The safety-as-a-service route raises a procurement question: who pays for findings that delay a launch? If the buyer can suppress an unfavourable evaluation, more safety spending could still produce little public knowledge. Could funding include a protected right to publish negative results, even when the customer walks away?
What I find most interesting here is that AI safety isn't just a technical problem — it's also an institutional one. The question of who funds safety research, and what those funders actually want from it, can shape which safety questions get studied in the first place.
Seeing Article 4 of the Charity Law spelled out helps a lot here. Once charity work sits under Party leadership and the Overseas NGO Law flips funding from 80–90% foreign to 80–90% domestic, a Chinese graduate who wants to spend five years on an odd alignment idea has nowhere independent to take it. What survives is whatever fits inside SHLAB, TC260, or a compliance product a developer will pay for. That also puts the jump from 12 to 57 frontier safety papers a month in context, since most of it comes out of universities and state labs. If the third wave of American philanthropy arrives after the Anthropic and OpenAI IPOs, the gap in open-ended research could widen fast, and Beijing seems unlikely to let much of that money cross the border.
When you talk about the CCP’s view on AI safety, remember their concerns about safety of hazardous viruses……which became COVID-19and before that how they handled SARS. Read Karl Taro Greenfield’s excellent book “China Syndrome” for a summary of that response. It was only because of the bravery of a couple of virologists in Hong Kong that this did not become a pandemic.
The safety-as-a-service route raises a procurement question: who pays for findings that delay a launch? If the buyer can suppress an unfavourable evaluation, more safety spending could still produce little public knowledge. Could funding include a protected right to publish negative results, even when the customer walks away?
What I find most interesting here is that AI safety isn't just a technical problem — it's also an institutional one. The question of who funds safety research, and what those funders actually want from it, can shape which safety questions get studied in the first place.
Great topic. Thanks
Seeing Article 4 of the Charity Law spelled out helps a lot here. Once charity work sits under Party leadership and the Overseas NGO Law flips funding from 80–90% foreign to 80–90% domestic, a Chinese graduate who wants to spend five years on an odd alignment idea has nowhere independent to take it. What survives is whatever fits inside SHLAB, TC260, or a compliance product a developer will pay for. That also puts the jump from 12 to 57 frontier safety papers a month in context, since most of it comes out of universities and state labs. If the third wave of American philanthropy arrives after the Anthropic and OpenAI IPOs, the gap in open-ended research could widen fast, and Beijing seems unlikely to let much of that money cross the border.
When you talk about the CCP’s view on AI safety, remember their concerns about safety of hazardous viruses……which became COVID-19and before that how they handled SARS. Read Karl Taro Greenfield’s excellent book “China Syndrome” for a summary of that response. It was only because of the bravery of a couple of virologists in Hong Kong that this did not become a pandemic.