown on “industrial-scale distillation” to prevent rapid catch-up. Yet Anthropic didn’t reject open weights entirely; it conceded the U.S. should lead in open models, but with guardrails. This reveals a “managed openness” stance—leadership is fine, so long as you can control the rails. Anthropic’s business depends on closed Claude models, so the safety narrative becomes a moat against commoditization.

Developers speak up: don’t make us scapegoats. On August 2, a letter from SAFEGEN (including independent developer Simon Willison) entered the fray: “Securing America’s Software Supply Chain.” Instead of picking sides, it highlighted a blind spot—the AI supply chain involves API providers, model deployers, application builders, and end users. Blanket regulation that dumps all responsibility on deployers and builders would strangle small innovators. The letter cites the AI Index Report 2026: training costs tripled, but inference costs dropped 1400-fold, meaning anyone can deploy a model cheaply. Forcing every deployer to shoulder heavy safety audits would kill innovation. They urged “smart regulation” that allocates responsibility appropriately.

Three hidden signals you might have missed. First, OpenAI’s pivot on distillation is telling. Last year it slammed DeepSeek for suspect distillation; now it signs a letter lauding distillation as legitimate. This hints at a nuanced IP strategy—distinguishing malevolent mass distillation from benign research use—or a prelude to licensing certain models.

Second, Anthropic’s pro-safety stance still leaves room for open weights; it wants to define the terms of “safe openness” and who gets to distill. The power to set those rules is what’s being fought over.

Finally, the developer letter exposes a macro trend: as inference costs plummet, model deployment will spread like electricity. Mismatched regulation would be like holding every faucet user responsible for water source purity—only giant utilities could survive. For AI builders, this trio of letters is a policy weather vane. Will distillation become a legal minefield? Will open models be shackled? Will your product be tagged “high-risk” and buried in compliance costs? Joining industry voices, as Simon Willison did, might tilt the scales. The regulatory pendulum is swinging—your engagement could nudge it toward a future where innovation and safety can coexist.