Can you say more on the chem/bio testing part, specifically the "it's the law" part? Other than this mention in the Kimi-K2 report I don't think I've seen that in other technical reports (doesn't appear to be in K3) and I'm not aware of any regulation that specifically mandates chem/bio testing. The second-hand word I'd heard was that the labs in China aren't testing for these or loss of control risks, but I could certainly be wrong on that.
CBRN risks have been standard assessment since GB/T 46347, and very much on the radar at the labs. WAIC sessions discussed this at length, and TC260 is about loss of control but targets « info on CBRN » as the thing that can be lost.
I can say for sure that this is very much a mounting concern, and by NeurIPS this year I expect you’ll see safety evals come standard (driven by cyber but will be a complete package). I’m working on a project to help speed this along, and care about it very deeply.
Yes, Zhou Bowen presentation at WAIC, for example, covered these risks, RSI, etc and lots of discussion of all of them during many open and closed door sessions at WAIC and on margins. The Chinese AI safety community has come a long way here since the last WAIC, but you have to be part of the ongoing discussions or you will not grasp this, as media coverage and understanding of China AI safety community is very limited. See the Concordia 2026 report and my Substack for more. Great piece here Caithrin and thanks for doing this, we need to have real as opposed to fake discussions of the frontier, not just more FUD....
The claim that Chinese AI models deny that Tiananmem Square happened is indeed weird and does not create much confidence in the article. They usually just politely say that they can’t talk about controversial topics. Plus, all this models can be abliterated because they are open weights, I.e. all limitations can be removed post-training.
Can you say more on the chem/bio testing part, specifically the "it's the law" part? Other than this mention in the Kimi-K2 report I don't think I've seen that in other technical reports (doesn't appear to be in K3) and I'm not aware of any regulation that specifically mandates chem/bio testing. The second-hand word I'd heard was that the labs in China aren't testing for these or loss of control risks, but I could certainly be wrong on that.
CBRN risks have been standard assessment since GB/T 46347, and very much on the radar at the labs. WAIC sessions discussed this at length, and TC260 is about loss of control but targets « info on CBRN » as the thing that can be lost.
I can say for sure that this is very much a mounting concern, and by NeurIPS this year I expect you’ll see safety evals come standard (driven by cyber but will be a complete package). I’m working on a project to help speed this along, and care about it very deeply.
Love your writing btw!
Yes, Zhou Bowen presentation at WAIC, for example, covered these risks, RSI, etc and lots of discussion of all of them during many open and closed door sessions at WAIC and on margins. The Chinese AI safety community has come a long way here since the last WAIC, but you have to be part of the ongoing discussions or you will not grasp this, as media coverage and understanding of China AI safety community is very limited. See the Concordia 2026 report and my Substack for more. Great piece here Caithrin and thanks for doing this, we need to have real as opposed to fake discussions of the frontier, not just more FUD....
The claim that Chinese AI models deny that Tiananmem Square happened is indeed weird and does not create much confidence in the article. They usually just politely say that they can’t talk about controversial topics. Plus, all this models can be abliterated because they are open weights, I.e. all limitations can be removed post-training.