I have to say something as a chinese

You need to know, as a chinese, we be teached don't say that much since we were kids or you will get in trouble unless you have to.

As someone familiar with China’s AI community, I have a few thoughts on Anthropic’s recent accusations. A lot of the framing strikes me as either misleading or badly lacking context. A few points:

First, the idea that a cheap model like DeepSeek would routinely have its requests forwarded to Claude makes little economic sense. It would be like ordering bottled water and being given wine for free.

What often happens is the reverse. Because of Anthropic’s access policies, many Chinese users who need Claude rely on third-party API providers. Some shady providers “water down” the service by routing a fraction of requests to cheaper models to cut costs. So if ten requests are sent through a reseller, perhaps eight really go to Claude and two are silently diverted elsewhere. That kind of substitution is a real issue—but it points in the opposite direction from the narrative Anthropic is pushing. Not Deepseek or kimi route to Claude, is someone route Claude

Second, on the alleged military and biological-weapons examples: according to material already posted by other Chinese bloggers, at least some of the cited prompts were questions from university exams and can be found publicly online. And the only one thing i saw is anthropic

Third, on “massive distillation”: yes, Chinese labs study outputs from other frontier models. But if a model and its training pipeline are built independently while engineers benchmark competitors and study their behavior, that is ordinary competitive research. But Chinese company do distillation, maybe 10% distillation and do 90% by themselve.

Calling all of that “massive distillation” is like saying Ford engineers committed some grave offense because they bought and drove a Honda to see how it works. Competitor benchmarking is normal in virtually every engineering industry.

Fourth, the “tens of thousands of accounts” argument also needs context. Many Chinese AI users must use VPNs to access overseas services. Large numbers of unrelated users can therefore appear behind the same VPN providers, IP ranges, or exit nodes.

None of this proves that no abuse ever happened. It does mean the evidence should be interpreted carefully. API abuse, reseller model substitution, competitive benchmarking, and systematic model extraction are very different things—and shouldn’t be blurred together.

I am not defending the CCP; I am simply fed up with being suppressed by Anthropic at the same time that I am being suppressed by the CCP. They treat the countless people who simply want to use cutting-edge AI as if we were demons, demonizing all Chinese people and companies. They view every Chinese person as evil and regard all our attempts to access their services as a threat to them.

One additional point: if Anthropic is really claiming that Kimi randomly routes some of its harder user requests to Claude, that makes very little sense either technically or economically.

First, Claude and Chinese models use different tokenizers, and more importantly, their KV caches and internal model states are not transferable. You cannot simply let Kimi think halfway through a problem and then hand the same reasoning state over to Claude. At best, you resend the conversation or prompt to Claude and start a fresh inference. It is similar to how switching from Sonnet to Opus halfway through a difficult task does not magically give Opus the reasoning state Sonnet already built up; starting a fresh Opus session can produce a very different result.

Second, if a Chinese AI company actually wanted Claude outputs for distillation, it could obtain Claude access through third-party APIs and deliberately generate much cleaner and more valuable training data. Paying Claude-level prices to process a random subset of ordinary Kimi users' messy prompts would be an extremely inefficient and expensive way to collect distillation data.

Again, I am not saying they do not distill. They probably do. I am saying Anthropic is deliberately framing ordinary competitive behavior in the most sinister way possible, and this particular accusation makes very little sense once you actually think through the technical and economic incentives.