the ex-openai cto's $12B startup admitted it trained its first model on a chinese competitor's outputs. then said the next one won't do that.

Inkling’s data card lists Kimi K2.5 outputs in its post-training provenance. That’s Moonshot AI — a Chinese lab. Thinking Machines is the company the press tagged as the Western answer to Chinese open-weight AI.

Pre-training was from scratch, and Kimi’s outputs only touched an early bootstrap phase before large-scale RL took over. The “clean Western alternative” framing was always media-constructed — Thinking Machines never explicitly claimed it. But they raised $2B inside that narrative, didn’t push back when it helped them, and now the data card exists. The framing lands on them anyway.

Their response: the next model will use “fully self-contained post-training.” Future tense, delivered in the same sentence as the admission that this one wasn’t.

You raise $2B as the Western answer to Chinese open-weight AI, bootstrap on Chinese open-weight outputs, then someone reads the data card. The next one will be cleaner — the AI lab equivalent of “we’ve updated our policies.”