September 13, 2026
Garry Tan's idea, as TechCrunch reported, is simple on its face: if Chinese open-weight labs got ahead by distilling frontier models into smaller, cheaper ones, American open-weight labs should do the exact same thing to American frontier models. It's a plan built entirely on the premise that open weights are a strategic asset worth defending, not just a developer convenience. The part worth sitting with is what this says about where the real competition is happening. It's not just OpenAI versus Anthropic versus Google anymore -- it's open versus closed, and increasingly American-open versus Chinese-open. For a business picking AI tools today, that's actually good news: more capable open-weight options, distilled down to something a normal company can run without a frontier-lab budget, means more real choice and less lock-in. We've written before about what that access fight means for teams deciding whether to build on somebody else's model or their own stack. The tradeoff nobody in Tan's framing addresses head-on: distillation is legally and ethically murkier than it sounds, and the same frontier labs whose weights get distilled may not sit still for it.
Twenty-five mathematicians signed an open letter, per TechCrunch, arguing that AI labs are treating their published proofs and problem-solving work as free training data while threatening the field's intellectual foundations. This isn't a new complaint in spirit -- writers, artists, and coders have all made versions of it -- but mathematicians carry a specific kind of authority: their work is often the clearest, most literal representation of reasoning itself, which is exactly what frontier labs are trying to teach models to do. What should worry a business reader isn't the letter itself, it's the pattern. Every constituency whose output trains these systems eventually organizes and pushes back, and the labs' response so far has mostly been silence or vague reassurance. If you're building products or workflows on top of these models, that unresolved tension is a real business risk, not just an academic dispute -- future licensing fights, court rulings, or PR blowups could reshape what these models are allowed to do or cost to use. It's the kind of thing worth watching the same way we've tracked the industry's other credibility fights.
The timing here is what makes this story different from the usual doomer cycle. A researcher resigned from Anthropic this week and posted on X that the company is racing toward self-improving superintelligence and gambling with lives, according to TechCrunch. What should stop you is that the company's own alignment lead co-signed the message instead of distancing from it. That's not an outsider critic -- that's someone still inside the tent, or just outside it, saying the safety culture isn't holding. I don't think this means superintelligence is imminent, and I'd hedge hard against anyone claiming certainty either way. But it does mean the people closest to the work are less confident in their own guardrails than their public messaging suggests, and that gap between internal alarm and external calm is exactly what a business evaluating AI vendors should price in. If the safety team's own leadership is willing to co-sign a resignation letter this stark, that's a signal about organizational pressure, not just individual conscience. It's worth reading alongside the broader conversation about whether AI leaders actually mean it when they say slow down.
Three stories, one thread: trust in who controls these models, who gets credited for their inputs, and who's honest about their risks is fraying on every front at once. If you're choosing AI tools for your business this quarter, which of these three fights actually changes your vendor decision -- the open-weight race, the mathematicians' claims, or the doomsday warning?
Sources