高洪浩
Key Views & Dialogues
179: The Distillation Storm: A Technology Race No One Is Discussing in Public
o1 and DeepSeek R1 improved distillation economics by turning reasoning traces into training material and shipping six distilled models. Zhang Yiming’s ban on Seed distillation trades a possible short-term lag for a deeper AGI moat; Anthropic’s naming, behavioral fingerprints and possible Entity List exposure make compliance and frontier demand the larger unresolved variables.
View summary (7)
- Distillation is not copying answers in the literal sense; it is “copying homework with a barrier to entry.” The standard approach is to use a powerful model such as Fable or Claude—its answers, chain of thought and even full Agent task trajectories—to train a student model to reproduce its output behavior, which requires compute, data pipelines and know-how. Even where a practice does not qualify as conventional distillation, 洪灏 notes that using a frontier model’s output to improve a competing model violates the broad user agreements of Anthropic, OpenAI and Google DeepMind; legally, “it is currently a gray area.”
- Two technical milestones improved distillation’s return on investment, and the scale expanded sharply from 2025 onward. The September 2024 release of o1 showed that large-scale reinforcement learning in post-training could teach models reasoning strategies, while chain of thought became “better raw material for distillation”; in January 2025, DeepSeek R1 shipped 6 distilled models at once—4 based on Qwen 2.5 and 2 on Llama 3. Combined with open-source models such as K3 closing in on US closed-source models, distillation has moved to the center of the China-US technology race.
- 张一鸣’s ban on Seed distillation was a real strategic bet, with clear logic and costs. He believes distillation “can improve model performance in the short term, but is fundamentally copying Claude’s existing capabilities,” meaning it can at most approach, not surpass, the original; Manqi believes Seed 2.1, released in June, likely used some distillation before an internal debate led to a decision to stop. Meanwhile, reporting says ByteDance began reorganizing its data team in June and plans to create a first-level department, “AI Data and Safety,” parallel to Seed and Flow. The sharpest argument is about the organizational cost: “张一鸣 may not be the most technically knowledgeable person, but he is probably the person who understands human nature best.”
- The escalation from naming names to countermeasures is turning distillation into a compliance tail risk. Anthropic named DeepSeek, Kimi and MiniMax in February; in a June letter to the US Congress, it called Alibaba Qwen’s 28.8M interactions “the largest-scale distillation attack ever uncovered,” and said it had deployed traffic classifiers and behavioral fingerprinting systems capable of detecting cross-account coordination and chain-of-thought extraction. Locally deploying an open-source model approaching 3T parameters for distillation also requires substantial compute and electricity, which Manqi says makes the activity relatively easy to monitor. Inclusion on the Entity List is the potential cost 洪灏 poses as a hypothetical.
- Whether a student can surpass its teacher remains technically open, but staying ahead over time depends on innovation. Multi-teacher distillation can “take the best from multiple sources,” and a student can surpass a teacher on a specific task; but “right now, everyone leads for a few months,” and if self-evolution works as assumed, the first model to reach a given state will accelerate faster and “throw the people behind it even farther back.” Staying close through distillation is not enough reason to declare victory.
- The hypocrisy debate leaves behind a useful case-law boomerang. Anthropic was sued in a class action by writers for downloading books from pirated sites and reached a $1.5B settlement, while a California court precedent held that training on paid-for material was not infringement. 洪灏 asks whether, under the same logic, using Anthropic’s output to improve another company’s model would also be non-infringing. “This is a pretty interesting case.”
- The bigger variable than distillation may be a mismatch between intelligence supply and demand. On the “kill line” chart, every model below and to the right of V4 and Grok 4.6 is eliminated; one productivity-app founder says users can have “8 out of 10 tasks solved by DeepSeek V4 Flash,” raising questions about both the scale and pace of white-collar demand for more intelligence: “We used to say coding was a bigger market than video generation—what if we were wrong?” That would shake the commercial logic of frontier investment more deeply than short-term valuation volatility.
176: 姚顺雨, 300 Days at Tencent
The defining move in 姚舜宇’s first 300 days after parachuting into Tencent’s Hunyuan was not a model launch but an organizational purge. Durin…
View summary (7)
- The defining move in 姚舜宇’s first 300 days after parachuting into Tencent’s Hunyuan was not a model launch but an organizational purge. During the first 2-3 months, when he was at the peak of the boss’s trust, the heads of pre-training, post-training, evaluation, and infra were “basically replaced” with people he recruited himself; the entire infra team came from ByteDance’s Seed infra, regarded as the industry’s strongest. 洪灏’s maxim is worth remembering: “The smartest reformers do the most radical things when they have the boss’s greatest trust.” Tencent’s boss may be patient for “2-3 years.”
- Hunyuan 3, a roughly 290B MoE, met expectations rather than surprising anyone. It was internally positioned as a small model: “The core goal is to get the whole organization working together and restore team morale.” A preview in April was used to manage expectations, with the full release moving up one level. 姚舜宇’s own assessment is clear-eyed: Hunyuan 4 “may not truly enter the first tier” either; a genuinely influential model will come later. The catch-up narrative—“surpassing Qwen is only a matter of time”—is being sustained through release pacing and expectation management.
- 姚舜宇 chose Tencent based on the argument in his blog The Second Half: the next phase will be about “defining the problems truly worth solving,” which depends on a large company’s context. In his view, Tencent and Meta have the most context; ByteDance’s 800M DAU is offset by “homogeneous user profiles,” while an ecosystem built around recommendation and distribution was passed over.
- The context thesis has collided with reality: WeChat still has not given Hunyuan access to its data. 张小龙 is “extremely concerned” about WeChat’s private user data and would rather build a controllable 258B model in-house, making it likely that the 2 foundation models will coexist for the long term. With Chinese companies operating at least an order of magnitude behind overseas leaders in compute, whether Tencent’s “federal system” can pull together is Martin’s central question. Internally, the metaphor is Black Hawk Down: Delta Force executes the capture perfectly, the Rangers responsible for coordination drop the ball, and the mission ends in disaster.
- The highest-increment deployment initiative is a company-wide reinforcement-learning platform. Hunyuan is assigning post-training teams of 2-3 people to each business, with Work Body and Yuanbao as seed users; much of the data used to train Hunyuan will come from Work Body. Linking the model layer with the application layer is the second battle after the organizational overhaul.
- The C paradox is the episode’s sharpest investment debate. Everyone assumes ByteDance’s C team will stay in the game because it has the best resources, talent, money, and infra, yet it has not produced the kind of reputation attached to Kimi, DeepSeek, or Qwen; in coding, it “may not even be at the top of the second tier.” 曼祺’s explanation is that this confidence is merely “a memory of short-term historical momentum,” an essentially emotional source. The observable signal is that people leaving ByteDance raise money more easily.
- Coding commercialization is running into cold water. Enterprises buy by seat, but actual seat utilization is low; “whether you look at selling by seat or at C-end payments, neither is particularly optimistic.” The growth slope from programmers to white-collar users remains “very low.” 洪灏 worries coding will become SaaS—slow growth, weak revenue, and no winner-take-all dynamic—and asks whether “the imagination space in China is simply quite limited.” So far, commercialization has only been validated in coding and video generation; meanwhile, “since 2023, every autumn people have said the bubble was about to burst,” and it has happened 3 times.
162: Batch-Produce Creative Assets, Use Models to Screen Influencers | Talking AI-Era Global Marketing with 沈晨岗 of Feishu Shenno Meetsocial
China’s 2025 overseas expansion is shifting from products toward brands as rising traffic costs, homogenization, and tighter compliance narrow low-end competition. Feishu Shenno’s agentic flow scales creative production, but human judgment still gates high-risk budgets across its 100,000-plus customers.
View summary (8)
沈晨岗认为，2025 marked a clearer inflection point as Chinese companies going global shifted from products to brands, though the transition was not overnight. Rising traffic costs, homogenized competition, and tougher quality and compliance standards across the media and traffic ecosystem have closed off the path of competing ever more aggressively on price and execution. Companies must move beyond immediate sales toward user research, product iteration, and brand mindshare; ultimately, what goes global is not just a good, product, or brand, but “a company with soul and vitality.”
In Feishu Shenno’s D-MES model, Digital is not a parallel capability but a multiplier on sales, product, and brand capabilities. Digitalization improves the efficiency of payments, supply chains, and delivery while also raising the speed and quality of decision-making. 沈晨岗’s competitive view is that the faster a company reads the market, the more accurately it understands change, and the more effectively it acts, the more likely it is to win.
Marketing’s biggest AI consensus is that it is important and should be deployed now, while the biggest disagreement is what it can actually solve. The controversy over McDonald’s Netherlands’ AI-generated holiday ad showed that being able to generate an image or video does not mean being able to deliver commercial creative. If content does not drive likes, registrations, orders, or purchases, what is the point of making it merely for the sake of making it?
AI has materially lowered the cost and barrier to content production, but it still struggles to independently own high-value, low-tolerance-for-error budget and media-buying decisions. An optimizer makes hundreds or thousands of decisions every day about what to keep, what to cut, bids, media, and market budgets. AI currently contributes information from the back end and is gradually moving toward the front end; handing high-risk customer-service or media-buying judgments to AI means a single error could have fatal commercial consequences.
Feishu Shenno breaks creative work into agents for requirements understanding, strategy planning, image or video production, and quality control, with humans orchestrating them into an agentic flow. A creative professional who once produced 5–10 assets a day can now deliver hundreds or even thousands in standardized scenarios. During the Spring Festival, one short-drama client received 6,000 edit concepts in a week, while daily ad spend rose from several thousand dollars to the low six figures.
After mass generation, the scarce capability is judging and improving commercial performance, not generation itself. 沈晨岗 says that “a 1% gap in quality can ultimately translate into a performance gap of more than 10%.” AI can assign hundreds or thousands of tags to creative assets, but it cannot independently create the underlying framework for judging quality; people still need to abstract knowledge, define methods, and teach the model through feedback.
A service provider’s potential moat is not a general-purpose model, but reusable infrastructure, cross-client data, and a digital loop tied to customers’ businesses. Feishu Shenno says its capabilities can be spread across more than 100,000 customers, while its systems capture more than $20M in daily advertising-spend data. Fixing a client’s website or data-collection infrastructure could produce a 10%–20% improvement, far above the 1%–2% typically fought over at the media-buying layer—the effect is “multiplicative, not additive.”
沈晨岗 expects marketing over the next 3–5 years to evolve from standalone ad buying into the digital nervous system of the enterprise. User viewing, liking, registration, purchase, and repurchase will feed back into products, orders, and supply chains, while AI lowers the barrier to global marketing. But the first-mover advantage will depend on whether organizations can genuinely rebuild their processes: “We all think it is hard; others may find it even harder.”