Key Takeaways

  • Moonshot AI is accused of “large-scale covert industrial distillation” of Anthropic's Fable model for its Kimmy K3, prompting intense debate over AI intellectual property.
  • US policy is a tightrope walk: it encourages open innovation but strongly condemns industrial-scale theft of proprietary American AI technology.
  • The controversy highlights a “pot calling the kettle black” argument: frontier AI labs train on vast public datasets, creating tension when they then object to others distilling their own models.
  • The practice is compared to established legal reverse-engineering in physical products, like Ford disassembling Chinese EVs to understand their construction.
  • Consumers and small businesses could benefit from cheaper, potentially distilled models, creating a complex competitive environment for leading AI labs and their investors.

The Disagreement

The AI world is wrestling with a thorny question: when does "learning from" become "stealing from"? The spark? Allegations that Moonshot AI developed a "sophisticated internal platform to conduct large-scale distillation against US models," specifically targeting Anthropic's Fable for its Kimmy K3 model. John Coogan minced no words, stating that "large-scale covert industrial distillation aimed at stealing proprietary US technology and undermining American research is unacceptable." This side of the debate sees such actions as a direct threat to intellectual property and national competitive advantage, akin to traditional industrial espionage.

But a counter-argument quickly emerged, summarized by Coogan himself: “everyone's saying, 'Hey, Enthropic distilled on my GitHub. They distilled on my writing. They distilled on my blog post. They distilled on my YouTube videos. Everyone's distilling me. Why are you getting upset when China's distilling on them now?'” This "pot calling the kettle black" perspective points out the hypocrisy of labs that extensively train on publicly available data, then cry foul when their outputs become fodder for others' models. Jordi Hays echoed this sentiment, adding, “I mean, it would be very silly not to try to look at other models and try to understand how that they work.” This side frames distillation as a natural part of competitive research and development, similar to how Ford engineers “drive the crap out of” Chinese EVs, then “disassemble and reassemble to understand how they're built.”

Who's Right (and When They're Wrong)

Both sides carry significant weight, and the truth, as always, lives in the messy middle. Coogan is right that blatant, industrial-scale, covert theft designed to clone a competitor's core proprietary model without substantial independent innovation crosses a line. When a company invests billions in R&D to create a unique piece of technology, deliberately circumventing that investment through large-scale copying undermines the very incentive structure of innovation. This is especially true when national security or strategic technological dominance is at stake, as the US government implies.

However, the "pot calling the kettle black" argument can't be dismissed. If your frontier model was largely built by consuming the internet—including the creative output of millions—then expecting absolute immunity from others learning from your model's output is naive. The line likely lies in intent and method: Is the goal to learn and innovate, drawing inspiration and insights, or to systematically reproduce a proprietary advantage? The act of “looking at other models” to “understand how they work” (as Hays suggests) is fundamental to progress. The ethical and legal gray area arises when that "looking" morphs into wholesale replication via sophisticated, covert platforms. For consumers and small businesses, cheaper, distilled models can be a boon, democratizing access to powerful AI capabilities, but this benefit doesn't automatically absolve potential IP infringements.

What to Do With This

As a founder building in AI, assume your model's output will be a data source for others. This week, audit your AI strategy with two things in mind. First, if you're building proprietary models, protect your data moat (unique, hard-to-replicate datasets) or your application layer (proprietary UX, integration, or vertical expertise) more than the raw model weights. Your competitive edge will shift from model architecture to what only you can train it on or how you deliver value. Second, if you're building on top of frontier models, plan for dramatic cost deflation. Cheaper, distilled versions will emerge, making basic AI capabilities a commodity. Your differentiation won't be model access, but how you integrate, customize, and deliver a superior customer experience on top of those increasingly commoditized capabilities.