China's AI Upstarts: How Z.ai Builds, Benchmarks & Ships in Hours, from ChinaTalk

Hello and welcome back to the Cognitive Revolution. Today I'm honored to share a special cross-post from the China Talk podcast hosted by Jordan Schneider, China Talk analyst Irene Zhong, and Nathan Lambert of AI2 and the Interconnects Substack, featuring a conversation with Zhixun Li, Director of Product and GenAI Strategy at Zii AI, also known as Zhipu AI, about the culture, incentives, and constraints Shaping Chinese AI development. Now, I imagine that many, even in our AI-obsessed audience, will not be familiar with Zhi.ai. But their model releases demonstrate that they are a significant player worthy of our attention. As of today, their latest GLM-4.6 model holds the number 19 spot on the LM Arena text leaderboard. Its Elo rating is roughly 65 points behind the current leaders, which means that it still wins 2 out of 5 head-to-head comparisons with the leaders, And it happens to sit right next to Quen 3 Max, Kimi K2 Thinking, and DeepSeek's V3.2. Together, these 4 models, all from China, are indeed the top 4 open-source models available today. Though it should be noted that Mistral is not too far behind. On the web development leaderboard, GLM 4.6 does even better, coming in at number 9, making it competitive with GPT-5.1 and meaningfully behind only Gemini 3 and And the new Claude 4.5 Opus. All that said, this conversation goes way beyond benchmarks and touches on a number of important topics, including why we should understand Chinese companies' open weight strategy not so much as an ideological commitment, but as a practical marketing tactic from companies that are seeking to gain global mindshare while recognizing that Western enterprises simply can't and won't use their APIs. The culturally distinct AI use cases such as roleplay that matter in China and how these drive different fine-tuning priorities than what we typically see from Western companies. The role that Silicon Valley thought leaders play in establishing credibility for Chinese companies, even in their home market. The market for AI talent in China and why very few people at ZAI even know that Zixuan studied at MIT. Zixuan's view that there is a wall and that further architectural breakthroughs will be needed. The extreme velocity with which ZAI releases models, which often involves shipping within hours of completing training. What, if anything, people in China fear about the AI future, and how Chinese companies generally still view themselves not as peers or rivals with leading American companies, but as upstarts that are just trying to keep pace and who will be quite happy if they can secure for themselves a meaningful niche in the global AI marketplace. This is a fascinating, detail-rich conversation, and regardless of your attitude on U.S.-China competition, it's clear to me that we in the West don't hear nearly enough from the actual builders inside Chinese AI labs. So I appreciate Jordan for allowing me to cross-post this episode, and of course I encourage everyone to subscribe to China Talk online at chinatalk.media. With that, I hope you find as much value as I did in this behind-the-scenes look at frontier AI development in China with Zixuan Li of ZAI from the China Talk podcast.

Zixuan Li, who studied in the US before moving back to China, works at Zhurupu AI or ZAI. We're going to let him introduce himself and his role. Co-hosting today, Irene Zhang, a longtime China Talk analyst, as well as Nathan Lambert of AI2 and the Interconnects Substack. Welcome to China Talk, everyone.

So personally, I've known of Zhipu AI or Z.AI for at least over a year, and I've been following the work there. And then this summer, it was like kind of in my mind, I was like, well, there's another DeepSeek when they released GLM 4.5. Um, you'll have to correct me on whether or not it was before or after the KIMI-K2 model.

After.

My guess is after. And that was like the sequence of order of weeks or days after KIMI-K2. And it's just like, wow, like, okay, there's a lot of people building great models. And it's fun to get to learn about some of them and how it compares between US labs and Chinese labs. And I think a lot of it is more in common than different.

Hi everyone, I'm Zixuan Li from Zedol AI, and I manage a lot of stuff like global partnerships, Zedol AI Chat, model evaluation, and our API services. So if you have heard of GLM coding plan, I'm actually in charge of this thing too. Yeah, nice to meet you everyone.

Yeah, thank you for introducing yourself. I'd love to hear more about how you ended up working in AI after moving back to China and Zedol AI specifically.

Yeah, actually I applied for multiple roles and companies like Moresha, like other minimax, but got rejected or got neglected because there are so many resumes going to them every day. And I studied AI for science and also AI safety at MIT. So I do a lot of research on AI application and AI alignment. That's not very relevant to what we're doing right now, but actually it gave me a sense of what's going on in the frontier area. So it helped me a lot, like any aspects in what OpenAI, Anthropic are doing, like at that time.

I think it's very innovative to have these sorts of ideas and Was it ever a debate for you whether to stay in the US or did you always know you were going back to China after grad school?

Yeah, I already know, like, I'm going back because my family's here. Yeah, but I got the job after the graduation because it's hard to get a job. I continuously applied for jobs, but finally after one month, Yeah, I got an opportunity to interview with Zed.ai and at that time I'm not in charge of the overseas department because we, our focus is in the domestic area, domestic chatbot. So I'm responsible for the strategy of developing a domestic chatbot. It's called ChatGOM.

Gotcha. So maybe let's do a little bit of Dripway AI backstory. Kind of when was it founded? How would you place it within the broader landscape of folks or teams developing models in China?

Great. So the DreamFuent and also like Zedal AI was founded in 2019 and we are chasing AGI at that time, but not with LLMs, but with some graphic network or graphic compute. So We did a thing like Google Scholar and it's called A-Miner. So we use that type of thing to connect all the data resources on the journals and like research papers and into a database and people can easily search and map these scholars and their contributions. So it's very popular at that time, but we shifted to exploration of large language models in 2020 and we launched our paper GLM in 2021. So that's like, I believe, one year ahead of the launch of GPT-3.5. So very, very early stage and we were the— we were one of the first companies to do the exploration of large language models. And after that, we continuously improved the performance of our models, tried new architecture. GLM is a new architecture actually, but we're going to explore more in the future. And I believe that we got famous by the launch of GLM 4.5. And also 4.6 because I think it's very capable in coding, reasoning, and agentic tool use. So yeah, that's more useful compared to the previous version. And like people may know us through ClockCode, KiloCode, and other tools. So we need to like combine with these top products and got us famous. Yeah.

Yeah, let's talk a little bit more about the evolution of GLM 3.5, sort of, I don't know, Nathan, this is your question. Why am I asking this question?

Well, it's like, okay, what does it take to transition from the models that you were early to this to things that get international recognition? So like, I have known of Z.AI and your work for years, and then it's like snap of the fingers and you're like, okay, now this model is on everybody's radar that's paying attention. And does this feel like something that was just going to happen for you overnight in developing the models or any, like, what does that feel like when you go through it or how do you get to that moment? Because there's a lot of people that want to do that at their companies.

Yeah, that's a very, very interesting point because In 2024, everyone is interested in Chatbot Arena, right? Like we see GPT-4, we see Gemini performing very well on Chatbot Arena. So that's our interest because like we pay attention to end users' experience. Like when there are 2 answers, which one you prefer? So we did a lot of things on that and we did perform very well on Cheb Arena. So ran maybe 6th or like 6th to 9th on Cheb Arena. But in 2025, with the launch of Mana's call code, we realized that like coding and genetic stuff are more useful or They can contribute more economically and also in terms of the efficiency for people. So chat, I think we're no longer like put that as our top priorities. Instead, we do more exploration on the coding side, on the agent side. So we observe the trend. And we do a lot of experiments on it. Yeah, we need to follow the trend and also predict the future.

Do you feel like you're better at executing for code versus chatbot arena? Because like, like I, like GLM 4.5 and like, I think the error version and 4.6 are like extremely renowned for this. And I think like the, the, when you train these models, the process can look very similar depending on what your target is. And I'm just wondering if like, do you feel like it was a shift or is just like sometimes Things work out better than others?

It's a shift, actually. Yeah, we pay more attention to the coding stuff. And like on the Zed.ai chat, we are free, right? So nobody's paying for the chat. Yeah, people pay for Cloud Code use and for agentic stuff, but we just let users chat with the chatbot freely. Yeah, so that's a shift, but we need to continuously improve the performance in normal chat and maybe role-playing, but not our top priority.

Hey, we'll continue our interview in a moment after a word from our sponsors.

So yeah, let's talk a little bit about the talent and the internal culture that allowed you to put out GLM 4.5. What do you think is different about, or what distinguishes AAAI from other labs both in the US and China?

I think the first of all, we are more collaborative, like inside the company. So everyone is working on this single target and maybe we have like some head of separate teams like pre-train team and post-train team, but they are working very closely. They just sit next to each other and working on a single target is trying to build a unified reasoning agentic encoding model. So we have built 3 separate models. As we have illustrated in our tech report, we then distill these 3 teaching models into one single model, is GLM-455. So that's our goal. And that's like, I believe that that's how we built GLM-455 more efficiently compared to other companies. And they are super young, right? And another point is that Like you mentioned talent, I believe that nowadays you need to do the research yourself. You need to do the training yourself as the head of the team. So you cannot let others do this stuff for you.

Why is that?

Because things change really fast. Like maybe during your training, there comes GPT-4, there comes like GPT-5. Anything can happen. So you need to feel the trend yourself. Like you need to combine the results from experiments, the trends, like what's going on within your competitors' team and to feel the move yourself. It's super important. Like even our founder, he did the experiments himself. He looked at the papers. Yeah, you need to do things simultaneously, not just set goals for people and let others do the stuff for you.

Yeah, it seems very fast-paced. I think before we started recording, you were also mentioning that there's a lot of PhD students involved. I was just wondering if these are people that are actively pursuing their PhD or kind of new grads or a mix of all of them, because I think that I work at a research institute which is very open source and we like have a lot of full-time students that are part of it. But where you look at other closed labs in the US, like there's not nearly as much intermingling with the academic institutions. And I think that could be a really powerful thing if you have this because there is a lot of extreme talent there. So I'm just wondering if you feel like it's a kind of open door between some academic institutions and your work.

Yeah, definitely. There are a lot of ongoing PhD students here. And I believe that they are both chasing their academia and the work of GOF simultaneously, but they can combine them together, right? If you are doing a really innovative job, like training a unified agentic encoding model, it's one of your greatest achievements ever, right? So people won't say— Okay, I need to do another research and let me finish this first and then we'll go back to the GLM. Like they will try to treat GLM as their single biggest achievement. So everyone is really devoted to this stuff. And yeah, we hardly see anyone not devoting to training GLM.

Could you talk a little more broadly about the talent market? I mean, you mentioned earlier that you had to put your resume in a lot of places. What does it look like right now? What's the kind of like hierarchy and what are, you know, what are folks looking— what are employers looking for and what are—

what's the talent looking for? So on the research and engineering side, I think They're looking for papers, looking for like GitHub code, looking for competitions. Yeah. And also your experience on using like GPUs, right? Your experience on like training models. But for the non-technical side, they're looking for how you're gonna grow the model performance, expand your branding, and also a lot of stuff like If you're going to be a product manager, like your live coding skills, your vision on this area, and also how you do the stuff yourself, yeah, those are very, very important. I think it's pretty similar, but you mentioned hierarchy, right? So in terms of hierarchy, large companies choose the people first because they have more money. They can pay more, like ByteDance, Alibaba. Yeah, but for startups, like we need people to fight together. You need to like fight against other competitors. You need to drive yourself to finish the goals because you don't get paid that much. Like you need ambition. You truly enjoy like working with really young, talented people and try to build something like GOM. Like, it seems come from nowhere and try to beat other competitors' models.

Yeah. How big would you say the number of people that are actually training the model is? I think in the US, it's accepted that the core research or engineering staff normally doesn't get to be more than 100 to 200 people. at the likes of like OpenAI or something. And then there's a lot of support around them in terms of product and distribution and stuff like this. Do you feel like this is similar where there's kind of a core small research team?

Similar, like 100 to 200 people. I think that's enough. Yeah, because you need to be focused, right? Like there are people preparing data and there are people doing the product stuff. But for the core team, like you don't need that much because you need to stay focused and these people need to be really talented. They cannot make many mistakes, right?

Do you know that's different at bigger companies?

I think for bigger companies, there might be different groups, right? They have like more GPUs and they can do more exploration. Like, for example, like in ByteDance, they're chasing the top performers, not only in text generation, but also in video generation, speech, other areas. So they can allocate the resources to multiple teams, right? But inside these teams, the core members, I think, is still the same, maybe 10 to 20 and the other like 80 or 100 doing the training stuff or data preparation.

There was a lot made in Chinese and Western media about how DeepSeek was biased against people who'd studied abroad. I'm curious about any other broader dynamics you see with relation to returnees versus people who did their whole education in China?

I think there's no bias, like, because they want the best people and usually the interviewees, like, only, like, stay in China, like, perform the best in their interviews. Yeah, they have no bias, but maybe people coming from US or other countries just did worse in their interview. Yeah, I believe that that's not a bias because they're— I think they're judging very well. They have their standards, maybe their standards different from like what you do like around the world. But actually, I believe that that's not an issue. Like even inside China or inside our team is the same standard. Because I, like, I joined Zedol AI after coming back from the US, but I think nobody actually knows. And people will never ask, like, are you studying abroad? Or like, have you ever, like, master's degree from MIT? I believe that maybe just 10 people know about this in Zedol AI.

Yeah.

So there's no bias because people don't care.

Yeah.

Hey, we'll continue our interview in a moment after a word from our sponsors.

Let's talk a little bit about open versus closed source, um, in Chinese model developers broadly and with Z.AI in particular. How do you think people, um, what's the, what's the thought process behind so many models going open source in recent years?

Yeah, first I think like maybe generally we need to devote more to the research area. Like Llama's doing this, Quinn's doing this, and Kimi's doing this. We are also doing this. We want to contribute more to the academia and also the exploration of all possibilities. I think that's our Top priority, right? But beyond that, as a Chinese company, like we need to really be open to get accepted by some companies, right? Because people will not use your API to like try your models. Maybe they deploy on Fireworks, maybe they use it on Groq. And maybe they download to their own chips, right? So I think it's not easy to get famous in the United States because, yeah, people just don't accept your API. They need to be stored in the US. So I think it's necessary to be open right now for people to use GLM.

I mean, this is like, this is what our company does. Like where I work, it's like, I wouldn't be able to sign up for the API service of the enterprise, but like I distill from multiple Chinese models when I'm training. It's like, I'm using multiple models and might come across this. So it's not surprising, but it's good to articulate it.

Yeah.

We also learn from DeepSeek because we have closed source version in 2024. Like our flagship model is closed source back then, but when DeepSeek R1 launched, we realized that, oh, you can do this thing simultaneously. You can be really famous for open sourcing your model while get some business return, like through API or other stuff or collaboration. Yeah, you need to expand the cake first and then take a bite of it.

Um, maybe taking one step back, like, why is it so important for Chinese model makers to get, um, I don't know, famous in the US or just global adoption more broadly?

Yeah, because I think there's better ecosystem for developers, research, still, still in this United States. Yeah, you need to get accepted by the top researchers or you, right? Because if we don't open source our models, we'll never have opportunity to join this conversation. Yeah, it's also important. Yeah, because we learn from X, from YouTube, Reddit every day, and all the Chinese tech media is also paying attention. To, to US, like, KOLs or influencers.

This was very surprising, I think, both to Nathan and I, how kind of recursive it was, where, like, the Chinese media covers the models, covers the Chinese models that the Americans are talking about. It's just, it's a very curious trend.

Yeah, because you have, like, People like Andrew Karpathy and also Sam Altman, Elon Musk, they not only talk about their own models, but also what's going on there. So everyone knows. Yeah. If they post a tweet, everyone knows what's going on, what models they're picking. Yeah. What preferences they have, their views on maybe PaLM versus Codex. Like all the social media will trying to grasp their core idea immediately. So that's very important. We will also learn this from DeepSeek. Yeah. Frankly speaking, we used to neglect the importance of global economy previously because we think we need to sell our products, sell our APIs directly to Chinese enterprises. But nowadays, like Chinese enterprises are still paying attention to like the global brand and your global performance.

Yeah, this reminded me of something I've been curious about, which is we know the conversation is recursive. We know that the Chinese tech pays a lot of attention to what American like Silicon Valley is looking at. But is there anything about the AI debate or discourse in China that Western media tends to miss, in your opinion? Are there any kind of issues or debates or things that people are really interested in that people in the English-speaking discourse tend to not understand?

Yeah, so I just talked to a professor from Germany yesterday and he mentioned Some models that he knew people are talking about these days, like LLaMA, Qwiint, uh, even Mistral, but not GLM. So there are many people still missing out on a date.

In the, in the SF circles, more people are talking about GLM than Mistral and arguably LLaMA these days. So you've made a lot of progress.

Yeah, we've made a lot of progress, but we track, we also track the discussion on Reddit, other social media, but we still see a lot of people talking about what is GLM? Like, is it a good model? Like, where it comes from? It comes from nowhere or similar stuff. Yeah, still need to do a lot of things because We only have 20K followers on X. So that's quite few, right? Yeah. So nobody actually get a very deeper understanding of GOM compared to other models.

I think DeepSeek has like a million. It's crazy.

Yeah.

I know. Like, that's even big for a whole lot of American, like, for that It's like for a new American tech company, that would even be big. It's remarkable.

Yeah.

Also like Mistral, Cohere, I think they get much more attention compared to Kime and Zedoya. So we still need to do better in our branding or our engagement in the technical community.

You mentioned Selling API access to Chinese companies. Um, I don't know, tell us a little bit about adoption in China, what the sales process is like. Do they all just have VPNs and use cloud code anyways? Like, what's the, um, what's it like trying to, um, you know, do enterprise sales in China?

Yeah, so you have like 2 types of enterprises. One is the companies that can't use API because there are companies that need to deploy the model on their own chips and they cannot accept sending data to other companies, even ZhiDao or even Alibaba. So that's a requirement for those companies. They require DeepSeek. So there are a team like deploying DeepSeek for them, not from DeepSeek or any companies can deploy DeepSeek, right? For them. And they usually built on top of the DeepSeek model with RAG, like data storage and workflows and other things. Yeah. And the other one use APIs, maybe from tech companies and like media companies. Yeah, these companies accept API because they need standardized their workflows. So for the API company, I think they choose between the balance, the performance and the price. So ByteDance is doing great in that area. Like ByteDance, I believe dominates the API services. And also like Quin is still trying to sell their APIs because Quin 3 Max is a closed source version, right? If you have, you have like heard of it. So they have open source some models, but also keep some, some thing closed source for selling. Like for, for we, we have open source our faction models.

So.

We are always frequently asked, like, what's your service differ from the open source version? Because we can deploy the open source version ourselves. So we need better engineering team. We need faster decoding speed, right? So we need to do more on top of just a good model. So that might be our unique selling point. We need to do searches. We need to build our MCP.

Yeah.

Trying to get a competitive advantage over other GLM providers.

Is that annoying?

I think it's—

or fun?

It's fun. It's fun because I think it's necessary to open source your models. So how you get a bite in that cake is really important. Yeah, because we have been figuring out for a long time, but recently we found subscription is a good idea, a GLM coding plan, because with subscription, your users will become more sticky and yeah, they love the— this area because you have to— you don't have to worry about like how one prompt consumed in your dialogue. Maybe inside Claude Code, a round of like interaction will consume a million tokens, but you don't have to worry about it. So we'll figure it out for our users.

Do you think you have meaningful adoption there? Because in the US market, it's like I could start using Claude, Codex, um, Gemini and whatever, all for free, like some basic cursor. And that's why I was wondering, like, are people in the US actively— like, is, is this a growing market that you think you're gonna eat into? Because I mean, Quen has one and I might have tried them, but I think I'm always like, oh, I have my ChatGPT subscription. And I'm just wondering if you feel like on the ground if it feels optimistic as something that is Like really shifting the needle.

It's definitely very optimistic because we don't have to persuade 50 people, 50% of people doing this. Maybe you only need 5%. Yeah, but 5% is a huge market. If 5% of Cloud Code users ship their model to GLM, it's a huge market.

Yeah, and it's growing so fast.

But not for cloud, not just for cloud code because we're trying new ideas like role-playing. Yeah, many, many people on Solid Torment are using GLM, generative AI, because we did very good in role-playing. Yeah, so we were trying to have more markets, coding markets, agentic markets. Yeah, maybe one day like Matt is using our model.

All right, we gotta, we gotta take a step back and explain role-playing. Um, what is it? How do you make a model that's good at it? Um, what, what are people using it for?

So before GPT-4.6, um, for models like GPT-4.5, we are relevant— relatively weak in role-playing because we, we haven't trained on Those data. So we need to create some data and let the model follow the instructions because for role-playing, like very long system prompt, if you don't train on that kind of stuff, it will forget who it is and forget all the instruction and just use its like general performance. to do the conversation. But for role-playing task, if you've given very long instructions, it'll strictly follow these instructions and like show more emotion or show more behavior, like for following these instructions. Yeah.

So just to be clear, this is, this is You know, people having a conversation saying, you know, I'm a Japanese pirate, I'm raiding the coast of Taiwan in, you know, 1570, and I want to plan an attack to defend the fort. And, but like, people write out like 5 pages of background. And then, you know, these are chatbots, right? So you're having conversations where, You're, you know, it's like playing like a text-based RPG from the 1980s, except it's AI and it just generates.

All right.

Sorry. To be clear, I'm not sure everyone knows what role-playing means when it comes to AI. Yeah.

And also, like, we try something very interesting, like Family Guy. Like, we have our own Stewie. You just give a description of what Siri does, a history, and then you can create your own Siri. But we perform very well in text generation, but like if we have some speech model, like we can recreate Siri.

Was there a specific kind of Pre-training data or RL that you needed to do to get this, or did it just all of a sudden you're like, oh wow, like this is really good at pretending to be cartoon characters?

Yeah, I believe, uh, it's mainly post-training data.

Okay.

Um, there's a big discussion of late in the US about people being worried folks are falling in love with AI. There's this whole discussion about like AI psychosis where ChatGPT kind of like convinces people who like trust it too much to harm themselves. I'm curious your broad sense of that type of discussion in China broadly, and then internally in your firm about these sorts of questions where people are using AI to play or for emotional support.

Yeah, I just read the post from OpenAI yesterday because they invited a lot of experts on trying to train a model that I think are not addictive. They like train data that ask ChatGPT to say it's an AI instead of saying it's a human being, not letting people attach to ChatGPT anymore.

Right.

So I've read this, a bunch of people read this, so we can discuss. I think it's internally discussed when people find the relevant stuff. Yeah, we had the news. It's a hot topic and we can discuss. But from a broader audience perspective, I think not people, not many people are looking into this. Yeah, because we are not there yet. If we have a model that can perform like GPT-5, then we can move on to remove the addiction, but still performance is not on par with these like top closed-source models. We need to chase first. Yeah, because when we chase these models, we'll shift our focus on the data collection, data preparation, and sometimes the model behavior will change dramatically. So if we do some similar things on our previous model, it will be outdated in the next version. So the performance is still very, very relevant.

Currently, I'm guessing this is somewhere in the rundown, but like how is the how is the balance of optimism versus fear of AI as like a long-term trajectory in your lab versus China generally? Because I think there's a very big concentration in the U.S. of people that worry deeply about the long-term potential of AI, whether it's like a powerful entity or concentration of power or other things. Like there's people that just think that this is the most important technology that has ever been invented. Like, we have to be really serious about it. Like, I'm just wondering where on this kind of spectrum you think the lab has a culture of, or if it's not really something that's debated and you're just like, we're building a useful thing and we're going to keep making it better.

I think developers fear the most. Yeah, because when you use cold code, when you use codecs, you get the fear. Very, in a very concrete way. Like you can do all the tasks for you, especially for junior developers.

Yeah.

But like for writers, for other managers, I think it's simple because we have SaaS, like we have, we have like other technologies helping them. Already. So large language models like ChatGPT, just another helper for them. So I cannot fear, feel fear coming out from the general public. But specifically for developers. Yeah, but for developers or data analysts, so they fear the most because they try out the the new models, the new products more frequently than the general public so they can feel the power. So if you don't use tools, because many people use DeepSeek and other chatbots, but DeepSeek can help you brainstorm the ideas, can help you polish your writing, can do the translation for you. But they don't believe that this work can replace them. But for developers, like, it's a different story.

So what are the main fears? Just people's jobs getting taken away, AI taking over the world, whatever? For the people who are worried, what are they worried about?

Maybe jobs, jobs taken away.

Those are pretty different than the US. There's definitely a huge culture. There's not a majority in terms of the people, but a very vocal minority that influences a lot of the thinking of the risks of AI, well beyond just job loss. Job loss is almost an assumption for many people in the US. And then there's added fears on top of this. And I think that's a very different media ecosystem and thought ecosystem.

Yeah, I definitely know about this because I did the research.

Okay, you lived here. You lived through some of this, obviously.

Everyone at MIT are talking about like how AI will change the world, not on the positive side, but on the like negative side.

Why do you think this is? Is it that Chinese society is a little more practical or just that job loss feels more Imminent, or is it because it's less of a market-driven economy?

I believe that people just know about DeepSeek because maybe just 1 million people follow the latest trend and there are a billion people do their work daily and not impacted by AI.

Yeah.

So the more you learn, the more fear you'll have.

And then what's the vibe among these younger engineers that you were talking about, like junior folks who are a little scared? I'm generally just curious what gets them into this work in the first place and what makes them want to work at places like ZAI.

At ZAI, I think we lack people. Yeah. So there's no fear about losing jobs here because Yeah, we have a lot of things to do, but for other companies, especially the large enterprises, they have maybe tens of thousands of people doing the similar things like data analytics and also like backend engineering stuff. So they might think if Other people are using Copilot or agentic tool. Maybe they just need 50% of the people. Yes, but they can do nothing. They need to wait for their bosses or the founders to make the decision. Like what's happening at Amazon, right? So for layoffs, you can do nothing. You just wait for the results.

I wanted to jump in here and also ask about translation because Zhiyao's models are very strong in making very contextually rich translations from Chinese to English and deploying that to social media. Could you talk a bit more about the process behind that, if you know, and what's the secret sauce to translating memes?

Yeah, exactly. We are doing very great in translation and especially the translation for Chinese and English. I think we are on par with Gemini 2.5 Pro. Yeah, but you mentioned memes. Memes is also one of our weapons because we just prepared the data and we understand the culture. We can even translate emoji. Yeah, for example.

What does that mean?

How does that work?

Yeah, because if you like Tencent emojis to Apple emojis?

No, you could, if you enter a sentence talking about AI and you use a whale to replace DeepSeek, we might translate this back to DeepSeek. And if the sentence is about animals, we will translate into a whale. Yeah, you understand the context.

Is this because Chinese internet talk is just so cryptic?

Yeah, because people are very, they're very naughty. They're naughty. They sometimes use emoji and there's a lot of companies including some animal names in their There again, in their brand, in their logo, and we're trying to use that to replace what actually people use, and also people use abbreviation, right? So all those things need to be translated right.

I remember a few years ago there was all this discussion like, oh, it's going to be really hard to train Chinese models to speak colloquially because all the data is behind You know, walled gardens and like, you know, you can't get the Tencent. Tencent has the Tencent data. Xiaohongshu has the Xiaohongshu data. Ali has the— I mean, I don't even know what data they have. But like, is that— was that a problem for you guys doing this more kind of like colloquial internet speak? Or is there enough out there and, you know, you can just scrape stuff and figure it out?

We need static data. Right? We don't have the actual data. We cannot scrape something from other WeChat user profile log.

Yeah.

But we know people are talking about, especially in the public area. So in the open area, we can observe what's going on on Xiaohongshu, right? On TikTok, on other things. We especially pay attention to their comment area because people are really naughty there. in their comments. So actually when the TikTok refugees thing happened, we benefit from it because more people or more softwares need auto-translation and we were trying to conquer some large customers like through our translation capabilities.

Does anyone train on like Danmu data?

Like, definitely, definitely. Yeah, all the memes, we're trying to collect memes from everywhere, especially for our vision model, right? Because like for the memes, they're always in like image format. I'm trying to understand it with our vision model. I think it's very interesting. And it's also very necessary because if you cannot translate the comment in a very accurate way, they will not purchase your medal. Unlike YouTube, because like if you use YouTube's auto translation, it won't grasp like the exact meaning because people will just need to understand, oh, This English version is about this, and I can read it in Chinese. Oh, 80% is enough for me. But for apps like X, um, Rando, Xiongshu, WeChat, you need to understand like 100% of the, the comment area.

Is it a challenge to balance like I mean, let alone culture, but data across like US, like you're marketing to Western users as well and you have your domestic market. Like, is that a technical challenge to feel like you have to do both excellently?

I think it's a challenge, but we can do very well in Chinese and English and we're trying to explore more in French. And even Hindu. So we have data on Hindu. So we can perform very well in, I believe, in like 20 languages. But beyond that, we're still exploring the data, their software. So we need to register on their software to see like what people are doing out there. Yeah, sometimes it's hard to figure out. I'm trying to like learn from Gemini and GPT-5, like why do— why they do so great in translation.

Uh, could we talk a little bit about compute? There's all these rumors. We're recording this, uh, October 29th, the evening US time. Are you excited to buy some Blackwells if, uh, they come on the market in the next few weeks?

Blackwell's is great because not only the chip, but also FP4, right? So FP4 can reduce a lot of cost. And yeah, we're trying to use the best we can get. Yeah, that's a strategy. I think it's pretty clear.

Yeah.

For the model training side, for the architecture, we use the best. For the chips, not the best, but I think like we do the best trade-off between the performance and the cost.

Do you guys train outside of China as well, or only on domestic clouds?

Yeah, we do the inference from outside China. Yeah, but all the training is going on here.

How do you feel about Huawei chips and software? Are they— Are they going to make it?

Yeah, we are going to, because we have multiple models like GOM 4.6 and the upcoming 4.6 Air and also our previous version. So we need to find the best use case for all sorts of chips, domestic chips and NVIDIA chips. Yeah, we need to classify the use case because For one customer, maybe it needs like 30 tokens per second, and for another customer needs 80 tokens per second. Yeah, so maybe for one customer or one use case, some chips are enough, and for others, like we need better chips and better inference techniques.

Do you try to do any API sales or just enterprise sales in general outside the US or China? Because we mentioned having a lot of languages and whatnot. Do you see use cases coming from other places?

We have 2 platforms. Like inside China, our platform is called Big Model. Big Model, it's like a large language model. It's a simple translation, bigmodel.cn. And we also have a Zedot AI. It's called api.zedot.ai. It's our overseas platform. So I'm actually in charge of API.Z.ai. So all our services are hosted in Singapore.

Yeah.

So actually I'm an employee of a Singaporean company.

Oh, sorry, I wasn't clear. I meant, do you see much demand coming from, for Z.ai coming from non-US countries, like other countries?

A lot of countries, like India, Indonesia, even Norway, and also Brazil. Yeah, but it depends on who's using Reddit, who's using X, because we basically did our growth on X, on Reddit, maybe some on YouTube, like if people are watching it. Watching these materials or videos, like they will purchase it. But we, we, we're trying to do Telegram or other things. So it might shift the, like the proportion of our users. Yeah. But, but India, Indonesia are huge markets. So, but there are more revenue coming from US compared to other countries because they pay more. So they, they buy like Pro plan and Max plan instead of the Lite plan. So yeah, in terms of users, I think India has the most number of users, but like US market generate 50% of our overseas revenue.

Jordan, are you on the Chinese plans yet? What's your AI bill? How do we diversify this internationally? I'm on like $500 a month. It's not good.

I don't know, just charge it to the firm. Charge it to the Allen estate, Nathan. Come on, we gotta, gotta save you. Um, Irene, what was your, what was your question earlier?

Um, building off of what we were talking about earlier with like, turns out walled gardens didn't matter, which is that, um, Whether Zed.ai has any thoughts about doing AI search on the Chinese internet and what that would look like in China, where there increasingly is no unified open internet.

I think that's a challenge also for US product builders because Google don't have an API, search API. And Bing is trying to stop its search API. So there are other third-party providers like SERP, and they basically just scrape the data and they quickly send a request to Google and scrape the page, right? So also very challenging for builders like Perplexity and even ChatGPT. So we need to do the technical side nowadays using our own technology or trying to grasp multiple resources from different platforms. I think that's very reasonable. And there are other technologies like Manus. Like they just browse the internet themselves, like without using the API. I think that's more doable these days when you want to see multiple resources and trying to distinguish the best use case, the best resources you need to really log into account. And see the data yourself, read the page yourself instead of just using like whatever API gives you.

Nathan, maybe you want to ask some like broader research direction type stuff or whatever else is on your mind.

Where are you planning to take your models next? I think less in domain, but like, how do you make models better given that everybody has limited compute and data resources and we're changing from chat to agents and it's just like, How do you, like, how far out do you think? Or do you think about like the very short-term problems or like, there's just so many directions that you can take it. I don't know.

Yeah, I can give some names of the ideas we are exploring right now, like on-policy training, on-policy reinforcement learning, because we are quite mature in off-policy reinforcement learning. But for the on-policy learning, we still need to explore more and also multi-agents.

Yes.

So when you look at the ZeroAI Chat, it actually acts like a single agent. So one model do the search itself, comes back and do another round of search, then come back and it can generate slides and generate presentation or generate a poster, things like that. But it's all performed by a single actor, the one GOM 4.6. But maybe for all—

I think our models are like, do we think you have to change our models a lot in order to do this? Because I think like so much of 2025 has been changing the training stack away from Like, we are a chatbot to now we are an agent. And it's just like, what do you think we should change the most about our models given that that is like, it's almost like the air model, the faster model might be more useful because you can have more of them and things like this.

So that's the reason why we need to do a very solid evaluation because we have different product solutions and currently The single agent works very well on our platform, but we need to do more to try out different ideas and see whether we can improve the speed, the performance with multi-agent architecture and yeah, other possibilities. Because for single agents, it has better context management. Yeah, because you have the best model that can see all the context ahead of the current conversation and follow the instructions maybe up there. But for like multi-agents, you need to compress the context for each agent, and that might lose some context or info.

Or like orchestration is hard where it's like if you have all, you give 4 agents the same context, they might all try the same thing and they might not work together well and stuff like this.

Yeah. And maybe even one agent has hallucination, it will like ruin all of the research. But also like we are trying to make a longer context window and a longer effective context window because we all know that you said like your model can do 1 million context window, but actually it— It just performed very well inside 60K or like 100K.

You can release however size of context window you want, but it's like whether or not it actually works.

Yeah.

Like, do you see, like, like how much— I guess another question that people and we debate a lot is like, how much do you think it's going to be scaling the kind of transformers that we have, which is like making the long context better, like just improving the data, versus if there's like fundamental walls that this is approaching? And kind of like the low-hanging fruit question, like, do you think there's a ton to keep improving it? Is it kind of easy to find the things to do and you just don't have time?

It's not easy. It's not easy. We believe it's the architecture thing. Yeah, data can improve, but it cannot like cross the wall. There is a wall. So we need better architecture, pre-training data, and also post-training data.

Do you think you're starting to hit this wall, or do you just kind of see it coming already? Like, is this like something you're forecasting, or you're like, oh, this specific thing, like data alone is not solving it for us? Because like people in the US that are training these models just don't talk about it. They're like, I don't— I can't say. So I'm like curious. And it's like the models I train are smaller. I think our biggest model is like 30B scale. So it's like when you scale up, you start to see very different limits of what's happening.

But we need to do some experiments. Like, GOM is a 355 billion parameter model, right? But we cannot do experiments with this large model. Like, we need to do experiments with some smaller models, maybe 9 billion parameters or 30 billion parameters, and we test our hypothesis. Like 90% of the time we just failed. Yeah, because experiments, you cannot win every time, but you need to do a lot of scientific stuff to finally get the right answer. Yeah. So if you're talking about whether the Geant4.6 architecture will hit the wall, actually there's a wall. But we need to shift our focus and start from maybe a new architecture or a new framework for doing this stuff.

So it sounds like these bigger runs were kind of like, I don't know, it's not necessarily barely making it, but definitely stressful for you.

Yes, it's stressful, but we are going to use some engineering stuff to try to compress the text windows. to make our users happy because you don't normally need that much. You don't normally need 1 million tokens. Yeah. So if it cannot perform very well, you can compress the context window to 60K or like 30K to make it work.

Zhixuan, you mentioned earlier that you guys, all the inference is abroad, but training is at home. What's behind the rationale of that decision there?

I think the rationale is very simple because we provide services to overseas customers. So I think it's a requirement to store the data overseas, right? So it's a very strict policy for our Zedal AI endpoint. Yeah, we change that privacy policy every month to make it stricter and more coherent to people's expectation. Yeah, but for inference, for the training, I think it's more simple because we don't have many resources. So we only have these resources and we need to utilize it.

But like doing it on, on Nebula or AWS and, you know, Malaysia or Singapore, it's too expensive, it's too slow. You guys already have enough chips at home. What's the, what's the thinking there?

I think it's not, it's not very slow. It's fast because we don't, we not only change the, the location of GPU, but also CPU.

Yeah.

And like database. So if they're all in Singapore, it's still very fast. But if you like have to go back from Singapore to mainland China and then go back to Singapore, it will be slow.

Okay.

But on the training side, it's on the training side in particular.

On the training side, I think it's very simple because we're not open there yet. We're not Anthropic. We don't have to like choose between Amazon and Google and their own infra. So they're doing very complicated stuff. But for us, I think we're still in the initial stage. Yeah. Don't have many like complicated structure with these large inference providers. So things just very simple here. Yeah.

For now.

For now, for now.

Irene or Nathan, any more training questions before Irene wraps us up?

Only sensitive questions that no one— that I don't expect to have an answer to. Like, how big is your next model? How many GPUs do you have? It's like, I don't know, it's not a real question, it's just curiosity.

So for our next generation, uh, we are going to launch 4.6 Air and I don't know whether it's called, it's maybe Mini. It's a 30 billion parameter model. So it becomes a lot smaller in a couple of weeks. And I think that's all for 2025. And for like 2026, we are still doing experiments. Like what I said, I'm trying to explore more, but we are doing experiments on smaller models. So they will not be put into practice in 2.56, but it gives a lot of idea on how we're going to train the next generation. Yeah, we'll see. So when this podcast launched, I believe we already have like 4.6 Air, 4.6 Mini, and also vision model, the next like 4.6 vision model.

Yeah, I guess a good question is like, how long does it take from when the model is done training till you release it? Like, what is your thought process on getting it out fast versus like—

Get it fast, get it fast. Several hours. In several hours. Yeah. So we just open source it.

I love it.

When we finish the training, we do some like evaluation. And after the evaluation, just— so we don't have some arrangements like Like sending the endpoint to LM Arena or to artificial analysis and trying to let it evaluate first and then release the model. We don't have this and we don't have a nano banana thing that trying to make it famous before it's launched because we are very transparent and we believe that like even if you want to open source the model, the open source itself is the biggest event.

So you just want to time it to some specific event or anything?

Yeah, because we're trying to do some marketing thing. From my side, I want to make it longer. We want like a week for me to collaborate with inference providers, like benchmark companies, coding agents, and let everyone trial the model before its release. The forum, the company's perspective, if open source is the most important thing, you need to only prepare for the materials for open source. You need the benchmarks. You need maybe a tech blog. And it's very stressful for me because I need to negotiate with multiple partners within several hours. Like, we have a new model, it's coming in 2 hours, maybe 3 hours, maybe you're sleeping. But this is huge. Yeah, sorry, don't give you enough time for you to like connect to the model or into the integration, but yeah, we're trying to post your tweet afterwards.

Can you talk a little bit about ours? I mean, this was, we now have America, we got our own thing, 002.

What is 0-0-2?

Yeah, what?

Nathan. Midnight to midnight with a 2-hour break. So dumb.

I think hours vary a lot, even inside the company. Like someone would just leave the company at 7 PM, someone will never leave the company. Yeah, for me, I work 18 hours a day because I need to negotiate with, with like US large firm CEOs or the founder of Coding Agents. I need to discuss with Fireworks, with LM Arena, and with maybe KiloCode, their CEOs. So I need to follow their time. And do the meetings maybe at 2 AM, 3 AM. It's all possible.

Oof.

Yeah, but like for our researchers or for the engineering team, I think your brains can only work maybe 8 hours a day. So if you feel tired, you need to get some rest. Yeah, I think. It's impossible to ask a top researcher to work 40 hours a day because that means that you are working really not efficiently or you just attend meetings, right? Because if you join meetings, you can join like 20 hours a day. You just sit here and listen to other people talking. But if you want to read papers and doing experiments and Right, cold, I think 8 hours is enough.

It's very sensible.

My PhD advisor always said that you can do it, like totally change the world if you do 4 hours a day of top technical work. Just like go, go walk in the sun after that. You did a good job.

I guess I'll ask a couple final questions then. I've always been curious. I've always wanted to ask Chinese AI folks this because I feel like the conversation on Value propositions can be really Western-centric. How do you explain the kind of value of your work to, let's say, like kids in high school in Beijing or your grandmother or my work? Yeah. Or the AI's work or the industry's work. How do you explain the value of that to other people, like to kids, to older people in China?

It's hard. It's hard.

I can only say I. do the similar thing like DeepSeek. So we are just a company like DeepSeek. We do the similar thing because DeepSeek is so famous. Everyone in high school and even in kindergarten knows about DeepSeek. For other companies, even Qwin, yeah, let's say Qwin cannot explain itself to kindergarten kids or high school Students. Yeah, because you can say I'm just an alternative of DeepSeek or other stuff, but for like developers, it's simple. We are like one of the best coding LLMs you can find, especially in China. Yeah, but for high school students, they always ask like, we have DeepSeek, like what are you doing? Why we need you? So are you doing a similar thing or are you better? Are you faster? Like, if I'm not using DeepSeek, I have DoBo, I have other apps. Like, why do I need your app? Yeah, that's very tough. So we still need to improve the model performance. I think that's the top priority. Yeah, product is the second, or your experience, the product experience is the second. Yeah, but without a solid model, like nobody will try to pay attention to you because like we are at the same level. Only the most famous one gets all the attention.

So you think the salience of AI models generally in society came straight out of DeepSeek and the kind of Nationalism associated with that?

Yeah, I think there, there is a hype. Yeah, it got so famous even in China, so we got unknown even here. Yeah, I believe that a lot of students in Tsinghua University haven't tried out GLM or even haven't heard of this company. Yeah, because everybody reads news, but not everyone like goes to this building, like to visit Zed AI, right? But like Deepsea is all over the news and social media. So it's really tough to explain over-contribution, over-value, because like we said, we have a gen hook. Model or a gigantic tool use model. So what is tool? Well, what is search?

Yeah.

But we are trying to do more in the future.

Do you think Chinese society is starting to find AI to be more valuable or scarier? Do you have a kind of sense?

Valuable.

Yeah. Because we are not there yet. So AI is not so strong to make people fear about it because there's still hallucination, still like not following the instructions. So all those issues make people feel, oh, it's still very silly for me. Or there's an agent, but it has hallucination. How can I use it?

Yeah.

So still a lot of issues to solve before it gets more fearful or terrifying for people.

So we end every episode with a song. Does Dripu have like a theme song or what do people listen to when they code? Around the office?

No, actually, because our founder loves running. Like, he is a he is a pro in marathon.

Yeah. What's his time? What is he? What is this? What's his marathon time?

Marathon time like below three hours. It's a solid run because because the founder of Morshav. Really loves song, but like our founder doesn't have much interest into song. So for our 4th anniversary, we have a half marathon. Yeah, to celebrate the anniversary. It's crazy.

I gotta go do this. I'm gonna go run the ZAI half marathon next year.

So I have an intern. Um, who did 3 hours and 50 minutes. It's a girl to finish the half marathon. It is crazy. I just wait— I just waited for her at the finish line and like, she's almost dead. It's super crazy. But yeah, we need to work very long hours, so energy is very important.

So no music, just—

You need to be very—

I don't know if this makes you a good boss for waiting or a terrible boss for making her do it in first place. Those interns, man, they gotta earn their slot, show their dedication.

I love it.

She is, she is actually, she is actually the product manager of Zed AI Chat. So she, she built this.

Good.

She earned it. Well, after, after making her do the, the, the half marathon, I'm glad you guys gave her a job at the end.

She ate 2 hamburgers after that. Okay, good, good.

All right, well, this was really fun. Thank you so much for joining the show.

If you're finding value in the show, we'd appreciate it if you'd take a moment to share with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts, which is now part of a16z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI Podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, Check them out and see my endorsement at aipodcast.ing. And thank you to everyone who listens for being part of the cognitive revolution.