Moonshot's Kimi K3 model accused of stealing Anthropic technology

A new controversy is brewing in the world of AI: White House advisors have accused China's Moonshot of illegally copying technology from the US-based Anthropic's Fable model to create its Kimi K3. Amid these allegations, Washington is considering imposing restrictions on Chinese open-source models. This is reported by Techcrunch.com reports .

White House science advisor Michael Kratsios stated that the Kimi K3 model was created not only by stealing American intellectual property but also by utilizing high-tech chips that are subject to export bans. According to him, such large-scale industrial espionage causes serious damage to American research. US Treasury Secretary Scott Bessent also confirmed that "watermarks" characteristic of American developments are being found in many Chinese models.

Expert skepticism: Technologically impossible

Although political circles speak of theft, industry experts are skeptical that a model could be created simply by copying in such a short time. Snorkel AI founder Braden Hancock believes that given the Fable model was introduced in early July, it is technically impossible to reprocess such a massive database and successfully launch a new model in just two weeks.

Experts also touched upon the "distillation" process, where one model is trained based on the principles and responses of another. However, Allen Institute for AI researcher Nathan Lambert says that Chinese models are already approaching global standards and simple copying no longer yields significant results. He believes the secret to powerful models like Kimi K3 may lie not just in data theft, but in complex training methods.

Complex training and infrastructure issues

Supervised fine-tuning using another system sometimes causes a new model to behave like Claude or other popular AI. This situation may have triggered the perception of "theft" among politicians. But Lambert emphasizes that achieving real growth requires complex infrastructure involving tens of millions of agents and reinforcement learning.

So far, Moonshot has not provided an official response to questions about its training processes. Nevertheless, this situation is expected to trigger a new phase in the technological cold war between the US and China. If a ban is placed on Chinese open-source models, it could completely shift the balance of power in the global AI market.

According to TechCrunch, stealing data via the Anthropic API is an extremely expensive and time-inefficient process. Therefore, it is highly likely that the success of the Kimi K3 model lies in unique engineering solutions rather than just theft.