Google DeepMind Yao Shunyu's insights, big tech updates, China field trips
candid interview with a Google DeepMind researcher on why the AI stack is heading toward cloud economics, Alibaba and Tencent AI updates and everyone's in China
Yao Shunyu interview takeaways
A Google DeepMind researcher talks candidly about what frontier AI work actually looks like, why individual genius matters less than people think, and why the AI stack may end up looking a lot like cloud computing.
As you probably know by now, I’m a fan of Zhang Xiaojun’s podcast (张小珺). It’s one of the best long-form conversation shows for researchers and tech leaders in Chinese, with episodes lasting 3 to 4 hours each, which is both the appeal and the problem. I’ve been listening to her interviews on and off for years, and I'm so glad to see it gain mainstream traction more recently. I usually listen selectively because, not going to lie, it’s hard to squeeze in 4 hours of concentrated time. But this episode with Google DeepMind’s Yao Shunyu (姚顺宇), his last character being ‘universe’ vs the Tencent Yao’s ‘rain,’ 姚顺雨, I’d actually recommend sitting through if you can. He was a very candid speaker, less polished than many business executives, but there was a sense of earnestness and humbleness from him, and most importantly, a rare glimpse into what frontier AI research actually looks like from the inside.
Quick note on the name before I get into it, because it’s genuinely confusing. There are two well-known “Yao Shunyu” in AI right now. Both went to Tsinghua and actually around the same time. One is the former OpenAI researcher who joined Tencent last year and now leads its model effort as chief AI scientist. The other one, the protagonist today, studied physics at Tsinghua, did theoretical physics at Stanford, joined Anthropic during the Claude coding and large-scale RL push, and then moved to Google DeepMind. They’re in fact good friends and used to “just hang out” when they were both in Silicon Valley. What are the odds? Honestly, even people in the industry mix them up sometimes.
For personal preference reasons or company policy reasons, frontier researchers almost never talk this openly. In a humbling way, maybe a byproduct of cultural upbringing, or just his candid view of the world, when asked what he thinks he has contributed to the projects he’s worked on, he said if it wasn’t him, it would have and could have been anyone else. At this point in life, only a handful of people can be credited for the progress of the technology, whereas the majority are just part of the natural evolution. Things would have become how they are anyway, maybe not with the exact same timing or the same people, but the direction was already there. If it weren’t him working on large-scale RL for coding, someone else would have. If it weren’t Anthropic, another company would have figured out coding agents. If it wasn’t Boris, Claude Code might not have looked the same, but someone would still have realized the next product after a chatbot is an agent harness (OpenClaw was developed independently, Gemini apparently had internal teams building similar things).
His metaphor was, AI is a wave. Whether you surf it or not, it still hits the shore.
Once the paradigm is found, history starts driving people, and not the other way around.
individualist contribution
We usually talk about AI in terms of individuals, right? Tech moguls, star researchers, transfer rumors, who left OpenAI, who joined Meta, who’s at Anthropic now. Better yet, who has Chinese tech companies courted recently? In many ways, it feels like celebrity sports league stories now, and much of the focus has been shifted to a few names and the speculated drama. (Rui has something amazing cooking here, a tracker of all the Chinese researchers)
Yao makes the opposite claim. He says that for LLMs, once the paradigm was found, the work became more industrialized. Pre-training has a paradigm. Post-training and RL have a paradigm. The industry now broadly knows what direction to go, and that was in many ways echoed by Chinese researchers too in February; they used the ‘copying homework’ analogy then, that’s all. Different labs have slightly different techniques and approaches in their trial and error process, different infrastructure capacity, different data buyouts (which you can question; they may be somewhat the same at this point), and different tastes, which is supposedly what really makes the difference now, but the general direction itself is no longer a mystery.
He argues that the transformer architecture needed the pioneers to invent it—the decision to really bet on pre-training scaling needed heroes. At every paradigm shift, key people and key teams still matter a lot. But once the paradigm is established, it starts to feel less like people are driving history and more like history is driving people. I found that framing quite eye-opening, given how much is mystified about the industry.
He basically said that AI research doesn’t necessarily need that much “brain’ anymore. What matters more now is being reliable, careful, hardworking, and responsible. In many ways, it’s about the ability to execute meticulously, which many may argue is why there seems to be a dominant percentage of Chinese researchers in the field. This wasn’t the first time I’d heard this, but I’d always assumed it was genius-talking, the kind of thing brilliant people say to sound modest. A researcher from one of the leading Chinese labs told me earlier that the way to train LLMs is already quite streamlined and that even interns, who make up the bulk of their workforce, can do it now. Kevin Xu has confirmed this, too. But the first time I heard it, I genuinely thought they were just flexing. But the reality is that many of these Chinese labs have very small teams that are relying on smart architecture and quietly upending the traditional playbook of how to train a model. The training itself is streamlined. It’s finding the next breakthroughs in efficiency and capability that requires out-of-the-box thinking, and that’s truly within the capabilities of a much smaller group of people.
‘clean’ and ‘responsible’
The next paradigm shift may still be in the hands of a few people, like one researcher at Microsoft told me, “There are a few of them, where the industry just orbits around them.” But where we are now is a mature system.
Throughout the interview, he kept emphasizing “干净” and “责任'’ which means clean and responsible, to the point that even got the interviewer to ask for clarification. Paraphrasing what he said, he explained, the difficulty is that doing the simple things cleanly, at scale, inside a giant system, is very different from academic research.
In academia, you care about your own results. You want the cleanest paper, the best individual contribution, the most impressive result you can claim. But in a frontier lab, the unit is no longer the individual paper. The unit is the whole model training system.
A researcher can find a method that looks amazing on a small experiment, but if it only works because of a better data mix, more sampling FLOPs, different constraints, or something that will not transfer to the real production run, then it may not actually help the company. In that case, the honest thing is to say: this is interesting, but it is not useful enough for the big system.
But that is a very difficult task, because it goes against human nature. Everyone wants credit. Everyone wants the result that looks good. Everyone wants to say, “this was my contribution.” But in scaled AI research, personal heroism can easily become system-level noise.
He said that a good frontier-lab researcher is not just someone who can make a metric go up. It is someone who understands whether the metric going up actually matters and what the limitations are, and has the integrity to stay honest.
Why coding worked first
The second big shift Yao described is that the game has moved from “can we make the model better” to “what should we build with the model.” And he said something that honestly surprised me: among OpenAI, Anthropic, and Gemini, no one is really worried that they can’t catch up on model capability, they’re all somewhat aware of what to expect next. Rather, they have committed to different use-cases, and what everyone is trying to figure out now is what’s the next valuable task? Can it be well-defined? Can you build the right training environment and feedback signal around it?
Coding is the perfect example, and it's what almost every lab ushered to do in its herd-like behavior. It didn’t work just because models got smarter. It worked because the task itself is unusually trainable. The reward signal is clean (code runs or it doesn’t, tests pass or they don’t). GitHub provides a vast repository of high-quality data. His point is that good programmers have surprisingly aligned tastes, and they broadly agree on what clean structure, good abstraction, and maintainability look like, but what makes something a new breakthrough is through quite literally trial and error, and sometimes a bit of serendipity.
That’s where Claude Code, Codex, and Cursor come in. The task of coding is a combination of good data, clean feedback, professional users who are willing to pay, and a product workflow where agents can do economically valuable work. What actually made it work was model, plus harness, plus workflow, plus feedback loop.
LLM business is looking more like a cloud provider model
‘Base models are IaaS. Agent harnesses are PaaS. AI apps are SaaS.’ Okay, so this is the framework from the interview that I think is potentially most useful for investors.
If Yao is right that frontier labs will eventually have similar model capabilities, with differences more about timing, implementation, taste, and productization than about some fundamental intelligence gap, then the base model business starts to look less like a pure technology monopoly and more like cloud infrastructure.
Think about it. Cloud is also extremely capital-intensive. Only a few companies can do it on a global scale. But AWS doesn’t win because it runs “cloud” in some magical way that Azure or GCP fundamentally can’t understand. The moat is elsewhere entirely: scale, cost structure, custom chips, data centers, enterprise distribution, developer mindshare, and all the PaaS and SaaS layers built on top over time. (You can apply that same thinking to ByteDance, Alibaba, Tencent, Huawei, and the SOE cloud providers) I wonder what cloud expert JS Tan think about this.
I think the LLM world could end up looking similar, where base models become the new IaaS. Agent harnesses (Claude Code, Codex, the orchestration layer) become PaaS. AI applications built on top become SaaS. The analogy isn’t perfect, but you get my point.
If the base model becomes infrastructure, then selling raw tokens is not a great business by itself, and Yao basically said this. API token resale becomes a price-war business, which is why I’ve been saying I’m not sure what the long-term strategy for Chinese labs is if they keep shipping 'nearly as good but very cheap models’, it feels like a short-term solution, and thus why the hyperscalers (if they can catch up in model capabilities or swallow the labs) then would have a huge advantage as they have the full stack. Now, I should flag here that Yao works at Google DeepMind, and this framing happens to favor Google’s structural position pretty conveniently. If inference becomes a commodity utility, having your own TPUs and a lower cost structure lets Google profit where others can’t.
For OpenAI and Anthropic, the strategic question is different. They can’t be model API companies forever. They need to move up the stack. That’s why memory, agents, workflow orchestration, and enterprise integration all matter. In the cloud, the money didn’t stop at compute; it moved into databases, observability, security, identity, developer tools, and business applications. In AI, I think the same thing happens. The model layer may be huge, but it may not be where all the margin stays.
What Yao is working on now
He said he’s focused on two things, and both are worth paying attention to.
The first is AI closing the loop on AI research itself. Models already write most of the code for researchers (Yao said a conservative estimate is 90% of his code is model-generated, and for experiment implementation, his speed is something like 20 to 50x what it was a year and a half ago, which is kind of wild). But the loop isn’t fully closed yet. The next step is for AI to define the problem, design the experiment, write the code, run it, evaluate the result, figure out what worked and what didn’t, propose the next hypothesis, adjust the experiment, and run it again. The full research loop. Yao thinks this can start becoming real in the next 6 to 12 months as the entire experiment cycle compresses. It becomes a recursive loop, much like Anthropic CFO Krishna Rao was describing on Invest Like the Best, too, and that happens, the pace of everything else accelerates, too.
The second is long-horizon memory. Or as he put it, train with finite context, use as infinite context. Everyone wants AI workers, but a real digital worker needs memory. It needs to keep track of long-running tasks, know what to remember and what to forget, retrieve the right information, and maintain state over time. Without that, a chatbot can never actually become a worker. Yao’s point is that humans don’t have infinite context either, as we forget most things, but we remember important things, at least we try to. So maybe the answer isn’t literally infinite context windows. Maybe it’s context management, selective memory, and retrieval capabilities, and actually knowing what to ‘forget’. I actually think this is one of the most commercially important ideas in the whole interview, because if it gets solved, you’re suddenly talking about something much closer to an actual digital employee than what we have seen commercially available so far.
In terms of model progress, he sees no end to progress and believes that features that feel differentiated today may become default model behavior tomorrow. The question for every app company becomes: what do you own that compounds faster than the model improves? How do ‘wrappers’ survive? I think maybe vertical industry knowledge and know-how is still something defensible here.
On China
On China, he prefaced a few times saying that he himself has not worked within a Chinese lab, so it’s all observations but not from direct experience. But surely with how small the circles are, he knows what’s up, or at least is quite tuned in. Yao’s view aligns with much of what I’ve been thinking (and writing about). The model gap has narrowed, but China’s compute constraints may force genuinely different methods. Less compute means Chinese labs have to be more efficient. They have to distill, reuse, open-source, optimize, and move faster through product loops.
Yao said that there is ‘dumb’ distillation and ‘smart’ distillation. ‘Dumb distillation’ is just taking another model’s output and training on it. That’s not a moat. It may even be a sign the lab doesn’t really know what it wants. ‘Smart distillation’ is more interesting: using another model as a judge, having your own model generate answers while another evaluates them, combining multiple models into a broader synthetic data and evaluation pipeline. That starts to look like real multi-model training, and Yao even suggested Chinese labs could become pioneers here.
Not all distillation is the same, and this part was pretty funny because Yao apparently named specific labs doing each kind, but the names were beeped out in the podcast. However, he did say Doubao is not distilling for a fact - hence why everyone seems to revere them secretly. If anyone listening caught who he was referring to, please let me know.
The other big China-US difference, as always, is product. The US has this massive enterprise productivity profit pool, so everyone naturally focuses on coding, workflow, and professional software. The math is quite clean in that sense. I help you write code, it costs me $150, and I charge you $200.
China has always been better at complicated consumer products that monetize indirectly. The product might look free and unprovocative. But then it monetizes through ads, commerce, games, live streaming, payments, recommendations, or some loop that only makes sense once it’s already spinning. If you’ve followed AI Proem, you know this is something I come back to a lot, because I genuinely think China’s AI consumer path is going to surprise people. The US may produce the best enterprise agents. China may produce the first weird AI superapp that looks like a toy until it suddenly isn’t.
In other news
Alibaba reported its March Quarter 2026 earnings yesterday, with cloud growth as the main highlight, though investors are questioning whether the heavy AI spend has a clear strategic path forward.
Business highlights:
Cloud Intelligence Group’s external revenue growth accelerated to 40% this quarter as enterprise adoption of AI services deepened.
AI-related product revenue delivered triple-digit year-over-year growth for the eleventh consecutive quarter and accounted for 30% of Cloud’s external revenue.
Customer management revenue grew 8% year-over-year on a like-for-like basis.
We also touched on the fact that Alibaba opened up Qwen app to Taobao this past week as part of its push into agentic shopping, but it’s still unproven whether this becomes a sticky habit or fades as a fad. My personal view: utility-purpose goods may well move to agents, but for discretionary spending, the choosing and selecting is part of the experience and the joy.
Tencent just announced that users can forward chat history into Yuanbao, and users are again asking why Yuanbao isn’t embedded directly into the WeChat interface. The answer remains: security.
Beyond that, I’m watching Alibaba’s Accio for AI merchant support, and Tencent’s WeChat Mini Programs. There are now reportedly AI-powered features that help merchants build mini programs within WeChat (think Cursor-lite), letting businesses easily plug into WeChat Pay, marketing, and distribution.
China (AI) visits in vogue
And on China, everyone seems to be in China these days. With the obvious Trump visit to Beijing, AFP interviewed me about my thoughts on how AI will fit into the conversations.
The thoughtful Rui Ma wrote a marvelous analysis of the state of physical AI and robotics in China after a week-long tour of factories and companies. I’ve been brewing on a piece on physical AI for a while, too, but this is a banger to showcase the reality of where we are at, which is nowhere close to mass consumer adoption yet. On the industrial side, robotics have been making more strides in adoption, I believe.
The @SAIL team seems to have had a fruitful China trip and I was glad to see Jasmine Sun’s article noting that partly why the Chinese public doesn’t seem to be as anti-AI is because the labor force makeup is different, knowledge workers simply make up a much smaller percentage of the Chinese workforce in comparison. I think that’s an acute observation and something I’ve been sharing too. Kevin Xu’s article gave a rundown of the cultural differences among the labs, which we’ve also touched on in the deep dives but this was an updated fun read. Azeem Azhar’s sharp economic analysis breaks down the potential backfired strategy of the export controls, stating that US export controls have left Chinese AI labs roughly 2-3 years behind in deployed compute (and an estimated 8x gap in capacity), yet Chinese frontier models trail the US by only 6-8 months on benchmarks. That implies Chinese labs are extracting 4-7x more capability per unit of compute than naive scaling would predict.
That efficiency edge is flowing through to inference economics: DeepSeek, Kimi, GLM, and Qwen are priced 3-28x below comparable US models while reportedly maintaining 50-70% gross margins, and they dominate the open-source and on-device tier. The implicit thesis is that the controls have backfired strategically, and by forcing ruthless efficiency, they’re cultivating exactly the capabilities (cheap inference, small-model performance, algorithmic ingenuity) that will matter most as AI scales. A very interesting take.
Last but not least, Nathan Lambert came away with some humbling observations about the Chinese AI ecosystem and how maybe open source, once again, proves to be the way to propel an industry forward. In his words, “the Chinese system is designed around quickly learning from your peers and avoiding double-spending research compute — or infra effort” as the labs compound on top of each other’s efforts.
He added, “The Chinese labs, through incredibly thorough technical reports and intentional knowledge sharing across labs, effectively are de-risking ideas for their peer companies to not necessarily need to invest as many resources in.” When we talked casually offline, he said that he also felt some sense of what I wrote here about how DeepSeek is beginning to serve the whole sector as a base technological foundation. So I’m very excited to have him on AI Proem’s Differentiated Understanding podcast next week to unpack more. So stay tuned!
Oh and two other good reads this week that were insightful/ enjoyable: Viola Zhou’s The Chinese whiz kids of Silicon Valley in Rest of World and Yi-ling Liu’s op-ed The Shared Feeling of Being Harvested by the Future in the NYT.
Next week, I’m in Singapore speaking at SocGen’s annual macro investor event, so no deep dives will likely be published. Thanks for your understanding. Toodaloo~
"GitHub provides a vast repository of high-quality data. His point is that good programmers have surprisingly aligned tastes, and they broadly agree on what clean structure, good abstraction, and maintainability look like"
Which is probably why AI-generated code is still mostly crap and 47+ percent insecure.
Because that's why human programmers produce.
People are slowly starting to wake up that AI-generated code is both a software apocalypse waiting to happen - both in terms of maintainability and security - and an exposure of how badly written most human code actually is.
This will have profound effects on the software industry - just not those everyone has been assuming over the last two or three years.