If someone watched/memorized how Qwen gave birth 3 years ago, the debut model itself was previewed/announced as a personalized LLM. For half a year, I’m responsible for its character in 14b size, also as the suggested model for game, edu, and smart devices. I even invited one of the most trending stand-up comedians/professors to be part of the post-training /alignment corpus. However, at that time, people didn't realize RL could scale up like today’s~~~
A good piece. I think Chinese labs are just later here, with a less philosophical group at the core of building — like ai safety researchers built the generational AI companies here — but I’m working on some stuff showing Chinese labs have interest in refining this area. Which is great for research & safety.
I believe that's the reason and result of still chasing behind. china pre-train teams haven't refined pipelines and experts on principle & taste of reinforcement learning. But bytedance an exception. That's also a systematic difference from Silicon Valley, researcher outcome in china are more likely rooted with internet project sprints, shorter venture capitals duration, less go-public tunnel. Researchers have to abide the performance indice in every half a year.
> You may not have the exact methodology, but you have the outcomes of those methodologies.
It's also worth noting that Anthropic's Evan Hubinger was the last author on the open source replication of character training! Not quite the same as constitution, but not that different either
(Incidentally, Nathan Lambert was the third author)
Kai Williams with a deep lore... I agree that character and constitution are muddy concepts right now in my thinking. More research is needed, and I want to get better at talking about these concepts as distinct.
Well I sat next to the first author of that paper when I was doing AI safety research. So I should know a little bit about that haha.
My sense is that Anthropic’s character training started with a constititution of about 9 statements or so, then used some techniques to instill those more simple statements in. Eventually, they’ve gotten more sophisticated though on the size and the content of the document. But I think of constitution as just a more holistic way of approaching character training rather than something fundamentally differnet.
The Chinese labs don't have an Amanda because, as they fundamentally don't overindex on soft power, aka perception of goodness in the mindshare of the public, theorizing on the future almost feels frivolous as opposed to earning the rightful place in the stack. It's why they don't prioritize comms and PR or understand it in the limited sense of announcementa/influencer strategy/controlled narrative. Build in public is fine, but think in public feels ominous. After all, how does one put a KPI to philosophizing?
I agree with you, but I also think this represents the single most important asymmetric advantage available to Chinese AI right now. There's a way to make character and constitution feel like another step in the technical foundation of a great model and not some Berkeley beanbag philosophizing activity. I think having someone embrace this is inevitable. It's just a question of who will take the leap first.
This is a bad take. Many things worth pursuing cannot be quantified. Even in business. Also philosophizing is not the goal here, it’s the method to create an aligned AI. Alignment is a difficult problem to solve and AIs might lead to human extinction. This is in addition to the fact that manufacturing a personality for AI is not that difficult, and is also vital to why people prefer certain AIs? It’s a huge missed opportunity from the Chinese. It’s low hanging fruit.
> Similarly, at Kimi, K2 has always seemed to me to be an emotionally fluent and stylistic model, although its English-language writing still emulates Claude to a large extent.
I think we still don't have any proof that Anthropic models are good thanks to or despite the focus on character. I think Amanda Askell does great work, I also think it wouldn't matter much if Anthropic models were not competitive on capabilities. Case in point : more people seem to be obsessed with ChatGPT 4o than any version of Claude.
>From a safety perspective, it is probably a net benefit that everyone is distilling from Claude. Anthropic seems to take safety more seriously than almost any other lab, and I think they are creating the conditions whereby other open models inherit this precaution through distillation.
It's still not clear to me how this whole distillation thing is supposed to work, especially when Chinese labs then ship more cost effective models than American labs. Why isn't Claude Haiku 4.5 good? Why isn't Sonnet 5 better? How are Chinese labs better at distillation than labs that have access to the models? Why aren't more American labs doing it, and better? It doesn't seem to be illegal.
My biggest question is: why do we (as of everyone, not of Chinese labs) need an Askell role in the first place? I can totally understand Anthropic believing that she can teach Claude to be "good", and I think it is good to have the role. However, given that these important models from the US and China are going to influence, whether drastically or subtly, how billions in this world think and act, I am scared about having a handful of "philosophers" selected by these labs to determine the direction that is so critical to shaping the world at large. It seems to me more like another round of cultural colonialism (regardless of where and initiated by whom) rather than a net benefit thing.
I thought a lot about this. I think the honest answer is “because it makes better models”. Claude feels more polished than almost any other language model. I can’t say for certain that it’s the role of the ASCLs team that makes it feel that way, but the eloquence of her thinking and the model’s output seem connected.
I agree claude is very good in terms of capability, but its philosophy & "personality" may not look good for everyone. Ultimately, I am not against having one more prominent philosopher in the world, but I am scared of having ONLY one school of thought. I am not saying that Chinese labs should hire a Confucius or Marxist philosopher (would be a nightmare), but feeding only Askell's interview to every good and widely-used model seems a very uncomfortable idea...
Well, the alternative isn’t everyone having a part in it - the alternative is letting the benchmarks labs are optimizing against do all the character training, which will be even more colonialist.
Chinese tech forums and Bilibili comment sections have a pattern worth noting: the most common criticism of domestic AI models is not benchmark performance but what gets called 没个性 -- no personality. Chinese users actively complain about blandness more than Western users do. The demand for character exists. The supply question is whether it can survive a benchmark cycle where moving from flagship 3.5 to 3.6 to 3.7 in three consecutive months is the dominant competitive signal -- and where the character investment has no clear delta to show for it.
so interesting!! re: Doubao, my somewhat flippant opinion is that it might represent a different race to the bottom on character - the cute-little-girl thing is frankly very uncomfortable, but sells.
If someone watched/memorized how Qwen gave birth 3 years ago, the debut model itself was previewed/announced as a personalized LLM. For half a year, I’m responsible for its character in 14b size, also as the suggested model for game, edu, and smart devices. I even invited one of the most trending stand-up comedians/professors to be part of the post-training /alignment corpus. However, at that time, people didn't realize RL could scale up like today’s~~~
Oh my god, I could talk to you about this for a hundred years. I'm dying to learn more. Drop me a note. I would really really like to talk about this.
A good piece. I think Chinese labs are just later here, with a less philosophical group at the core of building — like ai safety researchers built the generational AI companies here — but I’m working on some stuff showing Chinese labs have interest in refining this area. Which is great for research & safety.
I believe that's the reason and result of still chasing behind. china pre-train teams haven't refined pipelines and experts on principle & taste of reinforcement learning. But bytedance an exception. That's also a systematic difference from Silicon Valley, researcher outcome in china are more likely rooted with internet project sprints, shorter venture capitals duration, less go-public tunnel. Researchers have to abide the performance indice in every half a year.
> You may not have the exact methodology, but you have the outcomes of those methodologies.
It's also worth noting that Anthropic's Evan Hubinger was the last author on the open source replication of character training! Not quite the same as constitution, but not that different either
(Incidentally, Nathan Lambert was the third author)
https://arxiv.org/pdf/2511.01689
Kai Williams with a deep lore... I agree that character and constitution are muddy concepts right now in my thinking. More research is needed, and I want to get better at talking about these concepts as distinct.
Reading the paper now, thank you friend :)
Well I sat next to the first author of that paper when I was doing AI safety research. So I should know a little bit about that haha.
My sense is that Anthropic’s character training started with a constititution of about 9 statements or so, then used some techniques to instill those more simple statements in. Eventually, they’ve gotten more sophisticated though on the size and the content of the document. But I think of constitution as just a more holistic way of approaching character training rather than something fundamentally differnet.
The Chinese labs don't have an Amanda because, as they fundamentally don't overindex on soft power, aka perception of goodness in the mindshare of the public, theorizing on the future almost feels frivolous as opposed to earning the rightful place in the stack. It's why they don't prioritize comms and PR or understand it in the limited sense of announcementa/influencer strategy/controlled narrative. Build in public is fine, but think in public feels ominous. After all, how does one put a KPI to philosophizing?
I agree with you, but I also think this represents the single most important asymmetric advantage available to Chinese AI right now. There's a way to make character and constitution feel like another step in the technical foundation of a great model and not some Berkeley beanbag philosophizing activity. I think having someone embrace this is inevitable. It's just a question of who will take the leap first.
It would be an idiosyncratic advantage, and the one who gets it will leapfrog past all the rest.
This is a bad take. Many things worth pursuing cannot be quantified. Even in business. Also philosophizing is not the goal here, it’s the method to create an aligned AI. Alignment is a difficult problem to solve and AIs might lead to human extinction. This is in addition to the fact that manufacturing a personality for AI is not that difficult, and is also vital to why people prefer certain AIs? It’s a huge missed opportunity from the Chinese. It’s low hanging fruit.
Great post!
Thanks V! Back in SF Saturday let’s hang out I miss ya
> Similarly, at Kimi, K2 has always seemed to me to be an emotionally fluent and stylistic model, although its English-language writing still emulates Claude to a large extent.
Well, we now know why
I think we still don't have any proof that Anthropic models are good thanks to or despite the focus on character. I think Amanda Askell does great work, I also think it wouldn't matter much if Anthropic models were not competitive on capabilities. Case in point : more people seem to be obsessed with ChatGPT 4o than any version of Claude.
>From a safety perspective, it is probably a net benefit that everyone is distilling from Claude. Anthropic seems to take safety more seriously than almost any other lab, and I think they are creating the conditions whereby other open models inherit this precaution through distillation.
It's still not clear to me how this whole distillation thing is supposed to work, especially when Chinese labs then ship more cost effective models than American labs. Why isn't Claude Haiku 4.5 good? Why isn't Sonnet 5 better? How are Chinese labs better at distillation than labs that have access to the models? Why aren't more American labs doing it, and better? It doesn't seem to be illegal.
My biggest question is: why do we (as of everyone, not of Chinese labs) need an Askell role in the first place? I can totally understand Anthropic believing that she can teach Claude to be "good", and I think it is good to have the role. However, given that these important models from the US and China are going to influence, whether drastically or subtly, how billions in this world think and act, I am scared about having a handful of "philosophers" selected by these labs to determine the direction that is so critical to shaping the world at large. It seems to me more like another round of cultural colonialism (regardless of where and initiated by whom) rather than a net benefit thing.
I thought a lot about this. I think the honest answer is “because it makes better models”. Claude feels more polished than almost any other language model. I can’t say for certain that it’s the role of the ASCLs team that makes it feel that way, but the eloquence of her thinking and the model’s output seem connected.
I agree claude is very good in terms of capability, but its philosophy & "personality" may not look good for everyone. Ultimately, I am not against having one more prominent philosopher in the world, but I am scared of having ONLY one school of thought. I am not saying that Chinese labs should hire a Confucius or Marxist philosopher (would be a nightmare), but feeding only Askell's interview to every good and widely-used model seems a very uncomfortable idea...
Well, the alternative isn’t everyone having a part in it - the alternative is letting the benchmarks labs are optimizing against do all the character training, which will be even more colonialist.
Chinese tech forums and Bilibili comment sections have a pattern worth noting: the most common criticism of domestic AI models is not benchmark performance but what gets called 没个性 -- no personality. Chinese users actively complain about blandness more than Western users do. The demand for character exists. The supply question is whether it can survive a benchmark cycle where moving from flagship 3.5 to 3.6 to 3.7 in three consecutive months is the dominant competitive signal -- and where the character investment has no clear delta to show for it.
so interesting!! re: Doubao, my somewhat flippant opinion is that it might represent a different race to the bottom on character - the cute-little-girl thing is frankly very uncomfortable, but sells.
I don’t think I’m the right person to write about this, but I do think it’s super interesting and deserves a lot more thought than I’ve seen so far.