China’s AI-safety trajectory is not necessarily a delayed version of America’s. [1/6] People seem to have this mental model: Chinese models trail US models in capability by a given number of months and are involved in similar safety incidents after a delay. Whether this is true matters for what Chinese AI safety communities can contribute.

"Our investigation also revealed that user queries that Moonshot rerouted to Claude included sensitive information about various Moonshot customers. [...] These include: PLA-affiliated surveillance activity An engineer at a major PRC SOE" "DeepSeek rerouted requests from users that were attempting to use one of DeepSeek’s models through third-party or Anthropic coding harnesses, like Claude Code, the Claude Agent SDK, or OpenCode. [...] These cases include: A PRC technology company Russian defense agency PRC police surveillance"

I appreciate this as an initial step toward more transparent reporting on RSI. But there is still much further data we need to fully understand RSI. In particular, OAI disclosed some evidence about the inference compute usage, number of experiments/researcher, and to a lesser extent, how researchers interact with models of different capabilities (see the green boxes in the diagram). However, the evidence they provided is just on the input side (tokens, lines of code). We urgently need to know how these inputs get turned into algorithmic improvements and [probably internally deployed] AI capabilities. Even just a simple time series on those outcome variables would be extremely valuable. You can then know, as AI helps researchers do more AI R&D, how the corresponding algorithm efficiency grows (the blue arrow C -> A dot). Right now we are missing this key puzzle piece.

How do Chinese models' prices compare to US ones? I graph the "price frontier" of US and Chinese models. The graph tells you, for each ECI score (model capability), what's the cheapest US vs. Chinese model. Several takeaways: 1) Under ECI = 155, Chinese models tend to be cheaper. 2) above 155, US models are cheaper. 3) Alibaba and DeepSeek are the two main Chinese companies pushing cheap models.

Liang Wenfeng (DeepSeek CEO) on RSI: "Our model primarily aims to improve DeepSeek's own performance, providing more assistance as we develop the next version." "Our primary goal in creating a model isn't for everyone else to use it easily, but for us to use it easily. It has to be useful to us first. Once it's useful to us, developing the next version of the model will be much faster." (from the investor meeting leak, May 2026)

How much do Chinese AI companies pay their employees? In 2025: MiniMax pays $84.3 million for its 428 employees -> ~$197k per employee pays $202.5 million for its 1,094 employees -> ~$185k per employee Both figures include stock-based compensation. While WSJ says OpenAI’s stock-based compensation alone in 2025 reached an average of $1.5 million per employee!