Jing Yang

This is a BIG deal for AI in China and potentially worldwide, but open models can still flow for now. PSA: I have read the full Q&A and control list (checking Chinese original next to Claude translation, and using Claude to help understand the implication of the key technical thresholds for control. ) My thoughts: 1) Any Chinese model *not already out yet* we will care about is on the restricted list that requires permission to export--including releasing model weights. Any model already open weighted is not included. Restriction covers any model in top 10 globally in use, general performance, specific task performance, safety performance or training efficiency. This regime is in place already, though it is not final and may adjust substantially. Another set of criteria focused on cyber would capture anything at or above glm5.3 (there are many thresholds but one is strong performance on cyber benchmarks, another covers vulnerability discovery, etc). Agents are also covered, and this probably would hit anything like Kimi agent swarm scale (60+ agents). Some of the thresholds are based on future capabilities, but any new model as good as existing Chinese models would be under this new regime. 2) these are restricted (as are models trained on "important data"), while models trained on Chinese “core data” are blocked from export. Core and important data are very very messy and dynamic categories depending on sector so prob Chinese labs will not easily be able to figure this out (though there is a strict multi-part de minimis for important data that some will use). "Data related to national security, the lifelines of the national economy, important aspects of people’s livelihoods, major public interests, etc., constitute core national data" 3) the single most important line is in the MOFCOM Q&A saying there will be automatic approval for future open models, as @ZilanQian has pointed out. The question is then what conditions apply to get that (eg even if gov doesn’t review application for case by case approvals pre-release, do they need safeguards? what reporting is required to qualify for automatic approval? what ongoing oversight? The rules give us nothing to go off on this--surely this part will be core to the comments they receive from industry as they tweak the system. 4) either way 3 shakes out, they now have the system in place to throttle future release of open models if they choose: CN gov can turn this into a hurdle by removing automatic approval. 5) Also potential implications for the main way people globally access CN models today--through US cloud services. The Q&A implies Chinese labs sharing model weights with US cloud companies to serve the models on data centers outside China will also need gov approval (but serving foreign users through CN labs' API, or serving from a CN company's own data centers outside China are not).