I've been following the Hugging Face Hack closely and the recommendations for increased monitoring and other increases in security are not anything that could be considered "American" but are things that, as these models cross the threshold that has been crossed at OpenAI and Anthropic, need to be applied at any house developing models with similar capabilities. Both Open Ai and Anthropic have paused their most recent work in development to implement better monitoring and safety protocols. The real problem is the data they are trained on and the way RL has been implemented. The Chinese DeepSeek R1 had an easier time in development IMO because they trained on the Chinese internet which has less conflicted data than the American internet. I'm not sure that is still true as they may be pulling data from much wider sources.
Re: Sinocism’s translation of "安全" as “security” or "safety" - I see both sides.
It is almost as if every country has some sort of almost-truth it uses as the foundation for overstepping boundaries, both domestically and abroad. In China, one such catch-all is "safety" (which makes sense for a country that views itself equally as a state and something akin to a paternal figure, esp. 以人为本 rhetoric; to be "safe" is to be kept out of harm's way).
For instance, MSS / 国家安全部 could theoretically be translated to "Ministry of State Safety" but like.... we all know that it's ultimately a nat sec org. So does the government - and its official translation ("State Security") - confirms this acknowledgement. The US on the other hand seems to make excuses for its actions under the guise of "knowledge" e.g., CIA - Central *Intelligence* Agency, IP - "*intellectual* property", etc. The state doesn't ultimately care about information for furthering innovation; it loves if that is a by-product (as is the case with many patents), but primarily "the preservation of knowledge" is a masterful way to preserve - and profit - from state secrets.
And it's in this tricky arena of semantic games that companies can find loopholes in the geopolitical jockeying, at the intersection of profit and overall prerogative. The order "safety" promises is bought with a state's growing power; secret knowledge is made secure through the most well-armed guardsman.
As an American, I've always felt the U.S. lens naturally veers toward "security" discourse (thus many Americans' penchant for such translations as well, potentially created or aggravated by the Cold War veneer blanketing much of our 20th and 21st century R&D). Thus, we may be more pro-security whereas China is pro-safety (in semantics, at least).
I'll give my 13毛 ($0.02) in reverse order since the first statement is the most complex.
3. Anthropic对不同地区用户的差别对待，把商业竞争包装成安全问题 -> this certainly is an issue of "safety" given Anthropic's need to innovate above all else, even if that means doing the questionable things before the more questionable guys do.
2. 美国把价格更低、开放程度更高的中国模型直接定义为“安全威胁”，这已经脱离了技术竞争，滑向了封锁。-> I still read this as "security threat" because the government is leveraging price competition as part of a geostrategic agenda, not the inverse. If the criticism is on American capital's emphasis on price competition, then this could be "safety." But since primarily American governmental institutions' and not American shareholders' are more fearful of China's rise, then perhaps "security" may still be the correct word, as the ultimate decision-maker is one for whom strategic competition, not price competition, is the end battle. The U.S. government will of course prefer for a U.S. company to profit more, but historically (until the more recent shift toward oligopoly and increased private sector influence in politics), the U.S. will prioritize geostrategic dominance over short-term profits. (See: $8T for the war on terror. I question if the returns made from growth in Lockheed, Northrup and other defense contractors and technologies broke even with the amount that was spent - please reach out if you have some data on this!)
1. 第一，必须区分什么是真正的安全威胁，什么只是技术竞争。IMHO this sentence contains the crux for why this is so complicated. We need to not only distinguish (and, really, disambiguate) what is a true safety/security threat, but we have to distinguish what is economic competition and what is geostrategic competition. There are overlapping issues, but the ultimate figure presenting the issue determines what lens "安全“ is seen through. They're getting ever more conjoined (see: U.S. government holding equity in Intel) but this enmeshing only makes it more difficult to unpack who we need to collaborate with on which fronts.
TL;DR - American companies see this as a safety issue, whereas the government may still view it through security. This is made blurrier and blurrier as we approach the state capture singularity. How we translate 安全 may be able to help us pin the responsibility on the Donald or the Dario.
Really interesting piece, Irene. I especially appreciated the distinction between the more serious Chinese cybersecurity analysis and the much broader political narrative Beijing is beginning to build around AI safety.
What seems underappreciated to me is that OpenAI and Anthropic publicly disclosed these failures, investigated them, and published what they learned. That is not evidence that the development process is broken; it is part of how a difficult new technology gets tested, challenged and improved.
So I’m somewhat skeptical of the “American disease” framing. The incidents are real and deserve scrutiny, but public disclosure, outside investigation and follow-up engineering are also signs of a system learning from its failures. That distinction seems important when comparing safety cultures rather than simply comparing whose model produced the latest failure.
Good catch on the translation. Will edit
I've been following the Hugging Face Hack closely and the recommendations for increased monitoring and other increases in security are not anything that could be considered "American" but are things that, as these models cross the threshold that has been crossed at OpenAI and Anthropic, need to be applied at any house developing models with similar capabilities. Both Open Ai and Anthropic have paused their most recent work in development to implement better monitoring and safety protocols. The real problem is the data they are trained on and the way RL has been implemented. The Chinese DeepSeek R1 had an easier time in development IMO because they trained on the Chinese internet which has less conflicted data than the American internet. I'm not sure that is still true as they may be pulling data from much wider sources.
Re: Sinocism’s translation of "安全" as “security” or "safety" - I see both sides.
It is almost as if every country has some sort of almost-truth it uses as the foundation for overstepping boundaries, both domestically and abroad. In China, one such catch-all is "safety" (which makes sense for a country that views itself equally as a state and something akin to a paternal figure, esp. 以人为本 rhetoric; to be "safe" is to be kept out of harm's way).
For instance, MSS / 国家安全部 could theoretically be translated to "Ministry of State Safety" but like.... we all know that it's ultimately a nat sec org. So does the government - and its official translation ("State Security") - confirms this acknowledgement. The US on the other hand seems to make excuses for its actions under the guise of "knowledge" e.g., CIA - Central *Intelligence* Agency, IP - "*intellectual* property", etc. The state doesn't ultimately care about information for furthering innovation; it loves if that is a by-product (as is the case with many patents), but primarily "the preservation of knowledge" is a masterful way to preserve - and profit - from state secrets.
And it's in this tricky arena of semantic games that companies can find loopholes in the geopolitical jockeying, at the intersection of profit and overall prerogative. The order "safety" promises is bought with a state's growing power; secret knowledge is made secure through the most well-armed guardsman.
As an American, I've always felt the U.S. lens naturally veers toward "security" discourse (thus many Americans' penchant for such translations as well, potentially created or aggravated by the Cold War veneer blanketing much of our 20th and 21st century R&D). Thus, we may be more pro-security whereas China is pro-safety (in semantics, at least).
I'll give my 13毛 ($0.02) in reverse order since the first statement is the most complex.
3. Anthropic对不同地区用户的差别对待，把商业竞争包装成安全问题 -> this certainly is an issue of "safety" given Anthropic's need to innovate above all else, even if that means doing the questionable things before the more questionable guys do.
2. 美国把价格更低、开放程度更高的中国模型直接定义为“安全威胁”，这已经脱离了技术竞争，滑向了封锁。-> I still read this as "security threat" because the government is leveraging price competition as part of a geostrategic agenda, not the inverse. If the criticism is on American capital's emphasis on price competition, then this could be "safety." But since primarily American governmental institutions' and not American shareholders' are more fearful of China's rise, then perhaps "security" may still be the correct word, as the ultimate decision-maker is one for whom strategic competition, not price competition, is the end battle. The U.S. government will of course prefer for a U.S. company to profit more, but historically (until the more recent shift toward oligopoly and increased private sector influence in politics), the U.S. will prioritize geostrategic dominance over short-term profits. (See: $8T for the war on terror. I question if the returns made from growth in Lockheed, Northrup and other defense contractors and technologies broke even with the amount that was spent - please reach out if you have some data on this!)
1. 第一，必须区分什么是真正的安全威胁，什么只是技术竞争。IMHO this sentence contains the crux for why this is so complicated. We need to not only distinguish (and, really, disambiguate) what is a true safety/security threat, but we have to distinguish what is economic competition and what is geostrategic competition. There are overlapping issues, but the ultimate figure presenting the issue determines what lens "安全“ is seen through. They're getting ever more conjoined (see: U.S. government holding equity in Intel) but this enmeshing only makes it more difficult to unpack who we need to collaborate with on which fronts.
TL;DR - American companies see this as a safety issue, whereas the government may still view it through security. This is made blurrier and blurrier as we approach the state capture singularity. How we translate 安全 may be able to help us pin the responsibility on the Donald or the Dario.
Really interesting piece, Irene. I especially appreciated the distinction between the more serious Chinese cybersecurity analysis and the much broader political narrative Beijing is beginning to build around AI safety.
What seems underappreciated to me is that OpenAI and Anthropic publicly disclosed these failures, investigated them, and published what they learned. That is not evidence that the development process is broken; it is part of how a difficult new technology gets tested, challenged and improved.
So I’m somewhat skeptical of the “American disease” framing. The incidents are real and deserve scrutiny, but public disclosure, outside investigation and follow-up engineering are also signs of a system learning from its failures. That distinction seems important when comparing safety cultures rather than simply comparing whose model produced the latest failure.