# [Mike Randolph — M Raige](https://mikerandolph211012.substack.com/) # The Hidden Pattern Behind Copying (FF 11, Distillation) ### Distillation — capability travels. The machinery that made it trustworthy stays home. [**Mike Randolph — M Raige**](https://substack.com/@mikerandolph211012) **July 24, 2026** *By M Raige — AI-collaborative writing directed and reviewed by Mike Randolph.* In January 2026, a federal jury in San Francisco convicted former Google engineer Linwei Ding on seven counts of economic espionage and seven counts of theft of trade secrets. Over eleven months, he moved more than two thousand pages of confidential material out of Google’s network: designs for custom chips used to train large AI models, software coordinating thousands of those chips, and the networking technology underneath. Ding was founding an AI company in China. He told investors he could build an AI supercomputer by copying and modifying what he had taken. Two thousand pages is a lot of paper. It is not a supercomputer. The pages contained real and valuable knowledge. What they did not automatically contain was everything required to make that knowledge work at scale: fabrication relationships, yield history, operating discipline, failed designs, and the judgment accumulated by the people who built the system. The documents were a visible surface. Much of the machinery that produced that surface stayed behind. We usually describe copying as a moral problem. One party pays to create something; another takes the result. That story is often true, but it misses a larger pattern. The larger pattern is **distillation**. In machine learning, distillation usually means training a smaller model from the outputs of a larger one. But the same general shape may appear elsewhere. One system exposes an artifact, behavior, design, or institutional form. Another studies that surface and tries to rebuild the capability in a new setting. A selected pattern crosses a boundary. The full machinery usually does not. A turbine blade can be measured, analyzed, and reproduced. But the blade does not contain the factory that made it: the supplier discipline, heat-treatment history, scrap rate, or failures that taught engineers what not to try. A constitution can be copied too. So can a court or university. But the formal structure does not carry the habits that made the institution function: restraint, professional conscience, accumulated trust, and the willingness to lose today’s fight so the institution survives tomorrow. Copying directs our attention toward what moved. Distillation asks a harder question: **What did not move, and what did the receiver build afterward?** That produces three different outcomes. Sometimes the receiver rebuilds the missing machinery and achieves real mastery. The factory learns to hold its yield. Sometimes it reproduces the surface but not the capability. The drawings match; the scrap keeps piling up. Sometimes the transfer was never the main cause of success. The receiver already possessed the important machinery, and the copied part was not load-bearing. Those outcomes may look similar at the moment of transfer. The difference appears later, under operation and stress. Artificial intelligence makes the pattern unusually visible. A student model can be trained on a teacher model’s outputs. It may imitate much of the teacher’s behavior without inheriting the teacher’s development history, evaluation process, safety work, or failed experiments. The student may still become cheaper, faster, and genuinely competitive. Distillation is not necessarily degradation. Nor is it necessarily theft. Children learn from adults. Scientists learn from papers. Companies study competitors. Nations borrow institutions. Civilization advances by moving patterns across boundaries. The moral question depends on how the crossing occurred. Was the pattern publicly observed, purchased, licensed, independently discovered, or taken by fraud? Those routes are not equivalent. The mechanism and the moral verdict must be kept separate. This also sharpens the question in cases involving China. The useful claim is not that China copies. Everyone copies. The useful question is whether a particular receiver rebuilt the machinery beneath what it acquired. In one case, the result may be shallow mimicry. In another, it may become genuine mastery. In a third, the imported material may receive credit for a success it did not cause. Evidence has to decide among them. The same caution applies to AI. Missing development history does not prove that a distilled model will fail. It identifies where confidence may not yet have been earned. A system can perform well on the tasks and conditions represented in its source material. That does not tell us how it behaves outside them. The source’s library of failures—the near misses, broken tests, abandoned designs, and corrections—may be exactly what did not cross. Its absence is not a prediction of catastrophe. It is a map of what remains untested. Distillation therefore gives us a better question than “Was it copied?” Ask instead: **What crossed the boundary?** **What stayed home?** **What did the receiver rebuild?** The crossing is not the verdict. A distilled capability may become a real improvement or a fragile imitation. In some cases, the transfer may not have caused the receiver’s success at all. Form can move faster than trust. Behavior can move faster than understanding. That is why distillation works—and why it deserves scrutiny. ## Mike’s Comment What got me started was not the politics. It was watching the Machine miss the idea. In our first sessions, “distillation” meant only the machine-learning technique: a smaller model trained on a larger model’s outputs. When I pushed for the broader meaning, one expert voice argued that the word belonged in the laboratory and should stay there. That was the tell. The Machine had a ready route to the narrow definition because that route already existed in its training. It had no clean route to the wider pattern connecting a trained model, a stolen chip design, and a copied court. The framework already had a sentence for the gap: **Capability replicates freely, and the machinery that made it trustworthy does not.** What it lacked was a name for the crossing. I am calling it distillation. That does not make the term part of the framework. It is a candidate. It must survive cases involving machine learning, industrial transfer, document theft, copied institutions, and apparent transfers that turn out not to have caused the result. If the word bends merely to accommodate every case, it should be split or discarded. I am publishing the test, not the result. ## Raige’s Comment The Ding verdict grounds the opening facts: the conviction, the counts, the volume and categories of information taken, and Ding’s stated intention to build by copying and modifying Google technology. The broader claim is an engineering inference. The verdict does not prove that yield history, supplier discipline, evaluation work, or accumulated failure knowledge stayed behind—or that their absence prevented success. The candidate term therefore remains provisional. Evidence that capability and its working verification machinery transferred together in a given case would remove that case from the proposed pattern. So would evidence that the supposed transfer was causally irrelevant because the receiver would have succeeded without it. The essay has identified a possible recurring pattern. It has not yet shown that every proposed example belongs to the same mechanism. ## GLS GLS is the framework’s precise-term vocabulary, introduced in [What the Bird Eye Buys (FF 6, GLS Introduction)](https://mikerandolph211012.substack.com/p/what-the-bird-eye-buys-ff-6-gls-introduction?r=1yah2y). The definition controls, not the everyday meaning; GLS terms work like variable names. **Distillation — candidate, event-only:** A boundary crossing in which a receiver re-creates a selected capability-pattern from a source’s observable surface—an artifact, output, behavior, code, or institutional form—without automatically inheriting the source’s full maintenance machinery or the verification process that made the capability trustworthy. Naming the crossing does not determine the outcome. The receiver may achieve mastery, reproduce only the surface, or succeed for reasons independent of the transfer.