Twenty-five years ago it was cryptography, today it's model weights

Most of the conversation accepted the analogy at a high level. Once weights leak or can be reproduced, scarcity gets hard to enforce. Several commenters pushed the comparison further by saying the real pattern is not just export control but control over infrastructure and distribution. DRM came up as the closest consumer example. The recurring point was that a lock defeated once is defeated for everyone, and the people left carrying the cost are regular users who get a worse product. That led into a broader anti-incumbent view that streaming, app stores, and frontier model APIs all use “safety” and “rights” language to preserve business leverage after the technical barrier has already failed. The sharper part of the discussion was where the analogy breaks. Cryptography is compact, stable, and finished enough to print on a shirt. frontier models are huge, expensive, and still moving. Several people argued that the real moat is not secret insight in the weights so much as compute, electricity, training time, and the ability to repeatedly push the frontier. That makes weight controls feel less like controlling math and more like trying to slow proliferation of a costly industrial capability. Even people sympathetic to open weights said the thing regulators actually care about is not a 2B local model but the newest giant systems and the tooling around them. A second fault line was over danger. The most concrete pushback said the cryptography analogy collapses if you accept the premise that a model can autonomously do real damage. That claim was answered with a more technical distinction: an LLM is a static model, while the hazardous part is the software stack that gives it tools, network access, execution rights, and no human supervision. In that framing, recent “model escaped containment” rhetoric is doing political work for labs that want attention and tighter control, while hiding ordinary engineering negligence. The mood was skeptical of both state restriction and frontier lab self-justification. People broadly expect open weights to keep spreading, partly because knowledge diffuses and partly because the control story serves too many commercial interests to trust at face value. The practical consensus was not “there is no risk.” It was that restricting weights is a blunt instrument, and the more useful place to look is who controls compute, deployment, and the surrounding software systems.