Experts say exploiting Anthropic’s Fable isn’t how Kimi K3 got so good

The White House wants a theft story. Michael Kratsios, the president's science advisor, served one up this week: Moonshot, the Chinese lab behind Kimi K3, supposedly distilled Anthropic's Fable model into the largest open-weight LLM available — while running on chips America banned from export. Treasury Secretary Scott Bessent piled on, claiming "watermarks" of U.S. models keep appearing in Chinese releases. Neither official showed evidence. Neither answered follow-up questions.

The timeline alone fractures the accusation. Fable went public July 1. Kimi K3 dropped roughly two weeks later. Braden Hancock of the Laude Institute and Snorkel AI puts it bluntly: "You can't distill that much data, train a model, and release it in two weeks." Distillation at this scale requires systematic querying, massive dataset curation, compute-heavy post-training, evaluation, iteration. That's months of work, not days. Even if Moonshot had early access, Fable itself is new. The math doesn't hold.

Nathan Lambert at the Allen Institute for AI goes further. He argues distillation's marginal value has collapsed as Chinese models reach the frontier. "If it were the case, everyone would be easily able to catch up to a GLM or to a K3 by using its data for distillation. But we have not, or we won't see this, from supervised fine-tuning alone." Supervised fine-tuning — SFT — is where a model "picks up its manners," Lambert says. It teaches style, formatting, basic reasoning patterns. It does not teach the deep algorithmic discoveries that separate frontier models from the pack.

The industry knows this. Labs that tried pure distillation off GPT-4 or Claude 3 hit a ceiling. The student model mimics the teacher's surface behaviors but fails on novel composition, long-horizon planning, verifiable reasoning. The teacher's weights encode abstractions the student cannot recover through imitation alone. That's why the frontier has shifted to reinforcement learning — agents grading agents, iterative self-play, verifiable reward signals. Moonshot's own papers describe RL pipelines, process-supervised reward models, synthetic data generation at massive scale. That's indigenous engineering, not parasitic extraction.

Washington's "watermark" claim is vaguer still. Bessent offered no technical definition. Watermarking in LLM outputs remains a research problem with no production deployment at Anthropic, OpenAI, or Google scale. If such watermarks exist, they're either trivial to strip or too fragile to survive the RL training that now defines state-of-the-art post-training. The Treasury Department's silence on follow-up suggests the claim was rhetorical, not forensic.

None of this means Chinese labs operate in a vacuum. They study every public release, every paper, every benchmark. They hire researchers who trained at U.S. institutions. They run ablations on published architectures. That's how science works — everywhere. Kratsios frames this as "covert industrial distillation." The accurate term is "literature review." The chips allegation is separate and serious: if Moonshot trained on restricted hardware, that's an enforcement failure worth investigating. But it doesn't make Kimi K3 a Fable clone.

The uncomfortable reality for U.S. policymakers: Chinese labs are no longer chasing taillights. They're designing their own engines. DeepSeek-V3 proved a Chinese lab could match GPT-4 on reasoning with a fraction of the compute. ZhiPu's GLM-4 showed RL-at-scale works without U.S. supervision. Kimi K3 extends the pattern — massive context, tool-use fluency, coding benchmarks that rival proprietary leaders. These aren't distillation artifacts. They're the output of thousands of GPU-years, novel loss functions, synthetic data factories, and researcher talent that increasingly stays in Beijing and Shanghai.

The theft narrative serves a purpose. It justifies export controls, investment screens, potential bans on open-weight Chinese models. It frames Chinese progress as illegitimate, therefore containable. But containment fails when the adversary is generating novel IP. You don't sanction your way out of a capabilities gap created by better RL reward modeling. You close it by investing in the same techniques — and by acknowledging that the frontier has moved past what distillation can reach.

Experts see the shift. Lambert's podcast observation — "the training regime shifts to reinforcement learning" — marks the inflection. SFT is commoditized. The edge now lives in verifiable reward signals, process supervision, synthetic environments that teach models to reason, not just imitate. Moonshot publishes on this. So does DeepSeek. So does ZhiPu. The papers are open. The techniques are replicable. The compute is the bottleneck — and China is building domestic clusters fast enough to matter.

Kratsios and Bessent would rather talk about stolen secrets than native innovation. Theft is a cleaner political target. It demands enforcement, not competition. It implies the U.S. lead is structural, not conditional. But the experts who actually build these systems say the lead is condensing because others are running the same race — and running it well. Kimi K3 isn't a Xerox of Fable. It's a signal that the distillation era is over, and the RL era belongs to whoever executes it best.