176: 姚顺雨，来到腾讯300天

1. Key Themes

DeepSeek as the Catalyst for Big Tech's AI Awakening

DeepSeek's release was the pivotal moment that forced Chinese tech giants to recognize their organizational deficiencies. Seeing a young, inexperienced team produce a world-class model triggered a strategic rethink across the industry.

"When DeepSeek came out, Tencent's senior executives realized that a team with not much experience, full of young people, could actually produce such a powerful model. That's when they realized their own team was no longer young — they were mostly older folks with backgrounds in search, advertising, and recommendation systems." [00:09:15]

Yao Shunyu's "Organization-First" Leadership Philosophy

Yao Shunyu is universally described not as a technical visionary but as an organizational leader — a surprising trait for a 27-year-old researcher. His core belief is that most problems in building LLMs are organizational, not technical.

"He's not the type who is very technically oriented. He values organization a great deal. The thing he says most often internally is that when building large models, the organization is extremely critical — many problems are organizational problems at their root." [00:35:38]

The "Window of Trust" as the Most Valuable Resource for Reformers

The podcast surfaces a non-obvious insight: a new leader's greatest asset is the period of maximum boss trust — and the smart move is to execute the most aggressive changes in that window, not gradually.

"The smartest reformers will always do the most aggressive things during the period when the boss trusts them most, rather than slowly consuming half a year of patience and then doing things. The more time passes, the greater the resistance, unless you can continuously produce results to keep extending the boss's patience." [00:43:53]

Tencent's Federated Structure is Both its Moat and its Obstacle

Tencent's long-standing "federal system" — high autonomy per business group — is directly in tension with the unified command required to compete in AI. Martin Liu Zhiping reportedly leans toward unification, but organizational inertia resists it.

"Tencent is like what people call the American federal system — each business group has very strong autonomy. There's no requirement to use Hunyuan or anything like that. For example, the gaming division is doing its own world model, and Hunyuan is also doing 3D and world models. There's no mandate from above." [00:11:00:38]

WeChat's Privacy Paranoia Creates a Structural Barrier to AI Integration

Zhang Xiaolong's (Allen Zhang's) extreme sensitivity to user data means WeChat's VLM team trains its own model from scratch rather than using Hunyuan — an enormous resource duplication that reflects WeChat's cultural identity more than technical necessity.

"Allen Zhang's attitude roughly means he doesn't view Hunyuan as a third party per se — it's not like DeepSeek or Kimi — but he is extremely protective of WeChat's user privacy data, so he still prefers to build his own model for more control." [00:01:04]

Hunyuan's Rebuild: From 2,000 Borrowed GPUs to Full Overhaul

The starting condition of Hunyuan was dramatically worse than outsiders understood — the team began with just 2,000 GPUs borrowed from the advertising department, an organization staffed primarily by search/recommendation engineers, and no native AI research culture.

"Hunyuan started with only 2,000 GPUs, and those 2,000 GPUs weren't even Hunyuan's own — they were borrowed from the advertising department. The ad department had GPUs because they were also doing large models for ads, using the older BERT-style system for ad recommendation and search." [00:15:17]

The "Second Half" Thesis: Context-Rich Platforms Win the Next Phase

Yao Shunyu's published "Second Half" blog articulated his reason for joining Tencent: the next phase of AI competition is about finding the right problems to solve, and that requires companies with rich, diverse real-world context — which he identified as Tencent and Meta.

"In his view, Tencent and Meta are actually the two companies with the most critical or the most context. Tencent has massive social data, office scenario data like WeCom and Tencent Docs, meeting data from Tencent Meeting, search data from QQ Browser and WeChat Search, content data from Tencent Video and QQ Music, and even gaming. So across every major category, Tencent has a top-tier application." [00:19:37]

The Activation Rate Problem is the Hidden Ceiling for Enterprise AI

B2B AI tools face a structural problem: companies buy licenses for all employees but actual daily usage rates are very low, creating a gap between GMV and real value creation.

"The activation rate is the issue — a large company purchases your product for 10,000 employees, but the percentage actually using it daily is not high. Looking at this activation rate, or at the C-side payment situation, it's actually not that optimistic right now." [01:30:53]

ByteDance's AI Reputation Runs on Historical Momentum, Not Current Model Performance

Despite Seed (ByteDance's AI team) not having a breakout model in text, coding, or multimodal, investors and founders still back ByteDance-origin people more readily — a halo based on past company excellence, not current AI standing.

"ByteDance alumni have an easier time fundraising, regardless of whether they have a research or product background. I think that comes from recent historical inertia — ByteDance was China's strongest tech company over the past five to eight years, with the strongest monetization and organizational strength." [01:23:36]

The "Coding to White-Collar Worker" Gap Is the Next Inflection Point

The AI adoption curve from programmers to general knowledge workers is the key unresolved question for the industry's commercial viability, and the speed of that transition will determine whether valuations hold.

"From programmer to white-collar worker to the broader professional C-side user — that path is still quite long, or hasn't reached an inflection point yet. And sometimes it's really hard to judge when an inflection point arrives." [01:31:22]

---

2. Contrarian Perspectives

Tencent's "Diverse Data" Advantage May Be Illusory — Business Units Won't Actually Share

While Yao Shunyu chose Tencent for its theoretically rich cross-domain data, the reality is that each business unit guards its data as its most valuable proprietary asset and will only share after Hunyuan proves concrete value to them first.

"Before, most departments wouldn't give their data to the large model team. WeChat definitely won't give Hunyuan its data — not now. Tencent Video's content involves copyright that must be purchased. The cooperation model is still individual project-based: 'I'll give you this batch of data only for this specific project, not for general pretraining.'" [00:23:57]

A Young Leader Without Deep Technical Sharpness Running a Frontier AI Lab Is Unusual — But May Be Exactly Right

Conventional wisdom says a frontier AI lab needs a deeply technical leader. Yao Shunyu's peers note he doesn't give sharp technical direction, yet this may be the correct approach for reorganizing a large corporate AI team where organizational dysfunction, not technical insight, is the binding constraint.

"Researchers who have had contact with both Liang Wenfeng and Yao Shunyu would evaluate the difference in styles: Liang Wenfeng goes quite deep into technical details and is like... purely technical. Meanwhile researchers tend to feel Yao Shunyu is not as technically sharp as they expected." [00:37:35]

Tencent's Federal Autonomy — Long Seen as a Strength — Is Now Actively Harmful for AI Competition

Tencent's celebrated decentralized culture, which enabled WeChat and Honor of Kings to flourish independently, is now a structural liability when facing vertically integrated competitors like ByteDance that can unify all data and resources.

"Internally they've discussed this — Tencent's previously celebrated way of working was cross-department collaboration, each team contributing its strengths. But taken to an extreme, that's 'extreme distributed leadership,' and when facing a company that is highly centralized and top-down with a fast pace, you need unified command, not coordination of resources and interests." [01:12:58]

The "Who Will Survive at the Table" Consensus About ByteDance's Seed Is Based on Vibes, Not Evidence

The widely held belief that Seed will definitively remain a top-tier AI player is not supported by current model performance — it may simply reflect the tech world's short memory bias toward recent historical winners.

"Everyone can easily say who has the best chance of success, who will stay at the table. But where does that basis come from? I genuinely don't know where this feeling comes from. I think one source is inertia from recent history — ByteDance was China's strongest internet company for the past five to eight years. That's a very emotional source." [01:23:36]

Pretraining Duplication Between Hunyuan and WeChat's VLM Is Pure Resource Waste — But Won't Be Fixed Due to Culture

From a pure efficiency standpoint, Tencent should train one foundation model and let WeChat fine-tune on top of it. Instead, they're running two full pretraining efforts. The reason isn't technical — it's cultural and political, reflecting WeChat's identity as an autonomous fiefdom.

"At the pretraining stage, Tencent could really just do one. From a technical efficiency and resource usage standpoint, just one. You're all in the Tencent system, the data is still under your own control. But this is just... Tencent. Each business group has very strong autonomy and Tencent wouldn't mandate everyone use the same thing." [01:10:01]

---

3. Companies Identified

Hunyuan (混元) / Tencent Large Model Team

Tencent's internal large language model team, rebuilt under Yao Shunyu's leadership. Mentioned as the central subject of the episode — a team undergoing dramatic organizational transformation, replacing most key positions across pretraining, post-training, evaluation, and infrastructure.

"Pretraining, post-training, evaluation, infra — basically all turned over. All recruited by him personally... Infra people came from Seed Infra: Xiao Xuefeng, Zhang Chi, Huang Qi — all from Seed Infra, which is considered the strongest infra in the entire industry." [00:27:52]

DeepSeek

Chinese AI lab known for building world-class models with a young, lean team. Cited as the single event that woke up Tencent and ByteDance leadership to the possibility that young teams could beat incumbents.

"When DeepSeek came out, the Tencent executives realized an organization with not-so-experienced, all-young-people could make such a powerful model." [00:09:15]

Kimi (Moonshot AI)

Chinese AI startup known for its frontier models. Kimi K3 was released July 17th and cited as a strong competitive release that directly pressured Hunyuan's positioning.

"Kimi K3 was released last Friday, July 17th — its performance was impressive, the feedback was quite positive." [00:05:13]

On the technical side, Kimi is highlighted as having contributed meaningfully to the broader Chinese model ecosystem:

"On the optimizer used in V4, Kimi made important improvement contributions. V4's residual connection improvement MHC's original version came from ByteDance Seed's HC, and later Kimi was inspired by MHC and developed Tension Residues — though only proposed this spring, it's already being used in the latest K3." [01:38:38]

Seed (ByteDance AI)

ByteDance's internal AI research team. Consistently cited by industry observers as likely to "stay at the table" despite not having a breakout model in text or coding. Mentioned as Hunyuan's key competition for talent — specifically, Hunyuan's pitch to recruits is that they can lead a team, rather than being a "screw" in Seed's large machine.

"Compared to Seed, Hunyuan has the advantage: there are still open positions here. If you join me, you can lead a team, or do more of what you want — you don't have to be another cog at Seed." [00:31:17]

Qianwen / Tongyi (阿里千问)

Alibaba's large model. Cited as having genuine open-source community reputation and real adoption, giving Alibaba credibility that Tencent currently lacks in AI model benchmarks.

"Alibaba's Qianwen — it may have a strong reputation in the open-source community and real call volume, it has credibility." [01:24:35]

WeChat VLM Team (微信VLM团队)

WeChat's independent large model team that published a 258B MoE model in January. Described as a parallel Tencent AI effort separate from Hunyuan, focused on serving WeChat's 1.4 billion users within severe compute and privacy constraints.

"They published a 258B MoE model in January, slightly smaller than Hunyuan 3. They haven't done an external public release — they use it internally in WeChat features like Xiaowei Agent and search." [01:08:01]

Workbody (腾讯)

Tencent's enterprise office agent product under CSIG. Described as the most tightly integrated product with Hunyuan, providing training data for the model.

"Workbody is very important — much of Hunyuan's training data comes from the Workbody side. So these two teams should have the closest relationship right now." [00:57:24]

Yuanbao (元宝)

Tencent's consumer-facing AI chatbot. Gaining traction inside WeChat through the comment section @Yuanbao feature, which drives real user engagement with summarization and Q&A.

"Users love asking Yuanbao things inside news articles — having it summarize the article, explain the news. They can see this is genuinely improving their product's user activity. So they are now very willing to share some of their data for collaboration." [00:24:54]

Anthropic / Claude Code

Cited as the model/company that triggered the Chinese industry's pivot to coding as a strategic priority in the second half of 2024.

"When Anthropic suddenly rose, which was in the second half of last year — around the time of Claude 4.6 — those companies were all already pivoting to coding." [01:32:20]

Zhipu AI (智谱)

Cited as one of the Chinese model companies that made an early strategic commitment to coding as a major direction from the first half of 2025, with positive results.

"Zhipu roughly from the first half of 2025 started making this a particularly important direction, and Kimi too. And you can see these companies have recently had relatively good success in coding." [01:33:19]

---

4. People Identified

Yao Shunyu (姚顺雨)

97-born (age 27-28), formerly at OpenAI, now head of Tencent's Hunyuan large model team. Remarkable for being an organizational thinker at his age — swiftly replaced all key positions within two to three months of joining, recruited top talent from Seed and Microsoft Research Asia, and manages upward to senior Tencent leadership effectively.

"He's actually a pretty shrewd person — he knows that during the period when the boss trusts him most, he should rapidly replace everyone in the key positions. Pretraining, post-training, evaluation, infra — basically all turned over." [00:27:52]

"He'll often go find the youngest researchers and interns, ask them for feedback, and the feedback they raise can very quickly reach the relevant business leads and trigger adjustments." [00:38:33]

Liu Zhiping / Martin (刘志平)

Tencent President, Yao Shunyu's most important backer inside Tencent. Controls budget, hiring authority, GPU allocation (through his oversight of both the strategy line and ad revenue line), and the critical resource of "time" — i.e., how long he'll protect Yao from performance pressure. Personally visited Silicon Valley to recruit frontier researchers.

"Martin would go to Silicon Valley himself to meet frontier scientists, researchers at Meta and Google, hoping they'd come to Tencent. And now Martin directly participates in their interviews himself — a change from before, when interviewees would only get as far as VP level." [00:10:35]

Liang Wenfeng (梁文峰)

DeepSeek founder. Referenced as the exemplar of a deeply technical leader — the contrast case to Yao Shunyu's organizational approach. Researchers who've met both note the sharp stylistic difference.

"Liang Wenfeng still goes quite deep into technical details... and he actually probably doesn't want to manage other things, though he now has to." [00:37:35]

Zhang Xiaolong / Allen Zhang (张小龙)

WeChat's creator. His extreme user privacy sensitivity is the primary reason WeChat refuses to share data with Hunyuan or use it as a foundation model. Reportedly had direct conversations with Martin about the Hunyuan/VLM question.

"Allen Zhang's attitude roughly means... he is extremely protective of WeChat's user privacy data, so he still prefers to build his own model for more control." [01:09:30]

Tan Xu (谭旭)

Joined Tencent Hunyuan in late 2024, formerly from Microsoft Research Asia (speech/multimodal direction), previously briefly at Kimi. One of the first external research hires to signal Tencent's shift toward recruiting frontier talent.

"The first person was from the Microsoft Research Asia side — Tan Xu. He'd also come from Microsoft. And LateTalk also broke the story of him joining Kimi before that. He didn't stay at Kimi very long." [00:15:47]

Feng Jiashi (冯家时)

Joined Tencent Hunyuan, also from Microsoft Research Asia background. Part of the wave of external research talent brought in under the new strategy.

"And Feng Jiashi also came from Microsoft Research Asia... Peng Houwen also joined around late 2024, also from Microsoft Research Asia, also the multimodal direction." [00:16:17]

Peng Houwen (彭后文)

Joined Tencent Hunyuan in late 2024, from Microsoft Research Asia (multimodal direction). Recently reported on social media to have left Tencent.

"Peng Houwen also joined in late 2024, also from Microsoft Research Asia, also the multimodal direction... Recently there's been news on social media that he has left." [00:16:44]

Liu Huidan (刘慧丹)

Heads pretraining at Hunyuan under Yao Shunyu's new team structure. Brought in as part of the leadership overhaul.

"Including Pretraining's Liu Huidan." [00:28:49]

Xiao Xuefeng, Zhang Chi, Huang Qi (肖学峰, 张池, 黄启)

Infrastructure leads recruited from Seed Infra — ByteDance's infrastructure team, considered the strongest in the industry. Their recruitment to Hunyuan is a signal of the seriousness of the rebuild.

"Infra, from Seed Infra came Xiao Xuefeng, Zhang Chi, Huang Qi — these are all Seed Infra, because Seed Infra is considered possibly the strongest infra in the entire industry." [00:28:21]

Tang Daosheng (汤道生)

Head of Tencent's CSIG (Cloud and Smart Industries Group). Stepped forward to consolidate all of Tencent's AI-native consumer products under CSIG when no one else would take responsibility.

"At that point, it was CSIG's Tang Daosheng who stepped up. Actually nobody wanted to take it on. Then he stepped up, and basically all the so-called native AI products were consolidated into CSIG." [00:53:57]

Lu Shan (卢山)

President of TEG (Technology and Engineering Group), Yao Shunyu's formal reporting line manager. Actually primarily oversees the Microsoft Research Asia (multimodal) portion; GPU allocation decisions run through him.

"GPU allocation at Tencent is basically decided by two people: TEG's Lu Shan, and David Lin Jinghua — who is also a senior executive and Tencent's head of strategy." [00:42:28]

David Lin Jinghua (林景华)

Tencent's head of corporate strategy and a senior executive. Co-controls GPU allocation decisions alongside Lu Shan. Both strategy and advertising lines report up to Martin.

"The strategy line and the advertising line both go up to Martin — do I need to say more?" [00:42:56]

Dai Zihang (戴子航)

Former Grok researcher. WeChat VLM team reportedly reached out to him this year as a potential candidate to lead their large model effort — indicating WeChat is looking for external AI talent despite its historically insular culture.

"People like Dai Zihang from Grok — did WeChat actually try to recruit him? They've been in contact, around this year. They're looking for someone who could independently lead the entire large model effort." [01:02:18]

Wu Yonghui (吴永辉)

Based at Google, experienced industry researcher with a research background. Mentioned as an example of the type of senior industry-research hybrid talent that Tencent eventually learned it needed.

"They later found Wu Yonghui, who has been in the industry many years at Google — a research background. Including Yao Shunyu, who was at OpenAI, the best company in this field." [00:17:41]

Zhang Yiming (张一鸣)

ByteDance founder. Referenced for his philosophy that product success belongs to the system, not the individual — explaining why ByteDance has no "father of Douyin" but Tencent has a "father of WeChat."

"Zhang Yiming believed from the start that a product's success is not any one person's success, but the company's success, the system's success. So at ByteDance you won't see a 'father of Douyin' or 'father of Toutiao.'" [01:26:34]

Robin Li / Lu Qi (陆琪)

Former Baidu President. Used as a cautionary tale — his bold initial reforms at Baidu succeeded in strategy and org redesign, but the failure to account for cultural compatibility led to mass exodus of senior Baidu executives.

"When his ideas were actually implemented in practice, what followed was a mass exodus of core Baidu business executives. That shows your reorganization didn't account for cultural compatibility within the existing organization." [01:19:43]

---

5. Operating Insights

Act Maximally During the Honeymoon Window — Pace the Boss's Patience, Not Your Own Comfort

The most actionable leadership insight: a new leader's political capital is highest on day one and declines unless continuously replenished with results. The correct strategy is front-loaded aggression, not gradual reform.

"The smartest reformers will always do the most aggressive things during the period when the boss trusts them most. Don't slowly consume half a year of patience and then act. As time passes, the resistance to doing things only gets greater — unless you can continuously produce results to keep extending patience forward." [00:43:53]

Build Your Information Source at the Bottom of the Hierarchy, Not the Top

Yao Shunyu's habit of directly seeking feedback from the youngest researchers and interns — bypassing management layers — serves two purposes: it gives him unfiltered ground truth about organizational problems, and it builds loyalty with the people who execute.

"He often goes to find the youngest researchers and interns and asks them for feedback. The feedback they raise can very quickly reach relevant business leads and trigger adjustments — things like 'this data platform is hard to use.' He can very quickly surface these issues. I think this is a manifestation of his organizational instinct — your information source matters." [00:38:33]

Replace Key Positions Immediately, Then Leave Displaced Incumbents Harmless

Rather than a mass layoff (which generates backlash) or gradual displacement (which is slow), Yao's approach was surgical: swap all critical roles within two to three months, but allow incumbents to remain in the company with no specific mandate, letting them self-select out over time.

"Yao Shunyu's approach is interesting: you can stay here, you just can't touch the model — do whatever you want otherwise. So it wasn't a large-scale layoff or optimization. Most people stayed at Tencent. But they were marginalized, and gradually transferred or self-selected out." [00:28:49]

Use Empty Org Chart Slots as a Recruitment Weapon Against Larger, More Established Rivals

Hunyuan's rebuild created a structural advantage in talent competition: every senior role was open, allowing recruits from established labs to immediately run their own team rather than remain a specialist contributor at a mature organization.

"Compared to Seed, or Tianwen, Hunyuan's advantage is: there are still open positions here. When you join, you can be a leader, do more of what you want — you don't have to be a screw at Seed." [00:31:17]

---

6. Overlooked Insights

Tencent's RL Platform Could Be the Industry's Most Underappreciated AI Infrastructure Asset

Yao Shunyu is quietly building a company-wide reinforcement learning platform designed to enable every Tencent product to run post-training fine-tuning using their own business data. This is far more strategically significant than it sounds: it is the mechanism by which Tencent's otherwise-inaccessible data silos become a genuine training advantage. If it works, it solves the core problem Yao identified when choosing Tencent over other companies.

"He's building a reinforcement learning platform that the whole company can use — he hopes Tencent's various products can all do post-training on this platform, making the linkage between business and model more organic... Currently it hasn't been rolled out at large scale. The seed users are Workbody and Yuanbao, since they're most tightly integrated with CSIG." [00:49:37]

This platform, if it succeeds, would be the first time Tencent's legendary data diversity actually flows into model training — which is the entire thesis behind why Yao chose Tencent. Nobody in the conversation highlighted this as the crux of whether Yao's "Second Half" strategy actually delivers, but it is.

The Chinese AI Ecosystem Has a Cooperative Undercurrent That Contradicts the "War" Narrative

Buried in the closing remarks is a striking data point: Kimi contributed the optimizer improvements used in DeepSeek V4, which inspired Kimi's own Tension Residues technique, now already in K3. ByteDance Seed contributed the original HC residual connection concept. Three competing companies effectively co-developed critical architectural innovations across open research. This cooperative dynamic — competitors building on each other's published work — is systematically underreported and may be a structural advantage of the Chinese open-source model ecosystem versus the more closed Western approach.

"On the optimizer used in V4, Kimi made important improvement contributions. V4's residual connection improvement MHC — the original version came from ByteDance Seed's proposed HC, and later Kimi was inspired by MHC and developed Tension Residues. Though only proposed this spring, it's already being used in the latest K3." [01:38:38]