Ai2研究员考察中国AI实验室：算力短缺挤压中国模型长跑上限

Ai2研究员考察中国AI实验室：算力短缺挤压中国模型长跑上限 美中大模型技术代差预计在未来一年扩大至一年左右，算力与高端专业数据的双重匮乏将成为限制中国实验室冲刺前沿的硬性天花板。 Nathan Lambert在考察了阿里巴巴、月之暗面、智谱、美团、小米等中国AI团队后，分享了中美AI生态的深层差异。中国AI圈呈现出极度年轻化的人才梯队、极强的实用主义与工程执行力，以及以开放模型打响全球名片的战略路径；但与此同时，严重的GPU算力瓶颈与高技能数据生态的匮乏，仍制约着其冲刺最前沿模型的上限。与美国AI圈对未来的终极生存恐慌不同，中国团队更为务实接地气，将核心焦虑集中于算力与技术掌控力上。 中美在 GPU 算力上差距巨大，Meta 和 OpenAI 的拥有量超过中国 AI 公司的总和；但自 2025 年夏天起，全球开源模型主导权显著转移至中国，阿里通义千问（Qwen）已超越 LLaMA 成为全球研究人员最主要采用的基础模型。 (40:00) 中国 AI 实验室展现出极度年轻化的特点，领头人多为二三十岁且无传统学术定式；在强烈的“技术所有权心态”驱动下，美团、小米等非纯 AI 大厂普遍选择全栈自研基座模型而非外购 API。 (22:30) 研发节奏上，中国实验室倾向在强化学习（RL）完成后迅速发布模型，缩短了实际体验差距；而在心理状态上，中国团队深陷硬件层面的“算力焦虑”，美国科技圈则更多笼罩在对岗位被替代与 AGI 颠覆社会的“社会性焦虑”中。 (55:00) 受限于算力供给，国产芯片在推理侧已被智谱等公司大规模采用，但训练端仍高度依赖英伟达供应链；此外，中国严重缺乏高技能专业人士标注的高阶数据采购市场，成为限制 Agent 发展的致命短板。 (71:00)

艾伦人工智能研究所（Ai2）大模型后训练负责人 Nathan Lambert 在考察中国顶级 AI 实验室后指出，中国团队正凭借年轻的人才结构、极度务实的工程执行力和开源生态快速追赶，但依然面临顶尖算力匮乏与专业数据供应链薄弱的严峻瓶颈。 年轻人才结构与敏捷工程执行力 Nathan 观察到，中国 AI 实验室最鲜明的特征是人才极度年轻化，团队中大量充斥着二三十岁甚至在读的学生核心骨干（如月之暗面《Attention Residues》论文中的 17 岁高中生作者、智谱 20 岁出头的负责人及小米的在读博士等）。这批年轻研究人员没有传统深度学习的思维定式，能够毫不迟疑地直接迭代最先进的技术架构。清华大学等学术机构与周边创业公司之间的人才流动也显著比美国顺畅。在工程文化上，中国团队展现出极强的执行效率。不同于美国实验室在强化学习（RL）训练完成后往往需要数周进行安全评估与发布流程，中国实验室通常在完成 RL 训练后一天内即可上线模型。这种超快迭代周期大幅压缩了终端用户在实际体验中感受到的技术代差，尽管综合基准测试上中美仍维持约 6 到 9 个月的差距。 技术所有权心态与全栈自研逻辑 (20:00) 与美国依赖购买 API 或在云中间层协作的模式不同，美团、小米等非纯 AI 巨头均选择自研基座模型。美团通过开源通用基座吸引开发者，同时在内部利用独家业务数据微调不对外公开的专属 Agent；小米不仅将自研语言模型迭代至 MiMo v2.5，还布局了机器人与车载系统。Nathan 将这种现象归结为中国特有的“技术所有权心态”：企业深信依赖外部服务无法在激烈的长跑中战胜对手，只要数亿美元的训练成本可承受，便倾向于垂直整合而非外购。这种超级 App 生态推动了巨头内部正向商业飞轮的运转，但也对月之暗面、智谱 AI 和 MiniMax 等独立初创团队构成了生存空间挤压。此外，西方由 CoreWeave、Fireworks、Together 等构成的繁荣中间层生态在中国基本缺位，企业极少将推理或微调外包。 开源领导权转移与真实生态反馈 (40:00) Nathan 认为，自 2025 年夏季 GLM-4.5 和 Kimi K2 发布以来，美国实际上已将开源模型的主导权让渡给中国团队。以阿里通义千问（Qwen）为代表的中小尺寸模型（1B–30B）已在全球研究界取代 Meta 的 Llama 成为通用基座，同时阿里云对顶配旗舰模型（Qwen-Max 等）保留闭源以拉动基础设施消费。对中国实验室而言，缺乏闭源 API 垄断地位反而倒逼其通过高频开源维持全球影响力。这种密集的模型交付让中国团队积累了宝贵的隐形经验——能够深入理解全球开发者（如 Cursor 等应用）在真实场景下如何使用其模型。在当前开源前沿，Kimi K2.6 与 GLM-5.1 等模型展现出极高竞争力；而在美国，Meta 等巨头战线收缩，唯有英伟达（依靠 Nemotron-3 Ultra 等进行数据合成与蒸馏）仍具备重夺开源霸主地位的动机与能力。 算力焦虑与数据产业链的硬性制约 (65:00) 算力严重短缺是中国实验室面临的最大生存枷锁。据传 OpenAI 单个研究员拥有的显卡资源（约 5,000 张）相当于中国一家顶尖开源机构的全部算力预算，这导致国内团队容错率极低，一次长达数月的预训练失败可能直接压垮产品线。硬件方面呈现出“推理尝试国产化、训练依赖英伟达”的格局，智谱等头部公司每日处理数万亿 Token 的推理负载倒逼其引入华为昇腾等芯片，但国产硬件在训练大规模模型与软件生态配套上仍难以替代英伟达。更致命的隐形短板在于数据供应链：美国前沿实验室拥有数十亿美元预算向 Scale AI、Surge 等专业第三方供应商采购由医生、律师、金融专家构建的高技能环境数据；而中国实验室几乎没有外部采购高质量专业数据的生态，过度依赖内部团队搭建，导致研究人员不得不亲自模拟复杂业务场景，严重制约了 Agent 模型的深层进化。 中美认知温差与中长期产业分化 (80:00) 中美两地在对待 AI 的心理机制上存在巨大温差。硅谷充斥着对 AGI 毁灭人类、奇点临近以及人类沦为“永久底层”的哲学讨论与失业恐慌；中国研究员与大众则习惯了科技的高速剧变，心态更为务实，关注重点放在工程落地、系统安全与网络防护上，极少陷入存在主义危机。针对未来走势，Nathan 预测在未来 12 个月内，全球推理需求与 Agent 浪潮将持续高涨，DeepSeek V5、Kimi K3、GLM-6 等新模型会陆续推出；但中美顶尖模型的差距可能因前沿探索壁垒拉大至 6 到 12 个月。若将时间拉长至 3 到 5 年，AI 行业或将经历剧烈挤水分与市场矫正，部分面临高昂算力成本与融资压力的中国头部实验室可能陷入破产或并购危机。 中美 AI 竞赛呈现出“人力与工程执行力”对决“超级资本、顶尖算力与专业数据链”的复杂态势。随着地缘政治与芯片管制持续拉扯，中国依靠工程师红利和敏捷开源走出了一条特色路径，但缺乏商业化数据服务支撑以及底层算力天花板，仍是决定其能否最终冲刺全球最前沿（Frontier）的核心不确定因素。

[22:30] Nathan Lambert: Chinese capitalism is more intense than the U.S. because they're thinking of even longer term of what gives them the best margins... You don't beat somebody if you're relying on their service. And if you're hyper-competitive, you have to do it all the best and do it all yourself. [30:00] Nathan Lambert: In China, they don't have the infrastructure or the optimization layer... In order to support a company that operates in these middle layers, you have to be okay with offloading an important thing... And that just seems like a more Western startup style of things. [40:00] Nathan Lambert: If you put a gun to my head, I would say Kimi and ZAI as of today are the best open models in the world. But the competition is intense, right? That changes very fine margins. [55:00] Nathan Lambert: It doesn't seem like any of these cultural showing points has landed in China... They're like, 'I want to build this thing and figure it out later.' [80:00] Nathan Lambert: 12 months is still such a short timeline where we're seeing such a rapid rise in demand... Where if you ask me like three to five years, I'd be like... One of the major Chinese labs might have gone under due to finance... there's more corrections in the US market. [81:40] Nathan Lambert: I think that is core AI2's deepest culture is so scientific and like has this community value of building a lot of just like net good for the scientific community... [85:00] Nathan Lambert: In the Bay Area, AI feels like it's the only conversation right now... Like, I'm bought in as well. But you have to go back to stay in touch with what's actually happening.

[0:00] 美国AI研究员是怎么看中国AI模型发展的2026年4月一位在美国AI圈极具影响力的研究者专程来到中国走进北京、杭州和上海的AI实验室他密集拜访了阿里巴巴、玉智岸面、质朴、清华、美团、小米、蚂蚁和灵异万物等等团队回到美国之后他写了一篇文章在美国科技界是引发了不小的讨论那这位美国AI研究者就是Mathan Lambert他的观察也非常有意思谁有最好的中国的研究目标如果我把我头发挥了我会说"XXX" 现在的"XXX"最大的残忌是中国人仍然有现在的问题是的,是ComputeMeta和OpenAI有更多GPUs than all these Chinese companies combined如何是公共的车车车,像是Huawei chips所有的公司都说"如果我给你Huawei chips,我们可能会更多货帮"而大陆总监,Daryl Amaday在美国AI界,Nathan是一个很难绕开的名字。他参与组建了Hugging Face的RLHF研究团队,后来担任艾伦人工智能研究院AI2的大模型后训练负责人,参与打造了Omo和Tulu等开源模型。同时,Nathan也是Lex Freeman播客里的熟面孔。2025年DeepSeek引爆全球讨论之后Lex邀请了Nathan和Semi Analysis的创始人Dylan Patel进行了一场超过五个小时的深度对谈2026年Nathan再次登上Lex的节目用四个多小时盘点全球AI的最新进展Lex称他是AI社区中广受尊敬的研究者工程师和技术传播者之一也对全球开源模型有着紧密的追踪一个长期站在美国开源模型最前线的人真正走进中国AI实验室之后看到了什么中国和美国的大模型研究分别在发生什么样的变化这一期硅谷101视频播客我们飞到了西雅图与Nathan Lambert聊了聊他的中国型一路上的所见所闻他眼中的中美AI发展意统以及全球大模型竞赛的下一站同时不要忘记关注硅谷101不要错过AI最前沿的一线内容哦那以下就是我跟Nathan的专访用全英文进行Hello Nathan, Welcome to Valley101谢谢你来自我来的你回到中国的旅行,当时你曾经经过很多中国的主要AI的艺术告诉我们更多的旅行是一种有趣的一个系统的设计是有很多关注在美国的美国的美国人的关注在美国的组织的组织的组织的组织, 这些组织的织织的织织是在中国的织织的织织。这些社会的科技的技术的织织是在中国的织织的织织。我不是不容易合作的,但却成了融合我只是很兴奋的,我认为一些这些模式的模式是这么高的,但是当你学习更多的技术,它变成了一些比较稳定的。我自己的原因是因为AI是因为我会觉得在这些年代的未来,有很多危机的危机,特别的危机,如何人们是通知的。也许多,在这些危机的危机,它是当然一个问题的。所以,我们建立的区别人和人们的区别人,因为他们不做这些事情,不做这些事情,不做这些事情,不做这些事情,不做这些事情。或是极度的方式。建立这些连接是很重要的,一群人的一群人面对人们的背后好多天你去哪里?在哪里?我认为是有七个月的时间, 它是在一天的时间, 它是在一天的时间,所有的领域工程中的大陆,你不需要做任何的 [5:00] Bay Area feeling. And then we also went to Hangzhou for a few nights, which is where Alibaba is and I think DeepSeek is. We didn't go to the DeepSeek office. We talked to some people. and we did visit Alibaba. It's just great. my first time in China and nice to see them. So which labs or companies did you visit this time? Yeah, I'll try to list many of them. We visited also a bunch of robotics companies which I didn't focus on as much as writing but for language models we've visited Well, I went to two different Alibaba offices. I got off the plane in Beijing and immediately went to meet some people in their Beijing office. So there's Alibaba, Kimi, which is Moonshot as well. ZAI and then Florian who writes interconnects with me. We went to Meituan which has their language modeling effort. and we also visited Tsinghua University and we visited the Xiaomi campus in Beijing. And then when we went to Hangzhou, we visited the Alibaba campus there, the Antling team, and also Modelscope, which is also affiliated with Alibaba. This was a group trip, a part of mostly organized around this like SAIL, which is a bunch of sub stackers that write about AI, which is like my friends making a vehicle to do interesting things in AI, media and events. So I was mostly organized around this. And then the rest of the group in Shanghai talked to, I think, definitely Minimax and maybe Stepfun as well. So there's really a lot of the names. I think there's obvious missions like DeepSeek is the most secretive one and ByteDance is also like you visit them and they're like all the companies talk about DeepSeek as their like kind of mysterious leader. And then ByteDance is like, oh, the one closed AI lab that has a ton of resources that everybody's like worried about. The Dubow chatbot is the dominant thing. You pick up on these things pretty quickly when you're there because Everybody talks about it. So in your Substack article, you said you return with humility. What does it mean? Yeah, so I feel like As an American scientist, I never really expected to have anything to say about the Chinese tech ecosystem or U.S.-China interactions at a global level. I described it as like I knew I didn't know a lot and I left and I was like, I still have so much to see. It's like I saw these two cities in China and I was on like a busy work trip and feeling like you get teleported around in a van. We went to some great dinners and we saw some cute neighborhoods in Beijing and some of the older buildings and stuff. But it's like I feel like I don't know... so much about China and to like truly have immediate instincts and intuitions about something, you kind of need some more closer to full stack knowledge and appreciation for like what would a person's instinct here be, like what would a Chinese researcher's instinct be. And it's like it would take a lifetime for me to fully understand that as an American is things like how is the education system different. We went to like Xinhua and they would talk about like we want to They would acknowledge the reputation where it's like Chinese researchers haven't been as good at this like zero to one style science, which is kind of an arbitrary term. It's like, what does that actually mean? But they would bring it up. And part of this is in how people are trained and how the talent gets funneled through different layers from a much broader domestic network than in the US. These are things that people would talk about and actually seeing them is useful. This year I read Breakneck and Apple in China as these two wonderful books on the US tech ecosystem's interactions with China. But let's still like, a very limited view. So that's why I'm like, okay, I kind of expected this when I left, that I would be like, I don't really understand everything I left, but I still feel like I have more to learn. Were you meeting with the C-level C-suite people or the middle-level AI researcher or engineer-level people? Do you think those conversations actually helpful? Yeah, so we were targeting meeting with more researchers just to get a dialogue. We also targeted trying to get leadership, especially at startups. So it seems like mostly starting to talking to researchers. We talked to some executives, say like Alibaba or... some of these people that we talk to are very well known within companies and interface directly with the founders. But I think that when you're going to do like a formal visit. It's a lot of showing respect to the people that are there. But at the end of the day, leaders are more media trained and sometimes will tell you less than the kind of suppresses nuance to stay on the company line. So I think it was some of both the bigger companies I think were more open to show some of their senior cloud people without saying, without giving too much away. I think some of the conversations were off the record. I don't know, it's not the best answer, but it's a mix. Especially with the comms people around. Yeah, like most of these have comms handlers in the room. Like some of the best conversations were I announce that I'm here, I'm on WeChat, I message people, and then you go get Bobo with a random researcher not on the corporate office, and you're just like, What are you working on? What are you doing? And then it's like, okay, there's not the veneer of, like, I'm at the office. My manager is sitting next to me. The comms people are sitting next to me. So, Those are ones more fun, but more rare of the trip. Definitely. So what kind of questions do you [10:00] You mostly ask them. There's a distribution, so I think that I was interested in to start with something that's probably pretty accessible to them, which is like, why are you releasing these models? Like, why do you build your model? Why don't you just use DeepSeq's model? Right. A lot of these companies are very practical and it's like we need AI to serve our users and we want to own the full stack and therefore we're going to build it. And they're much more daunted about building AI models being this super expensive, impossible thing, where it's kind of an allure in the US of like, being a frontier lab is restricted to an elite few, but it's like all the companies. That's why it was like we wanted to go visit Meituan. It's one of our favorite random Chinese tech companies to have big models. So we're like, just go see what they're doing. They're like, yeah, we need this to serve our agentic products. So we build the model. We need it for products. It doesn't matter if it's open. We'll release it as well and get some feedback. And they're very practical about this open science. sort of idea. And this would... lead into different things like What are your current bottlenecks, which can obviously quickly be a dead end on compute or but a lot of them are like Huawei chips are good for inference, we use Huawei for inference, not really for trainings like a lot of people would say things like this. Also in our group people would want to talk about like what is the coming future of AI? So like, what do you view on how it's gonna impact labor markets and economics and like, do you feel like you have a responsibility as a scientist to understand, like think about these impacts. And that's where we saw in China that researchers are much less inclined to share a lot of views on this. Some of this is the corporate environment, but also I think it's just like environment that's much more structured than what they're encouraged to talk about in their education, which is like I build the thing. and there's much less of this podcast ecosystem in China where there's researchers that are rewarded for being visible and trying to sway the direction of the company. something a bit more practical minded, which I think I sometimes in my writing got overlapped with the humility, where it's like you can be a humble person, but also you could be in an environment where you're just not used to being asked certain questions. Cool. Let's talk about some of the companies you visited. In your opinion, how different were they from one another? Let's talk about Moonshot first, shall we? Yeah. do have pretty distinct feelings when you could go into them. I think Kimi was one where it's like These people seem so close and they're just in it for kind of the thrill of building a cool thing. And they're building a business alongside building this cool thing. But they're also just like, we're going to build these models. We use them ourselves. and they stay very close in a way that's like you just pick up intuitively when you go to the office. You're like, okay, these people are really bonded and they're tight and there just spend their lives together building this. I describe Kimmy as the best vibes company of the one that we visited, which is just like They seem like they're happy and they're going on their path. And it's not very corporate in a way where if you go to, obviously, Alibaba, it's like, this is a huge tech company. It's very corporate. The Ant Group, Ant Ling team has some of that because they're also a huge tech company. But even in some of the middle, it's like ZAI was very AGI forwards. and they had a lot of pride, but at least in the interactions we had, it wasn't as much of this like charm that Kimmy and Moonshot have. And some of this might just be like random day-to-day, like I met these people for a few hours over lunch. You can't conclude a company's trajectory based off this, but it's also interesting because it's people are expressive and you can learn things from that. There was a paper from Moonshot named Attention Residues with one of its authors being a 17-year-old high school student. You know about that one. I think that paper was reposted by Elon Musk saying impressive work from Kimmy. So this high school student got very famous in China now. So what does that tell you about the culture at Moonshot? It's like less of a moonshot specific thing, but the talent there is very young. These companies are all on a push to have more of a Western media footprint. And there are a lot of these researchers are on X. And it's like I knew a few of the people I think that it's like... people that do a mix of product and comms at the company. One is like Crystal, I think, at Kimi, and then Lu at ZAI, and it's like I met Lu, I don't know if I'm pronouncing her name exactly right, but she's like 20 and she's still a student. She's like, "Hi!" I'm like, "Wow!" These people are so young. And then one of the Xiaomi researchers we met Lai Li is also a PhD student and one of the lead researchers at Xiaomi. There are some people like this in the US, which are just extremely motivated, end up at a frontier lab, but it just seems like we're getting so many. one after another in China, these super young researchers deeply embedded in this lab. It felt like a bit of a structural, just, they absorb this talent much more quickly than other labs and are more open about it. It's wild. I mean, like, I'm in my early 30s, and these people are so young. It's like, I feel young in my life. Yeah. It's cool to see. But is it different here in the U.S.? [15:00] I think so. I think it is younger. You don't see that young talent here in US labs. Yeah, so something that defines AI2 a bit. It's like secret sauce is that like UW is behind you and we have a lot of PhD students that have been like core to almost and core to other major projects who are concurrent students. But that sort of open doors, at least like less fluid at some of the other labs. Like I know a handful of PhD students that work at labs, but these are like the absolute best of the best of the US, where in China it feels like, the intermingling of academia, especially at a place like Tsinghua, and these startups nearby is much more intentional and much higher overlap for the talent. Why does that matter? Students are young, they're hungry, they've have nothing else going on. They have much less going on in their life. Like I'm getting married this year, I think about family, like I'm 30, I have other things. But students are like, this is the technology of my life, I'm just gonna work all the time and absorb all the context. And when you build a language model, Like I work in post training, you need to know what the pre-training team is doing, you need to know the multiple stages of your recipe. and then you need to figure out what is the bottleneck and go address it. So there's a really big fixed context of technical knowledge that you need. And I think it's just like if it's the biggest thing in your life and you don't have a lot of distractions, it's just easier to put all this into your brain. One of the researchers told us like, oh, yes, new students don't have priors for what used to work in deep learning. They're not looking at it. a different architecture or anything. They're just like, "I want to know what the current state of the art is and make it better." And it's kind of like a simple approach that they can do. Some of the researchers in China told us this, they're like, "Yeah, that's what we're doing." I see, I see. Let's talk about Zhipu. How was the trip there? It was fun. I think he like So part of the thing that I learned is that Chinese companies have like a showroom culture, where when you have visitors, you like show the history of your company. showroom is all, like every panel has some different like AGI features. So it's like AGI loading or it's like their thing is that AGI, we're like 42% of the way here, which is from the Hitchhiker's Guide to the Galaxy reference. You just like immediately notice it. You're like, whoa. And But at the other time, they're I felt like I didn't get to know them quite as much as the Kimmy people. but they're very proud of their work and they're very successful and I think it's some things have like being public and being randomly entity listed by the US. They're just like immediately vaulted to a "Whoa, you're like a real big company status" type of thing that Kimmy hasn't quite hit that level of. But in my mind they're so similar. It's like when I look at the models they have different trade-offs. And it'll be interesting to see if that kind of diverges the cultures a bit more, if Kimmy stays private longer and Jifu staying public. Being listed on the secondary market probably... sometimes make a difference, just normally get more attention. Yeah, so Jipoo and Minimax are... public now and it's good because they can raise short-term, but it's like what if they need more capital to raise models? It's harder to do like you don't really do like a second public offering. It's kind of unclear. I And right now the markets are really rewarding any company that's public, so there's so much demand to invest in it, but that could turn... So right now it's good. I don't know. I don't make those decisions, but you can make arguments both ways. Earlier this year, there was a surge of API requests on OpenRouter for KME as well as JPU during the open call sensation. Do you think that trend will continue? Yeah, we're seeing kind of multiple markets emerge. You have the cloud code and codex, which is like the top end. I describe it as like knowledge work tools. But essentially there's going to be agents at price points for wherever they work. The thing that makes open models nice is that the price point is so low that the experimentation is much more doable. It's kind of a natural place where a lot of people will experiment on these models that are known to be strong and a step down in price while still having these established brands, which is where this whole deep sea, Jebu, Kimmy, Quen kind of circle lands. Personally, I think Open Claw itself is a bit of a fad, where the name, but the type of agent it represents will be sticking. So right now, Hermes agent is far more popular, but even this will come and go and there will be another, but it's just like... this type of agent that you can run it and it is not super expensive on these like hosted models on Open Router is is interesting. Open router itself is not a huge proportion of the actual inference market. So like in the US like fireworks and together are much bigger. They both started sharing some numbers on the amount of tokens they're doing every day. But Open Router is like more of like community oriented. So when you think about people that are following open models, that's like it was where they think and That obviously has an advantage and I think companies are very actively leaning into it because when their models surge on it, they use [20:00] It's like good co-branding and it feeds into itself. And let's talk about your trip to MainTuan and Xiaomi. Yeah, it's interesting, as we discussed, the story is very different. China is almost like Uber and Apple are developing their own open source models, but in the U.S. they are not. Why is this the case in China, though? When you ask Meituan, Meituan and Jaume's corporate position would be like, we're going to have agents that are used. you can think of it in Meituan, it's like an agent can do something for a user to figure out the ride to request or thing to order. Like it can do this. and the first instinct in the U.S. consumer would be, I'm just going to get an API key. Exactly. But in China, it's very much like, we can build this and it will be cheaper. Like, we have compute, we have talent, why don't we just try to build this? And... This isn't specific to Meituan, but there's a few companies that are like, we build these general models. and that's not where our use of them ends, but we can release the general model because more people will use it. And then what they do is they fine tune this to make specialized agents and specialized models, which they use for their services internally. And Meituan's a good example where it's like those specialized models they don't release, but they hope to use them in their apps. And you can imagine this with Xiaomi where it's like it can power features in the car, and I'm guessing that Xiaomi works like a Tesla or something, where it communicates with the cloud and there's all these interfaces and You don't really know what this is B, but AI is going to be intermingled with like all digital technologies. Yeah, I think Xiaomi is another good example of Big Corp. It's trying to do everything. It is doing like OLMs and robotics models, robotics hardwares. It is really like trying to do every single thing itself. Yeah, Xiaomi's models are pretty new too. They released a 7B, their original MIMO model in... April to June of 2025 and then come December when they had their /v2 model which was very popular. and then I think 2.5 was recently. It takes a long time for people to realize that these models are going to keep coming. And like, it's definitely surprising to see a new lab come from scratch. And they hired, I don't remember her first name, but like Fuji from Deep Sea. And it's like, yeah, and they seem on track to be one of the best model builders from China. People in the US are mostly still at like Deepsea. - Yeah. can we and ZAI. Some know of Minimax, some probably don't. It's such a big list, but I expect Having visited Xiaomi, I expect them to continue. It's very no-nonsense. It makes sense for them to do this. They've done so many incredible things in the past. It's like building an AI model almost seems easier than building some of the cars and stuff that they have built. I think for Meituan, they also invested in Moonshot's latest fundraising, which indicates Meituan is betting big on AI. and they love to build everything from scratch themselves. Do you think that's the right approach to go during the AI era? It's worth the link. try your hand at it. So I think AI, because it's so resource intensive, by its nature will benefit very large companies and companies that can get onto this flywheel of like we have a lot of resources. AI makes our products better. We get more returns on this. We can keep investing. We'll win so much. I think in the US it's really obvious to see like Google is winning. Open AI Anthropic will be fine. Like NVIDIA obviously like all these companies that are in this feedback loop are going to be so good. And then It's interesting to see how this works. in China, I think like Ant Group is along this way where they're building their models and they see how their platform spreads. Like right now it's spreading into health and they talked about this a lot and they will surely also use their models there. And it's like for them I think this actually makes a lot of sense but when you look at a Western observer you're like why is antling a thing but they don't realize the like broad service area that exists when you're using this Alipay thing. I think for these companies with very large footprints, it actually does make sense. worry that it suppresses some of this competition. So it's like, how did the smaller Kimi's mini maxes and ZAIs like fit into this ecosystem where such big apps exist. I forget which one of the companies, but they were telling us how they built an agent to navigate one of ByteDance's apps and then ByteDance banned them because they couldn't take the data. And it's like, there's... interfaces emerging between the small and the bigger companies or the small companies are trying to find their niche. And I feel like Meituan's in the middle where I don't have a good grasp from my perspective, like how big of a tech company it is and it's like, is it going to get swallowed by consolidation or is the spend going to be too much? [25:00] The language models, I don't really know, but I'm pretty bullish on AI as a business. So I feel like most of these companies that are early are going to be okay and sell. They sell a lot of tokens, they sell intelligence, and there's going to be a lot of demand for it. You wrote that Chinese companies have a technology ownership mentality. Can you explain that a little bit? This goes to some of the super app ideas and They just see that they can build things and make it better. And if they succeed, they'll put more money into it and keep making it more efficient and they'll tune it to their use cases. and... This is how many domestic industries in China have grown over time to be some of the largest in the world. This is-- how it's kind of like they want to build rather than buy if at all possible so they can kind of own the long-term trajectory. It's like a very... I think it's like Dan Wong of Breakneck said that China's capitalism is more intense than the U.S. because it's like they're thinking of even longer term of what gives them the best margins and the ability to provide products and beat competitors. And you don't beat somebody if you're relying on their service. And if you're hyper competitive, you have to do it all. the best and do it all yourself. That's why I think of it as this ownership mentality of where it's not currently an impossible line item. I would say that they're probably spending 500 million to billions a year to make these models, which is a substantial amount of money, but not impossible for these labs. It seems likely that the market will consolidate eventually, if progress on capabilities keeps going up and like capabilities is like log of investment where it's like you have diminishing returns but if you invest more it gets smarter. But it was an interesting time where it's like, I remember I interviewed Richard from Antling and he was like, yeah, it took us six months to set up this team to build these large models. That's because there's all this talent in tech already and they just like move the teams to do this. We're in the phase of like, I don't know how long they will keep doing it. But for now, it's kind of a fun thing. Let's talk about some other of the companies you visited. So Quinn and Alibaba. It's a big corp. It's different from One Shot and other smaller labs. Yeah, so then Quinn has won the hearts and minds of open source AI by releasing these small models consistently with incredible performance at all these sizes. were doing this before Llama 4 and before Llama Faded, but they really hit their stride right around the same time as Llama 4 happened and meta paused. they know this. So we asked them, it's like they think about things that developers will use. They want to keep building models that developers use and thrive on. But it's like when you look at Quen, they haven't released their biggest models for a while. So it's like Quen Plus and Quen Max have been closed source. And this is very much of a like, the cloud side of Alibaba talking where... just like the US, like you see Google Cloud's crazy expenditures and crazy earnings. It's like Alibaba is the default in China for that to happen. And they know this. So like, they're not going to release a model, like they might tell us they might not release it because no one will use it. But also it's like, I'm gonna be like, look, you can make a lot of money on this. And I think there's a lot of money to be made in AI. But I think the most interesting thing is hearing various people talk about them being intentional about these small models. And it's like not an accident. Maybe it was at the beginning or like, oh, let's try to do this thing somebody else is doing. But they have a very refined recipe of these 1 to 10 to 30 B models that are very popular. And that's intentional. And it's kind of nice to just like hear how they're like, yes, we do this. We understand it's a time sink. So it's hard to release models. It takes a lot of work in the ecosystem. that stress of like, we could build more models, but if we build a model, we have to support it. We have to go to VLM, we have to go to all these companies in the open ecosystem and like, the modeling team is still pretty small. And it's like they don't have a way to outsource that work. somebody very close to building the model still does this. There's the two ends of caution. you can release too many models and also like They haven't been releasing the biggest models because they want to monetize them in the cloud. Part of going into China for this trip, one of the other questions was like, is AI investment kind of like a SaaS, like the software as a service thing, or closer to a cloud spend? Because one of the whether or not it's a trope or actually true, is that the views in the US is that China doesn't spend on these services, where it's like, the slack equivalent is much smaller and people maybe just like, rebuild minimal software in-house or download something and much less willing to pay this per head per seat monthly cost on software. But all these companies do run on the cloud. Alibaba has a large business and a lot of modern software you need. via the cloud. I think that China will shift to be more viewing that AI spend is like this cloud spend. It might not be as exorbitant as the spend in the U.S., but they'll view it as needed spend to make the technology that they want to make. So therefore, it's these clouds will exist. [30:00] It's so interesting because when you arrive at the Shanghai airport, or you take the high-speed train to Hangzhou, you will see the Alibaba Cloud commercial banner. Yeah, I walked into Beijing and the first thing I saw was like Alibaba Cloud, Quen something. I'm like, "Okay, here I am." I think one difference between Alibaba Cloud, and AWS is actually a bedrock carry. every single big model. and their own model and smaller models at the same time. For China, cloud, they only carry their own models. That's a big difference. And I think that this neocloud market in the US where there's just like a lot more people reselling other people's models as tokens or reselling compute is like a huge part of the US market. You can name so many companies. It's like CoreWeave, Nebius. together, fireworks, Lambda. There's going to be way more. We talked about Open Router. I don't know any of the equivalents. of those in China. I don't think so. Those companies are largely looked at as being supported by NVIDIA in a lot of ways. Like, NVIDIA likes to prioritize the diversity of the ecosystem because it's better for them to not have most of their revenue be from opening AI. Like, for those few companies. And it seems like that dynamic is not the same. That one is interesting because you are missing this layer. in China, they don't have the infrastructure or the optimization layer. Because you look at the likes of what a trend we're seeing now is like cursor took the Kimi model and they post-trained it within the Fireworks like RL sandbox training environment and serve it with Fireworks. And that's a really interesting multi-party... collaboration to build a We have a very strong model and product. which might have to be more verticalized in China to get the same thing done. - Probably, Chinese companies just Won't you have everything in-house? Yeah, it goes against the like the ownership and the integration where it's like in order to support a company that operates in these middle layers, you have to be okay with offloading an important thing, whether it's GPU hosting or inference, because you want to focus on something else. And that just seems like a more Western startup style of things than otherwise. I do think a lot of the model labs in China are going to be more willing to try different things than the bigger and older tech companies in China, because Kimmy knows Cursor trains their model. They will go talk to Cursor and stuff like this. And ZeroOne.ai, founded by Kaifu Lee, They now experienced a major transition to a total like 2B kind of business. model, where do you see the future for China's to be market? during the AIR. It goes back to if these relationships can kind of So I think a lot of them are wondering if they can use this global to business software market. I feel like it's so hard for me to make a bet domestically whether or not it'll win. Yes, you can take AI expertise and build it within another company and it'll be very valuable for the company that had less AI expertise, but it's like whether or not this willingness to like you're losing ownership if you're relying on an external partner to do this. It's like the latest trend in the US is like forward deployed engineers. This is buzzword FDEs. It's like opening eyes, hiring a ton and stuff like this where it's going to happen in the US. But it's like, is that going to be a viable thing where you can go through the front door of a company and actually get welcomed to them is very interesting. It's like they're gonna try this and I think that Like if it works, being the first one to retry it in the AI era is going to obviously be so useful, but I'm like, I have no idea. I could flip a coin, it depends entirely on kind of this like Chinese culture of the business that I can't fully understand. Yeah, but did you, whom did you meet with? We met with Kai Fuli. Oh, yeah, yeah. How was the conversation? It was good. He's, I mean, it's like such a legend to be him and it's so cool to be able to see people like that and he's like, focus. Like he could do, he's done so much in his life. He can do whatever he wants or not work. And it's like, to see him like trying to solve a problem and figure out how to reimagined his business is like a very AI type thing where it's like it's just drawn all these people back to like try to think about what the next phase of building businesses is going to look like. Why Kai-Fu's lab failed to become the leading lab in China? I don't really know. You should ask him. That's a tough question. I think there's a lot of opportunity for doing things with AI and building models is a different type of culture. the building models is very much like turning the crank and continuing to ratchet up the resource rungs. And I don't think it's just zero on AI that has built a model and then stops. I think [35:00] In the US, Databricks was building models and then there were like "No need for us to do this. Other people have got the bill." I don't know if they're on the record, but I think a lot of leadership at Databricks would tell you this. They're like, "We're happy to not pay for building this model and we're going to do our business." And I think that's why I'm surprised that so many companies have kept ratcheting, or doing this like 1 trillion parameter MOE, which is way more expensive. Like, Zero-One AI and Databricks got out at like small MOEs or small dense models. It's like, way less investment So in some ways it might seem smart if you thought that you were not going to be competitive. It's better to stop earlier than later. Not everyone needs to have or to build the best model, right? - Yeah. - A cliched question. How many months do you think US SOTA models are ahead of the Chinese models? And for that gap, is it going to become narrower? or wider, why? So on traditional benchmarks, it's been like six to nine months for a while in real world use. especially early in reasoning models. It was very tight because it was a new paradigm. And it's like when you're in a new paradigm, the rewards to doing it are so steep. Both U.S. and Chinese labs were climbing this really fast rate of improvement and the Chinese labs release sooner. So you talk to somebody at a Our model's still in RL. It takes about a week to do this last RL round, and then we'll release it within like a day of it being done, where the U.S. labs have... some process that normally takes weeks to get the model out. So if you think about it, you're on like steep progress you have two different lines and it's like if China's behind but you don't have this like no progress, horizontal is when you're waiting to release it. So China releases it faster so it kind of like converges the performance that users get. So at the end of the day, the performance gap is like the models that are available to people. and Through this conversation we've talked about a few things. Like some is like this knowledge work domain which is all this US workers and things like this which I think realistically the raw utility of the US frontier models that are just like a bit more robust in a way that's been hard to measure. So it's hard to know on the benchmark numbers. I think the six to nine months on the benchmarks that are popular will become stable, but there's this like intangible unmeasurable where the US models will pull ahead. The best example is this Claude Code and Codex Revolution that started in December and January, where it'll be very obvious if an open weight model enables this type of behavior, because everybody in the world will be talking about it, which is like, I can pay $1 a month instead of open AI's $200 a month. A lot of people are going to switch. as of recording that's say like five to six months and I just like don't really expect that really clear watershed this year, which will like shed some light on the gap and how the way we're looking at the gap is a bit narrow to how AI is used in a really broad set of use cases now. Why haven't we seen a really good Chinese coding model as of yet? I think a lot of What's popular with cloud coding codecs is that they're just very easy to use and I think the Chinese labs will keep solving coding, so like the actual code quality of the models will be incredibly high, and that's based on public data and things like GitHub and stuff like this, and you create puzzles the model learns to solve there. But it's in this like ease of use thing that the US labs had a really big lead and there's a difference in how they operate where the U S labs get consumer data. So like, Cursor trains their model on people using Cursor and ChatGPT and Codex and CloudCoder are the same. Well, all these open weight labs, they release a model and they have their CLIs that are a bit less popular. But a lot of the usage of the model just doesn't go back to the developer. So some of it is a bet of like that real-world data is really useful to kind of making the model more robust across general circumstances. And that's where the closed labs have a huge advantage. Cool. And I do have some open source model questions. Why are Chinese labs releasing so many strong open source models? it's a path to relevance. I think they know they're behind and they know they want to become more in the conversation. And if no one knows about you and you have an API, nobody thinks about your model. but in a world where AI models are so valuable and pretty rare when you release them openly. There's a lot of people that are like, "Oh, I wish I had my own AI model," that are going to take a look and try it. So all the labs are really practical. It's like, this is our way to be in the story of what is happening at the frontier of AI, and they will keep doing this. And I expect this to be stable 'cause I expect this fast follow gap. to continue. And this is why it's like a very practical mindset. It's like, There's not much added risk to us doing it. It helps us accomplish our goals. It's the only way to get some customers, We've got to do it. If Meta fails to become the best open model, does it mean that U.S. actually is losing open model leadership to China. In the last year, the US has ceded the open model leadership to China, I think especially in the summer of 2025 when a bunch of these new Chinese labs were starting to get really prominent. That's when [40:00] DLM 4.5 and Kimi K2 came out as their breakthrough models. And like from that point on, like the US no longer had open model leadership. And that's been the case, I think, since summer of 2025. Anyone claiming otherwise is a little bit diluted on where the models are coming from that actually matter. Does the U.S. still have a chance to catch up? They do because they have NVIDIA, and NVIDIA's financial interest is to not have AI be owned by any one company. And the best way to do this is to release open models for the world. So the only hope is NVIDIA. Realistically, potentially, like in open models. Like I think there's the likes of Meta and Microsoft, where they have products, whether user enterprise. that are supported by AI, so they could release open weight models, it wouldn't really hurt their businesses, but they chose not to in recent years. Maybe it's a distraction. Maybe it's like a marketing and security risk or something, but So long as that stays, I think NVIDIA is best positioned to change this as of recording, a couple days ago or yesterday and Jensen announced their Nematron 3 Ultra model. The way I think about this is like that's the completion of the first arc to NVIDIA starting to try open models again where they had these Nematron 3 Nano, Super, and Ultra, which is their like a mixture of expert models and ultra is the biggest one, which they will then use to make synthetic data and all sorts of smaller models and distillation in-house. So now that they have their first big model established, they have a lot more flexibility in building other sizes of models and continuing to improve it. And this is where it will be more interesting to see their open model efforts. It's like once somebody gets to this big MOE, can they keep kind of scaling it from there? Where NVIDIA had a lot of models in the past that are like, were good but kind of fell out of the usage or coverage cycle a bit more quickly. How about some smaller players like Reflection AI? Or maybe you guys. Yeah, so I think there will be many more people building models in the open, and the open ecosystem can support a diversity of models. So these smaller, medium-sized models are very useful. because they're so cheap and if you can get them to do one task repeatedly, you're going to save so much money over using Claude or Codex. why there's kind of a fixation in open models of chasing the frontier, which is these big MOE models, which is an Emotron Ultra, KBK2, and whatever. So I think that's kind of something where over time that fixation on the largest model will fade. So like reflection wants to build a big model. But it's like people competing with Quen that will win the hearts and minds, which is a small model that every single researcher in the world is building off of like the extent that. AI research right now is done on QN is incredible. It was LLAMA for a long time, but it's like everybody, like all the researchers know how QN works at that scale. They're used to using it. And that's like a really remarkable achievement for them. I think it's like not talked about enough. And I think over time that distribution from small to big will have more levels of interest and I think for reflection, it might be better to pick a niche that's less served. And AI2 builds small, although models, which are great for research. And it's like, I don't think AI2 should... chase the frontier and try to build a 1 trillion parameter model. I think philanthropically is not possible, but also it's just like good to do something. strategically different than There's time to fill different niches before people realize they can't financially keep up with the frontier. Who has the best China open source model? It goes back and forth. I personally want to spend more time playing with the various models. I would say that the best models are like Kimi and ZAI. So like Kimi K2.6 and GLM 5.1 are probably the ones that I hear most as like actually making a difference at the frontier of usage. But it ebbs so fast, like I think all of 2025 DeepSeek was the like unanimous king with the best models, particularly like their v3 and r1 lines. And I wouldn't be have been closed. that has held back some of their adoption in Mindshare at the biggest sizes. But I don't know. If you put a gun to my head, I would say Kimi and ZAI as of today are the best open models in the world. But the competition is intense, right? That changes very fine margins. You mentioned a lot of Chinese researchers in your article. Tell us about your impression on them. How are they different from the US AI labs? There's so much talent in China where you go to a lab and it's mostly like Chinese researchers grew up in China. They're often younger. We went to a lot of places near Tsinghua and there's other great universities. So it's like you go to the labs. For one, it's way more uniform in who is there. A lot of Chinese researchers, like, here at AI2 or anywhere, I'm like, I feel like I work with all sorts of people. And as somebody who is, like, not an immigrant or parents were an immigrant in the US, I'm like, I feel a little bit unusual. I work with people from all over the world, whether it's Europe, Middle East, China, other [45:00] - This is in Asia, and that's, in the US labs, that's also the case. There's a lot of Chinese researchers in the US. But that's like the population level is like, that's definitely a different feeling. there and then I think there's some of these things we talked about earlier where a lot of the like top US researchers are like talking about where AI will be going, this kind of philosophical or artistic design of models and I think it's an example that the organizer of the trip, Catherine Rintoul, pointed out is that like the likes of Amanda Askell that are talking about like model character and really crafting the tone of the models, it's like that whole conversation didn't really come up in China. It's like there's some pockets there where the topics they talk about or like... how far they're willing to project where AI is going is pretty different. Otherwise, it's so similar. I think the things that are more interesting, China specific, is like the newest generation of researchers in China seem even better at speaking English. So many of these really young people speak incredible English and they all follow the Western AI ecosystem on Twitter and read blogs and they're all writing these papers where the opposite is not true. Where it's like the US ecosystem, a lot of people just have no... Like they don't think about what's happening in China. They don't, like it's hard to get reliable information or what I feel like is of what's happening in the Chinese tech ecosystem. Even if somebody is like really interested in it and wants to know how these companies navigate. So there's these very different information flows that will feed into how people think. But also it's like a fairly American character trait to be like very proud of the stuff around you and like slightly beautiful. It's like a very American thing and it's like the Silicon Valley kind of like rewards people from doing this and it's like a bit tunnel vision in a way where the Silicon Valley is obviously a developing its own narrow culture and discussion around AI with the permanent underclass nonsense, and now it's the recursive self-improvement and the singularity. There's sprinkles of this when you visit the Chinese labs, but way less. They got part of it from the US, and they're like, "Oh, that's a little weird, but we'll be entertained by it." But when you go to SF, it's like, that's really what people talk about. Senior PhD students are like, "Oh, I need to make my money now. I need to join a lab." There's this very intense culture and SF around the timeliness of AI and how their career is changing that didn't feel the same in China where they're like we're gonna do this. maybe a bit more down to earth in that regard, less anxious about things. How do you view Chinese AI talent in the current wave? I mean, if you even look at the leading frontier AI labs in the US, you will see a lot of Chinese researchers, right? Chinese American researchers working there after completing their PhDs. They might be immigrants, they might be second generation immigrants. What kind of roles are Chinese talents actually playing in the AI research today? They're in every company playing crucial roles. I think the US current political stances to limit high-skill immigration is bad for the country. I work with a lot of Chinese or international people and I think that I want them to be able to live in the country where I like to live. And I disagree with the policy that is going on. And I think anyone that works in tech... knows how crucial this talent is and I think there's worries as, for example, there was recent news about exit bans on passports for some of the top executives at AI companies and people in the U.S. that are sensitive to how important this talent is and the rising geopolitical tension will worry, like, what if they call back the top researchers because the U.S. companies will be severely hurt? Like, that's the reality of it. I don't know if that call gets made at all in the near term. That is not remotely near the circles of people who listen to this podcast. Like, that's not who makes that decision. But, like, People talk about this stuff in the US as being a major potential bottleneck on the tech industry's progress. It looks like you only need one person to lead the team, right? Then you have that model to be better. I think the vision is hard. So an undersold thing about building language models is that so much of it is an org chart problem where you have talent and you want to make the model better. And the hard thing is organizing the talent and figuring out how to break down the big problems into little problems where a lot of but it could be the way that Chinese engineers are trained, that they are more suited to making these org charts, where it's just like they're just going to grind on their little thing and be happy that it contributes to the bigger model. I think there's a lot of uncertainty in this because like how the frontier labs actually work is not that well documented. And it might just be more of chaos where it's just like people running around trying to make the model all the time. It's like not the best analogy, but it's like, But definitely org chart limited, or you have to get all these people in. [50:00] to this one thing. Are Chinese AI researchers more willing to drop old assumptions and just absorb the latest paradigm. If you look at MOE, RL, agents, coding models, and tools use, for example, MOE wasn't invented by DeepSeq, but the company optimized the technology in an extremely effective way. I think this is something that's more of just with new and like youthful motivated researchers there's a very grim saying that's like science advances one funeral at a time which is like the most bleak way of saying it but this is just that like new people are much more willing to accept the newest things and they have no reason to defend their old ideas because they have no established track record and they want to dive in and take these things. And to the extent that China is potentially younger, Amen. both through engineering talent and leadership, like those people just have less priors. and therefore it might be true. It's something that if you're trained in science, you're just like, the younger you are, like that's where a lot of the new ideas and change comes from. - For Chinese culture, does it mean it is harder to cultivate someone who is truly visionary? For example, Ilya Suskever, that level of leading expert. That's the impression that I have, and the impression is built on, like, understanding that the culture is built on standardized testing from such a young age and some of these tests rewarding memorization. Students know that if they make this test they get to go to the best school at the next rung and they always are this like super competitive thinking about this kind of gate in their progress. And correct me if my understanding of the Chinese system is wrong, but that's my understanding where the US ecosystem has like less of these gates where it's like your fate is decided as better or worse and it's more of a like a continuum where A lot of the best people go to the best schools, and the second best schools are also very good, and they blend through each other. Yes, there's credentialism, but maybe a bit less extreme. So I think it's very possible that that is the case. And some of the leadership would, off the record, talk about this, like this rewards memorization. very hard to rework an education system, despite if you want to. Some tech leaders you talked to off the record said that that was the case. And I was like, okay, I have no reason to not believe you. You describe Chinese labs as almost perfectly set up to be fast followers. Can you explain that? What makes them execute so fast? Yeah, this is what I was talking about, the various things on... the young talent and the culture just like being very technically driven and focused on their task and potentially like they're just gonna lock in on this thing and they have abundant talent where it's like, you can just have people working across the whole stack and funnel it into a model. And we've seen this many times, like once you know that something is possible, it's a bit easier to do this. And We've seen Chinese tech economies come For other things, it's like you go into Xiaomi's technology, you go into it, it feels like an Apple store. very likely they're inspired by the iPad and iPhone and their designs and they're good at it. They've done this not just in AI and I'm like, "I expect them to do it in AI again." the debate is whether or not like them doing this fundamentals to keep redoing it is enough for them to like people like to wonder, it's like, oh, is China going to leapfrog the USAI labs? And I think in an equal resource world they might be able to or they're just like they have more talent that's willing to work more hours over a five to ten year time frame but i think human talent is only one of the resources that it takes to build these models and just like capital through mostly compute and data is actually the biggest lever and it seems like they're disproportionately fewer resources than the u.s hyperscale companies so it's like i think talent will only take them so far but the technology diffuses, where the best models make it easier to make. more efficient. like So it just kind of like, is this going to keep... propelling through there. - Can a fast follower become the frontier leader. They can. I think AI is weird because of the resource requirement. China's path to leadership is by it taking a long time to get to the final destination and that's time where Huawei builds out its capacity and then its net capacity becomes bigger. It's a common scenario you'll hear in people talking about balance of power and chips. So I think that most people think that It would be a long haul for that to happen, but don't rule it out. as we talk about the leading AI scientists' culture difference. you know, here in the US, right? But I think back to China, maybe Xiaomi is an exception, with Luo Fu Li being quite active on social media, as well as podcasts. There was also Lin Junyang from Quinn. He was the top lead of Quinn. He left the company earlier this year. Especially for Lin Junyang, I think there was apparent conflict between what's called the individualism and collectivism, which prioritizes the company's [55:00] Trust. per se. Do you see more of that kind of conflict coming out from the China labs? Going into this Western ecosystem, I think some of the researchers will find that they are very well suited to this. So I call him Justin, the American name. And it's like, he was so good at this. And I kind of see it as something we've also seen in America, where it's like people in tech companies that can get a lot of influence. externally as well can push against leadership and the big tech companies are like we don't need to listen to you so then like they part ways like we've seen it like we've seen it in the US I think a prominent example was around like some of the AI ethics researchers at Google where it's like Tim and it sent an ultimatum to Google leadership and they said no. And it's like, there's walls that you can hit within big tech companies that are hard to hit and I think that some of these Chinese researchers will probably get fed into this ecosystem and we might see some of it again. But for startups, the startups like Kimmy or ZA, I would love to have a person that is bringing noise and bringing attention. You just figure it all out later. It very much depends on the company size. But... I think it would be good. If more of the Chinese researchers get into this because it's good to have more views like this whole part of the world if they can be seen in the US so more people understand how they think and how they approach models. It would be good for the kind of global equilibrium. Yeah, that's great. We're trying to do that. We're trying to invite more Chinese researchers coming to our show. So subscribe our channel. When you talk to the Chinese researchers, right, what was the mode inside the Chinese AI labs? Did you see more confidence, anxiety, curiosity, or something more practical? - I think there's definitely an anxiety of not being able to keep up and compute. all the Chinese labs want more NVIDIA compute to train the latest models and the researchers are the people who feel this most. They're the people that are hitting the resource constraints. But otherwise, there's less... social anxiety. I think in the US there's a big upswell of social anxiety, whether it's people I work with here that are seriously concerned about their job being replaced. And I'm like, you're one of the best AI researchers I know. But like, even people like this would be like, I'm really, really concerned about this. And this is in the broader discourse where American CEOs are going around talking about how people are going to lose their jobs, which is not a good messaging decision. But that's very deeply embedded in how people think about AI now and the data center discussion. It doesn't seem like any of these cultural showing points has landed in China. Some of the developers are like, we ask them what is AGI and they're like, when they take my job. And they're not like, It's not like a cloud hanging over them. They're like, "I want to build this thing and figure it out later." Why is that? There's some hesitancy around developers, or like this is changing a lot. They all use Claude, which is supposedly banned in China. They all use it to do their work and they told us very forthcomingly, like we love Claude, Claude's the best model. So there's definitely some hesitancy growing around like software jobs are changing, but it's much less of a collective like social opinion on AI is bad and disruptive. That doesn't feel like it's established in China at all. I think that some of them discussed how China's been living through so many technological changes in rapid succession as the company makes so much progress technologically and economically, where they're like, this is like the next one where they've had so many with these like the Chinese mega apps. like they have electric cars now. They're just kind of like, oh, this is just how it is. Where in the US, it's like these changes have come with more cadence in between them. And then like the latest one, and there's kind of like more inertia that it's going to have to move aside. - Okay. Were people talking about alignment at all, safety issue at all within Chinese labs? The word that's used more is security. - Yeah. The company is definitely care about this, but it was much... less front of mind. It was like if we're releasing model weights we want to make sure that it's not enabling any gigantic new risk. And we would poke at, like, what about multimodal models? And we're like, yeah, some of these things have moved to be risks a bit faster, but it's much less of this big text LLM agent that we're releasing is a major risk and therefore we need to not release it. They would say reasonable things about this, but in the US it can almost sound like ideological deep rooted concern that like AI, like the next day I is going to be the one that is, of a lot of concern. So I think it's measured. but not very front of mind. Yeah, probably like Chinese are just being more practical. I felt like it was reasonable. That's kind of my stance. I support open models, but I'm not an absolutist. I don't think everyone [60:00] Model should necessarily be open. Claude Mythos is a bit of a marketing stunt, but also it's an incredible breakthrough model. It would not be good if that was just uploaded to Hugging Face the first day that they had it. It takes time for technology to diffuse. access diffusing over time for transformative things is reasonable and the equilibrium we're in right now where the closed models are ahead and open models follow it like seems like a good way to get technology into the world and I think that a lot of these companies would feel similarly. Which parts do you think Chinese labs are genuinely ahead of the US labs today and which parts are they not ahead but moving faster than Americans expect? So a lot of people say the Chinese labs are building more efficient models, and I think that's kind of overblown because the US labs are serving way more tokens at inference, so they're way more incentivized, like the financial gain to make... Claude or ChatGPT 1% more effective, it's like billions of dollars. I think they have like more reasons to be efficient. But I think the Chinese labs by doing it longer have a bit better understanding of how people are using their open models and We talk to people at Alibaba and they're very intentional about this, but also now the likes of Kimi and ZAI and all these others with these established MOEs that have multiple cycles. They're following how people are using their models and This expertise is very transient. I try to figure out how are people using Guadamo, and it's like a mess. It's really hard to get the direct information. But it's... is what makes their model successful or not. Obviously there's benchmarks, but then it's like, Why did Cursor choose the Kimi model? Kimi's gonna think about that. really deeply or ask them and this is like very valuable expertise at the frontier of AI that's hard to get so they like know what makes a bit better open model or not. And you only get this from releasing models a lot. So I think like Nvidia is building momentum on releasing models and they'll start to get more of this feedback and then there's reflection in the US that's like, we'll release a model when it's ready, but they're gonna get none of this feedback of people actually use the model and yes, this benchmark is good, but this very related thing doesn't work, why? And I think that's like the, intangible knowledge of how the open ecosystem works is consolidating there right now. Let's move to the opposite case. What are the biggest disadvantages that the Chinese lab still face today? So still GPU. Yeah, it's compute. There's likely incorrect estimates, but you could say something like Meta or OpenAI have more GPUs than all of these Chinese companies combined, and they very well likely do. And I've heard rumors that say an OpenAI researcher can get 5,000 GPUs for themselves, and this would be a major proportion. of a Chinese open model lab's research budget. So I think their aperture for risk and amount of dynamic things they can do algorithmically is much more narrow. And I think their source C, like just optimizing the current path of the models is much more of a necessity because they know they can keep making this better. But it's a less dynamic ecosystem. One of the Chinese labs told us like their latest pre-training run took like six months. And this is a model that's out on Hugging Face. And it's like if that run didn't work, like that whole line of models just like wouldn't exist. And that is a very precarious position to be for the company and I think that they're exceeding most people's expectations with the amount of compute that they have, but that is the limiting factor. And yes, the Chinese labs are desperate for more NVIDIA GPU chips. How much does Compute Scarcity change their model strategy? the biggest thing is that Every model You essentially design it based on how you expect it to be used downstream, and then what the shape of your training clusters is. So like the training cluster will configure how efficient it is to train a different size of model, and then you also know you need to serve it. So based on probably having smaller clusters, which is going to change how they do make these types of decisions or maybe they can change their base model architecture less frequently because pre-training runs take longer. And I do think that we're going to see for some of these smaller labs, I think Alibaba and ByteDance with the resources they have is a bit different, but like Minimax, Moonshot, etc., like most expectations are clawed in GPTs. are at this like three to six trillion parameter range. And I think it'll be a lot harder for them to scale up unless they're getting these latest NBL 72 racks and Blackwell chips and things like this. So essentially there's like, there's ceilings to these, but it's a multi-optimization problem, whether it's like, what's your training cluster, what's your inference stack, they might want it to work on Huawei inference now. Like, who are your users? So they do all of these things. And we see the output as a model on HuggingPace of like 600 billion to a trillion parameters. It's hard to map, but it's It's one of the most interesting decisions in a lab and most people won't tell you. Yeah. So does computer scarcity force better efficiency or does it ultimately cap the frontier? It does. [65:00] But it's more of like you're changing how you work. Everybody wants to compute efficiency because if you have a more efficient research, if you make your experimental unit half the price you run twice as many experiments. There's a lot of very basic math where every AI researcher wants this. It's hard to say that the Chinese labs would necessarily be better at it. It's like maybe because they have less inference compute, they train a model that's more efficient at inference, and that has a different trade-off. I think DeepSeek's famous for these memory-saving innovations, some of which are used everywhere now. But it like might just be overkill when you compare. Like a US lab might look at that and be like, "Eh, it would give us a small gain, but it's not worth the effort," or something. So how good are domestic accelerators, such as Huawei chips, in practice? I heard many people saying it's ready for inference but not for training. That's mostly the consensus. We would go to some more random companies, not like Core LLM, and they're like, "Yeah, we have to have Huawei, but we don't use it." But most of the LLM, language model companies we talked to, are established enough, whereas they all have services that people are signing up for. ZAI serves trillions of tokens a day on their API. And all of those companies are like, "If you can give me Huawei chips, we will maybe get more customers." And they're figuring out. how to use them for inference. But on training it's only like very small models that could be trained on Huawei and I personally expect that to be the status quo for a while. That's why researchers focused on training early. I would like more NVIDIA trips. A while is how long? I mean, ask a semiconductor expert, but I would expect order of years where it's like still NVIDIA is... so far ahead. Because NVIDIA is channeling the global supply chain into making the top-end products versus necessity is the mother of invention at Huawei. And it's like, how far did they get? I don't know. How do you see the recent debate between Jensen and Darkash? For people who are not very familiar with the context, Jensen argued against the strict limits on Chinese access to NVIDIA GPUs, saying that China already has the energy infrastructure, chip production capabilities, and AI talent to develop advanced models. Indipitably, he said that restricting sales only cuts NVIDIA off from a massive market while failing to stop China's technology process. was holding the China threat theory. And of course, Jensen is biased because he wants to sell more chips in China, right? But whom do you think is right? I think realistically I'm in the middle and more on Dwarkesh's side. It ends up being a geopolitical discussion, which is like, do you think AI is a technology that is going to be used for this? And it's like, I would rather see... What is? traditional American values of freedom of expression and all these things spread into the world. And I think that giving AI chips to China does pose some level of risk in this capacity, on the balance of power. As an American, that's how I've fallen on this. It's hard to know to what extent. So when you listen to people within NVIDIA, they essentially are saying there's no more power to build in the US, no one in the US would buy these chips, so we want to sell them to China. But at the same time, you have US companies asking for more allocation and US policymakers talking to constituents who are like, the US companies would have bought more. So I kind of talked to both parties, which is these US labs that are like, I want more chips to sell my companies. And I talked to policy researchers at NVIDIA where they're like, the US can't support more volume. So it's so hard to get the ground truth information of like how much would the US buy and what is the diverted to this Chinese market. This is why I end up with big error bars because I don't know... who exactly, like, what the equilibrium is and therefore there's some chips that NVIDIA definitely wouldn't sell in the US, and we should tell them China and support NVIDIA's AI leadership. So I'm like, I don't know. The question I'm more curious is that will the ban encourage better AI chips in China, which is Huawei chips, right? It's definitely the case. We talked earlier about this like software as a service versus cloud, and I think like AI is going in the cloud direction. and this is going to support all sorts of new companies in China that need this cloud in order to build their viable products. growing inference demand. We talk about OpenCloud, we talk about Cloud Code. The inference demand is there. China is earlier on the adoptions, you talk to developers in China and they're like, I use Claude. Like this is getting routed to what is Western capacity. And it's like as China allocks more domestic AI inference capacity in these use cases. this demand will all funnel into how many chips can Huawei make? This is coming in... not years, like there will be [70:00] New apps like Cloud Code that are so enabling in China that people have crazy demand for them. So I see this happening and this capacity Jensen. is somewhat right, where it's like that is what's going to drive the local AI stack to be developed. Like the local labs have Huawei chips that they're designing inference for. Like in some ways it's happening and it's just a ratio of of how much and to what extent is okay for people. - So the bottleneck is really on the production capacity for Huawei, right? Yeah, and the ecosystem for how easy it is to actually run a model. on these chips, which is like the cooldown level. So I think, like, I'm very... bullish on AI demand continuing to increase. So therefore, demand creates this need. Anthropic CEO Daryl Amadei has consistently positioned himself as one of the most vocal tech leaders warning about the national security and geopolitical threats posted by Chinese AI development. How do you see his approach? I think it's a little... too aggressive. Like I think that Some of anthropic media is pretty careful about separating the concerns from Chinese government and Chinese individual scientists. But it's a... a little too aggressive in the near term about like risks posed by open-way models and so on. So I think it could be a little up-putting to people to hear him kind of be the tip of the sphere in that regards. And you mentioned that China's weakness in the AI data industry is more important than people realize. Can you elaborate more on that? One of the most surprising things is that the data industry is much less developed in China. So when we flew, this whole environment's... so to say, where, where, companies like Anthropic and OpenAI are saying that they have billions of dollars of data budget to buy environments from external vendors. We asked the companies and they're like, "No, we don't really have this. We build some in-house. People would point to the likes of ByteDance and Alibaba being bigger and being able to build internal data teams. but there's very little talk of like you go and buy some data from an external vendor, which is so normal in the US. Like the Frontier Lab playbook If you take it a little cynically in training the models, it's like the model is bad at something, then you go to try to buy some data for it and you see if the data makes the eval go up. And it's kind of a cynical take on U.S. frontier labs. It's like the boring stuff that they do. They do much more exciting things. But that dynamic doesn't seem to exist in China. That looks like the biggest difference. It's like obviously open AI researchers have more compute than Chinese researchers. And then therefore these Chinese startups, the people work on different problems. And maybe it looks like more efficient solutions and so on. one that was like, okay, this didn't necessarily need to be different and felt very different. How people currently view U.S. frontier models is that they're going into this knowledge work tool. So as a nominal knowledge worker, I use Cloud Code and Codex to now mediate the vast majority of my work. And people see this happening for healthcare, law, consulting, finance. And the U.S. models are all going to push into this capacity. a lot of the training data for this comes through these data company networks where they're giving money to professionals to create this training data. Yeah, there's a whole market of Scalier, Merkur, Surge, Turing, There's a whole plethora of companies here, and the market is pushing them away from offshored preference data for RLHF into this very high-skill market. vapor industry and business like... It's measured in billions of dollars of how much people are paying for this across each model lab. and They do this because it serves value and makes the models better. at these agentic tasks. And I think it was a question of like, do the Chinese labs unlock this domestic market for their own agents and people using their models for this because that demand is what will let them buy this data that's very expensive versus if the Chinese AI researchers and then knowledge workers are like, I'm using Claude. Like Claude just has all the more reason to keep doing this in the US and that's all for the US data industry. So like it's I talked to my friends that do post-training at labs and like these data companies are really important to getting this incremental performance bump on all of these domains. And so if it doesn't exist in China, you're going to have to be very creative to an AI researcher is then going to be making up an environment in data of say like a finance problem of like making an Excel spreadsheet. And I just don't think this is a task that an AI researcher is naturally suited to do. It's like a different world than they live in. I have a few more questions about Chinese demand and business model. During the mobile internet era, Chinese companies are so good at launching new apps, applications. People say companies like Meituan, WeChat, and Alibaba totally changed. [75:00] our daily lives. And assumption is that China will move faster this time as well on AI applications once the models are ready. Do you see that coming? I think both in the US and China there's going to be incredible dynamism in the applications. Some arguments are even like the US consumer is more used to adding new apps and then like going through the like getting very used to it and paying for it than in China where they have these mega apps. And it's a question of like, can the mega app support the same amount of diversity of new use cases where in the US you have things like Sora, which... is the number one app and millions of people try it and it's a flop. And it's in its own sandbox, it can be a more defined experience. So I would almost say in the software side, the U.S. is... more set up for like this dynamic. trial of new apps. Were Chinese apps worried about monetization at all? not super vocally, but we also didn't ask. I also don't think that many of the frontier AI labs are that obsessed with monetization. OpenAI and Anthropik want to IPO. But most people building AI models are just like, "I want to build the best model." if you assume the model is going to get better, like, they're so useful, the funding won't be a problem. Depends on whom you ask, right? If you ask a product manager, probably they're concerned about it. Or the leadership that's trying to Support buying more compute. Should US labs visit China labs more often? I think we'll see some US especially small and medium players that are trying to build economic value around open models, whether it's a fine-tuning API or specialized deployments or so on. want to build relationships with Chinese labs because the Chinese labs are so open to sharing their technology and expertise. After traveling to China, we've seen like I've, gotten inbound from a lot of companies are like, How do you get access to all these labs? We would go visit, and this is a, you could look up a known... either infra or training company in the US that works on top of open models where If this evolves, it'll be kind of a new era of these new companies learning to work. with each other across borders. I would expect more US companies to go to China because it's in the visa situation right now. So I traveled on the visa-free transit where you enter and exit through different countries. That's super easy to do. You pretty much just show up, whereas a Chinese researcher or engineer traveling to the U.S., you have to wait for the Visa games. So I expect more U.S. medium-sized startups that are interested in open models to go to China, which is probably going to happen in the coming weeks to months. This stuff happens fast. The summer of 2026. That's good to know. You've been a frequent guest on the podcast show with Lex Friedman. He was in China recently. I think one of his videos, catching a truck ride, really going viral on Chinese internet. Do you and Lex actually talk privately about your China trip or his interest in visiting China? We haven't done a postmortem of like, what do you think of China? Both of us are... just interested in seeing this large swath of humanity that influences tech that we live with so much and seeing it firsthand because that's the best way to understand things. I mean, Lex has done more world traveling than I have. He's done a lot of other world traveling to begin with, so I would guess, I mean, I can't speak for him, but... I would guess that it's similar in some of the motivations. Yeah, great. And can you give us some predictions, 12 months from now, What are we going to see? in the China labs and what are we going to see in the US labs? One of the defining things I would expect is that there's going to be some sort of new agentic product that is going to be what everybody is talking about. I think Cloud Code and Codex will still be around, but I think there will be others, whether it's from Google or somebody else that built a new app. I think it was so early in the story of complex kind of agentic workflows across your work. So that'll be a new thing. The unpredictable things are... Do the models just keep getting better as they are, or do we have a lucky breakthrough on something like continual learning? I would guess that It's more like you see the revenue of big and small AI labs. alike continue to grow. OpenAI, Anthropic are public and okay. But more interesting is this medium-sized companies, together, Fireworks, Space 10, in also some of these Chinese labs like Kimi, ZAI, mini Macs where they sell a mix of tokens and models and aren't as giant. They're also going to be doing pretty well in terms of selling intelligence because there's just so much demand for this. I would expect that accepted gap between the frontier models in the US and China to have widened. So if people were saying three months, they'll say six to nine months or they'll say a year. but it won't look like the Chinese [80:00] I've consolidated and given up so we will probably have a DeepSeq V5, we'll have Kimi K3, we'll have GLM 6 and all of these things. So like 12 months is still such a short timeline where we're seeing such a rapid rise in demand, where it's like the markets are still probably going to be going crazy. Like we haven't reacted, we haven't fully processed this demand cycle, so it'll still probably feel like a continuation of that. Where if you ask me like three to five years, I'd be like, uh... One of the major Chinese labs might have gone under due to finance, like, I don't know, like, Minimax or GPU goes under for some financing thing. there's more corrections in the US market and stuff. I think once you get to that three to five to ten year timeline. Cool. After this trip, are you more optimistic or more worried about the global AI race? I think I'm pretty neutral. I think most of my worries on AI are like very centered on the US opinion being sour on AI and how that could hurt I think the global AI race is really what I would expect geopolitically. In this biggest picture, it's like the US is ahead and the Chinese wants to catch up and there's this nuance in terms of how much chips are being sold. and what that pin ends up doing to the ecosystem. I think it's mostly just the benefit to be able to bring more humanity to this and do what I can to make it less of a very extreme circumstance for people on both sides. So people building and using the technology that have very little say on what the highest level decisions are. It's like I could do a small part to make that more understood for both sides. All right, thanks so much Nathan. I think all those are my questions. Anything else you want to add? Any question I didn't ask? No, thanks for having me. Hopefully sometime soon go for a second trip to China. There's plenty more to see. Yeah, we'd love to have the second interview after your next trip. Thank you. Thank you so much. Now tell us about what you guys do. So A2 does all sorts of kind of open source AI research and development. I think it's historically been fairly academic and over time it's been moving more into this artifact. defined development where you're building models and code bases and data sets that the rest of the community can build on. So I think Olmo in these open language models has been the most popular one, but there's also like Momo the multimodal almost Earth and other things. And then AI2 has historically been very well known for its integration with AI literature. So I think the application there has been Semantic Scholar, which is kind of seen as alternative to Google Scholar for searching the academic literature. And this team has evolved into building AI agents for science. So it's this ASTO brand where they're building tools to assist people in building and understanding the scientific literature. So I think that is core AI2's deepest culture is so scientific and like has this community value of building a lot of just like net good for the scientific community, which is really great. And that comes through in these papers that are very detail-oriented and caring, and I think there's just like so many different ways, which way we want to go. There's so many different ways that this comes through. And it's the best, one of the best things about working at AI too, is you just like, you get so focused on trying to build things that various researchers use when you travel as an academic and you're like all these people are building on ULBO and know of AI too and it's just kind of like a lasting thing that is great. When there's interns or student researchers and they go out into the academic community and all these other research institutions to kind of spread this scientific value, which is a great service that AI2 does. Yeah. How many people do you have right now? Probably about 200 people. Oh wow, that's a lot. So most of them are researchers, right? I would guess probably half of the talent is technical engineering on like the traditional AI research and then there's engineering teams that help support our clusters and some of our products. How is like an NGO lab different from other labs in Silicon Valley? So AI2 is a full non-profit, which means most of its money comes from philanthropies. So the Allen is on the door, which is from Paul Allen, one of the co-founders of Microsoft, and then also has large... grants from, for example, from the NSF. So it's all nonprofit funding and doesn't really sell any products, it just gives everything it makes away for free for people to learn from. Wow, so you guys have more research freedom, I guess. Yeah, so I think that on the spectrum of closed lab to academia, in terms of resources, it sits in the middle in terms of freedom. We have more resources and we feel pressure to build models that are actually good, so it does suppress the freedom a little bit, but much less than a frontier lab that is racing towards AGI. But if you go to a true academic institution, you'll see that they have a diversity and kind of intellectual spark that you won't get added. [85:00] company just based on the needs of paying the bills. How is Seattle vibe different from the Silicon Valley vibe? It's so much calmer. Like there's not events all the time with various people meeting up. I personally wish that more people in Seattle like there's major tech talent hubs at Amazon, Microsoft, and all the tech companies would have more of this. But it's what it is you can't fight it in some ways I think it's a better place to like settle down and have a family because it's like a bit more spread out and it's a a bit less intense and how bright everything is in terms of talking about AI. Like, I don't have AI in every conversation, where in the Bay Area, AI feels like it's the only conversation right now, which I understand why. Like, I'm bought in as well. But you visited the Bay Area a lot. I go regularly. I say I go Like three to eight times a year. Okay. I went to UC Berkeley, so I lived in the Bay Area for like seven or eight years. and then you have to go back to stay in touch with what's actually happening. In the Bay Area, you don't have the view like this, a very beautiful lake. What's the name of it? Lake Union. Definitely my dreamy office for sure. Yeah. So when you visit China, how do you think the office there, the AI lab office there? They don't have this view, but they're also tech company offices where it's like they're pretty similar in that regards or like I go in them and it feels like I'm in a tech company and I've been into a lot of tech offices in my life and it's like okay the showroom culture is different oh yeah sure but that's like unique in China that is a very unique China thing that a lot of people just like captivate for their first time and they're like okay this is the thing - Cool, cool. Thank you so much, Nathan, for having us. Yeah, thanks for coming.