The White House Says the Model I Run Every Day Was Stolen From Claude

I pay Anthropic $200 a month for Claude. I also run Kimi K3, the open-weight model from the Chinese lab Moonshot AI, as the daily brain of the agent setup that runs my projects. Those two facts got along fine until Wednesday, when the White House accused the second one of being built with stolen goods from the first.

So I've got a foot in both camps of this fight. Here's what actually happened, and what it looks like from where I sit.

Michael Kratsios, the White House's top science advisor, posted that the administration "has information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model." He said Moonshot built "a sophisticated internal platform" to query American models at scale while switching between access methods to avoid detection. He also claimed Moonshot got its hands on Nvidia's GB300 systems, chips banned from export to China, either directly or through servers in Thailand. A day earlier, Treasury Secretary Scott Bessent said the government could sanction Chinese AI companies over exactly this, and claimed officials are finding "watermarks of our U.S. large language models on many of the Chinese models."

Then on Friday, 25 companies, including Nvidia, Microsoft, Meta, Palantir, Hugging Face, and Mistral, published an open letter urging Washington not to respond with broad restrictions on open-weight models. Distillation is a normal, widely used training technique, they argued, and the answer to actual theft is targeted legal action, not bans. OpenAI and Anthropic did not sign. Neither did Google DeepMind. Elon Musk backed the letter on X, though SpaceX didn't sign it.

I'm not going to pretend the accusation is crazy

Back in April I wrote about a Reddit user who caught Kimi K2.5 calling itself Claude in long conversations. That happened one day after Anthropic published a report naming Moonshot, DeepSeek, and MiniMax for running distillation operations against its models, millions of exchanges, traced back through IP addresses and usage patterns that Anthropic said looked like deliberate capability extraction. I called the identity slip a tell, and I still think it was one. Models trained on another model's outputs pick up its manners. Sometimes they pick up its name.

So when the White House says it has information, my reaction isn't "no way." It's "show me."

The timeline problem

Because here's the thing the accusation doesn't explain. K3 came out last week. Per TechCrunch's reporting, Fable has only been publicly available since July 1st. Braden Hancock, a researcher at the Laude Institute, put it plainly: you can't distill that much data, train a model, and release it in two weeks. Nathan Lambert at the Allen Institute for AI has been saying distillation matters less and less as Chinese labs get closer to the frontier and shift to reinforcement learning.

And notice who hasn't confirmed the specific claim. Anthropic's policy chief thanked Kratsios for speaking out, but when TechCrunch asked Anthropic directly whether Fable was distilled to make K3, the company didn't respond. Kratsios himself shared nothing about where his information comes from. "We have information" is doing a lot of work in that sentence.

None of this clears Moonshot. Hancock's own guess is that earlier American frontier models probably contributed to Kimi's training somewhere along the line. But there's a big gap between "Chinese labs have distilled our models before" and "this specific model, released last week, is stolen." The government is threatening sanctions and bans over the second claim while showing evidence for neither.

Everyone's position maps to their business model

Follow the incentives and this whole story gets clearer. Nvidia sells the chips. Microsoft rents the cloud. Meta gives its models away. Of course they signed. Open models being everywhere is good for all of them.

The two companies that didn't sign are the two with IPOs coming, per CNBC's reporting, both valued near a trillion dollars, both selling closed models at full price while a free competitor a third of the cost trades punches with their flagship. Their absence isn't a mystery. It's a business decision.

And the technique at the center of all this? Everybody uses it. Musk himself testified earlier this year that his AI company distilled OpenAI's models to build Grok, per TechCrunch. Kratsios even admits legitimate distillation "plays a vital role" in the ecosystem. Replit's CEO was blunter: banning Chinese open models is as good as banning open models, period. He pointed out that Thinking Machines' new open model was trained with help from Kimi 2.5. It's one ecosystem, and the models cross-pollinate whether Washington likes it or not.

What a ban would actually take from me

Let me be honest about my own stake here, since I'm asking you to see everyone else's.

A week ago I wrote about running K3 inside my agent setup and changing my mind about it. That post ended with a prediction: this fight is about to get a lot louder. It took eight days. K3 costs me $3 per million input tokens where Fable costs $10. It handles the real, vague, figure-it-out-yourself work I throw at it daily. If the administration bans Chinese open-weight models, the person who loses isn't Moonshot's CEO. It's me, and every other builder who finally got a capable model at a price that doesn't require a venture fund.

There's also a question nobody in Washington seems to be asking: what does regular-person AI cost in a world with no open competition? I pay $200 a month already. Without pressure from K3 and models like it, what's stopping that from being $300? Closed labs with trillion-dollar valuations and no price competition. That's the world the non-signers are hoping for.

If Moonshot stole, prove it. Anthropic says it can fingerprint distillation in its API traffic, and the White House says it has information. Great. Put the evidence out, name it in court filings, sanction whoever the evidence actually points to. That's how this is supposed to work in a country that runs on proof.

But don't tell me the model sitting on my machine right now, the one I watched reason through real work with my own eyes, is contraband, based on a tweet thread and the word of officials who haven't shown a single document.

Show me the watermarks.