Elon Musk Testifies xAI Partly Distilled OpenAI Models to Train Grok

Industrial espionage is as old as industry itself—Samuel Slater’s 1789 theft of British textile know-how and the Soviet Union’s 1940s copying of U.S. nuclear designs both relied on extracting secrets to leapfrog costly R&D. Musk’s admission shows that, in the AI era, the extraction can happen legally through an API instead of a spy’s briefcase. It underscores two converging long-term trends: (1) the escalation of IP battles from hardware to intangible model weights and (2) the erosion of clear property boundaries in software when marginal copying costs approach zero. Over a 100-year horizon, moments like this determine whether AI knowledge becomes a guarded commodity (protected by law, sanctions, and gatekeeping) or a quasi-public good continuously re-cycled by competitors. If even well-capitalised U.S. labs resort to distillation shortcuts, future regulation may have to choose between policing an unenforceable line and accepting that frontier capabilities diffuse rapidly regardless of national borders.