@xiaogaifun: Zhu Xiaohu's Real Thoughts on Tencent's AI. Zhu Xiaohu's new interview yesterday had a few points that I still largely …

Zhu Xiaohu's Real Thoughts on Tencent's AI. Zhu Xiaohu's new interview yesterday had a few points that I still largely agree with. Summing them up: First, the model API business is hard to sustain high margins over the long term. Anthropic is currently achieving around 60% margins, mainly because their model's intelligence capabilities are relatively strong, especially in scenarios like Coding, where they've had a clear edge over the past period. But that edge has already started crossing the tipping point. When other models gradually catch up, and prices could be just a tenth of yours, how could you possibly maintain such high premiums forever? Even if Anthropic's model stays ahead for the next few years, if the intelligence capability is only 10% stronger but the price is 50% higher, will users really keep paying for it? Intelligence will ultimately become more and more like electricity and water. For the model API layer, long-term margins might settle at 10% to 20%. That's pretty reasonable. Once capabilities start converging, price wars become inevitable, and everyone will have more incentive to switch to cheaper models. Anthropic's rush to go public now might also be tied to this window of opportunity. Second, the AI office market is definitely much bigger than the AI Coding market. AI Coding emerged first, and a big reason is that engineers already have a high acceptance of AI. Plus, Coding is a scenario that's particularly well-suited to AI—tasks are relatively clear-cut, and results are easy to verify. But AI Coding is just one part of office work. The truly bigger market is the entire white-collar space, which is what everyone's competing over now with AI office tools. Taking it further, AI office tools are essentially general-purpose Agents. The earlier Coding was just a vertical scenario, while office work covers the whole white-collar market. It's like Microsoft's Office back in the day—you'll find that wherever there's a computer, there's Office. Third, the competitiveness of AI office products comes from Model + Harness + Product + Context. In the first half of this year, the industry talked a lot about Harness. But it's 100% predictable that as the Harness layer converges, the ultimate competition will shift to Context. Because the killer app scenarios for AI office tools are work-related—like analyzing data, writing PPTs, and producing reports—and the delivery of those tasks directly depends on how much of the enterprise's work context the AI actually grasps. It's like with a new employee: if you want them to do the job well but don't let them understand the background, no matter how smart they are, it won't help. Context will definitely become a very important moat in this round of AI office competition. Based on these three judgments, Zhu Xiaohu also talked about Tencent's AI development. It was relatively objective. Of course, I don't deny that Zhu Xiaohu has his own limitations—mainly, take a look at his thinking logic and see if it sparks any insights for us in understanding a company and trends. Tencent, in this wave of AI, currently looks like a very typical latecomer that turns the tables. At the start, the Hunyuan model—honestly—barely registered in the industry. But now, with Hunyuan4 preview, it's gradually squeezing into the domestic first tier. In this round of AI office tools, WorkBuddy has actually surged ahead and become a first-tier office Agent app in China. Zhu Xiaohu mentioned that Gold Sand River (Jinzhaijiang) also bought enterprise accounts for WorkBuddy. Why? On one hand, WorkBuddy's product experience is genuinely good. On the other, a very important reason is that Gold Sand River was already using Enterprise WeChat. Enterprise WeChat already has a wealth of沉淀 company org structures, historical communications, plus Tencent Meeting and Tencent Docs—all within the same ecosystem. For WorkBuddy, these are ready-made Context. If it were an independent AI office product, it'd have to reconnect knowledge bases, docs, meetings, and even rehandle org relationships and permissions from scratch. Zhu Xiaohu has another judgment I agree with a lot: Model capabilities are of course still important now, because we're in a phase of alternating leadership. But as long as the gaps between models keep narrowing, technical advantages will be hard to sustain as long-term moats. When that time comes, competition will revert to the classic internet metrics: who controls the entry points, who has user acquisition advantages, who can access historical user data. These are exactly where companies like Tencent have the biggest edges. If you compare last year's ChatBots and AI Coding to this year's AI office tools, you'll see that the competitive focus for AI apps has indeed been migrating upward. The early ChatBots competed on model capabilities—whoever had the strongest model had the best product experience. But with later AI Coding, everyone realized that a strong model alone isn't enough; Harness matters too. Because in long-sequence tasks, how the model calls tools and handles errors directly impacts the final outcome. Now with AI office tools, we're seeing that on top of Model and Harness, there's Context. So I've always thought that general-purpose Agents like AI office tools will probably end up as the domain of big tech giants or superstar startups. Because getting to this level, it's hard to win with just a single capability anymore—the downstream entanglements just keep multiplying.