Peter Diamandis wanted to pull on a string before the conversation moved on. Chinese models such as Kimi were being accused of learning from closed American systems, he told Alvin Graylin on Moonshots. What was being said in China about that?
Graylin’s answer was blunt: he called the distillation controversy “more of a PR tool” for certain companies. When he singled out one lab as pressing the issue, Diamandis guessed OpenAI. “No, Anthropic,” Graylin replied.
His argument was not that Chinese labs never learn from competitors’ models. It was that borrowed answers cannot account for everything those labs have built—and that buying access to a leading model is not the same as knowing how to make one.
What distillation is, and where the dispute begins
Distillation transfers capabilities from one model to another by training it on the first model’s outputs. Those examples can include answers or reasoning traces: intermediate steps a model produces while working through a problem. A lab can use the technique to make a smaller, cheaper model learn from a larger one. It can also supplement other training with examples generated by another system.
Graylin described learning across labs as universal. Anthropic’s allegation, however, concerns how access was obtained and what the outputs were used for—not whether distillation has legitimate uses.
In its February 23, 2026 investigation account, Anthropic alleges that DeepSeek, Moonshot and MiniMax generated more than 16 million exchanges with Claude through approximately 24,000 fraudulent accounts, violating its access rules. It reports over 150,000 exchanges associated with DeepSeek, 3.4 million with Moonshot and 13 million with MiniMax.
Anthropic says its attribution drew on request metadata, correlations between internet addresses, infrastructure indicators and some corroboration from partners. It also says MiniMax shifted nearly half its traffic to a newly released Claude model within 24 hours during an active campaign. The account does not provide a complete observation period. Its reported responses include detecting suspicious behavior, strengthening account verification and sharing indicators of abuse.
That is a more specific accusation than simply using AI to help train AI. Graylin’s defense of distillation as a general technique does not resolve the alleged fraudulent access or prohibited extraction.
A query bill is not a replication budget
Graylin approached the allegations through their price tag. He said he had estimated what the queries would cost using average response lengths and prices for Anthropic’s highest-end models. His answer was roughly $2 million to $3 million across the three labs, with DeepSeek’s share, as he recalled it, only in the thousands.
His recollection of the underlying volume—around 20,000 accounts and “a million or two questions”—was substantially below Anthropic’s published total of more than 16 million exchanges. The discussion did not provide enough of his calculation to reconcile the figures.
Still, the estimate led him to a larger challenge: “If you have a model that you spend a billion dollars on and somebody can distill and duplicate with a couple million dollars, then the whole economics of frontier AI doesn’t make sense.”
The “if” carries much of that argument. An API bill—the metered charge for sending requests to a model over the internet—prices access to training examples. It does not include building the model that learns from them, training computation, researchers’ salaries, experiments or failed runs. Graylin’s estimate was neither a model’s total development cost nor a demonstrated budget for reproducing a frontier system, meaning one at the leading edge of capability.
His concern was that cheap imitation could shorten the period in which a leading lab earns a return on expensive research. But the exchange also raised the question of how much imitation alone can achieve.
The lawyers’ objection
Graylin’s comparison was Meta. He put its spending on Anthropic’s service at $100 million to $200 million a month and argued that, despite abundant access, computing resources and highly paid researchers, it had only recently produced a model he considered somewhat competitive. If access to outputs explained Chinese success, he asked, why had Meta not achieved comparable results earlier?
Alex, another panelist, objected that access and permission were different things. Citing what he described as public reporting, he said Meta feared litigation and had warned engineers against allowing Anthropic reasoning traces into development of its Muse models. On that account, Meta’s purchases were not evidence that its researchers could freely use Claude’s outputs for competing-model training.
Graylin remained skeptical and offered xAI as another lab with access but, in his view, fewer reservations about moving quickly. Alex countered with a possible indirect training route involving Cursor, the AI coding environment: he described it as using a Kimi-based model refined with reasoning traces from Claude interactions, and linked that account to his explanation of xAI’s progress. This was Alex’s interpretation of Cursor’s and xAI’s training, not a provenance record established in the discussion.
The disagreement left Graylin’s spending comparison unresolved. Large purchases do not establish that a company distilled a competitor’s model—or that it was permitted to. His broader answer was that research and engineering quality still matter: learning from useful examples does not replace the work of designing and refining the system that learns from them.
What answers do not reveal
Graylin’s concrete example was a claimed twentyfold reduction in KV-cache use by Chinese researchers. A KV cache stores intermediate calculations while a language model generates text, allowing it to reuse earlier work instead of repeatedly computing it. Reducing that storage requirement can lower the memory needed to serve a model, especially for long conversations.
He did not identify the model, comparison baseline or test conditions in this exchange, so the twentyfold figure cannot be treated as a general performance result. His point was about the kind of improvement involved: training on another model’s answers does not, by itself, disclose how that model organizes its internal memory or supply a new architecture for doing so.
“These are not things that you get from distillation,” he said. In his account, Chinese success also reflects original algorithms, architecture and the later training used to refine a model’s behavior.
Graylin also argued that learning runs in both directions. He recalled Claude identifying itself as Qwen when questioned in Chinese, and researchers laughing about that during his visit to Alibaba’s labs. He said they suspected American labs were learning from their downloadable models, where use would be harder to observe. Neither the reported self-identification nor the conversation establishes which data trained Claude.
The cache example gave his argument a firmer engineering focus than the identity anecdote. Even a challenger with a large collection of excellent answers still needs to build a system that learns from them and serves users efficiently. In Graylin’s telling, reducing the memory required to deliver those answers is part of the Chinese contribution—not something the distillation accusation explains away.