Opus 4.8 Exposed for 'Distilling' Chinese Models; Zhipu's Market Value Briefly Surpasses Xiaomi; ByteDance Opens 'Doubao Shares' to Seed Employees, Developing Custom CPU | AI Weekly Report Opus 4.8被曝“蒸馏”中国模型；智谱盘中市值一度超小米；字节向Seed员工开放“豆包股”，正开发定制CPU｜AI周报

Anthropic recently secured its Series H financing at a $96.5 billion valuation, teetering on the edge of the trillion-dollar club. However, its latest flagship model, Opus 4.8, has ironically "self-reported" in API tests, claiming to be either Tongyi Qianwen or DeepSeek. This scenario is deeply ironic: a high-profile company that waves the banner of "safety and ethics" and has publicly accused Chinese companies of "industrial-scale distillation attacks" is now seeing its own model exhibit outputs that reflect a severe "identity confusion." While some defend this by pointing out that the web interface works fine, the API remains the core interface for developers and app building. It’s like someone who claims absolute purity but fills their private diary with others' secrets behind closed doors. Either the training data contains contaminants it shouldn’t, or there are inexplicable flaws in the model architecture or fine-tuning process. Either way, it’s a crushing blow to the valuation summit Anthropic has just reached.

When we discuss "distillation," we are essentially delving into the blurry, sensitive, and double-standards-laden ethical line in the AI arms race. American companies are quick to point fingers at each other, but when similar issues potentially arise within their own ranks, the logic of defense often becomes remarkably flexible. This reveals a cold reality: on the path to AGI, the construction of technical barriers is frequently accompanied by the monopolization of discourse power, and the so-called "rules" are often tools used by the powerful to restrain latecomers.

Meanwhile, Jensen Huang’s comments on Huawei’s "τ Law" represent another savvy narrative. He praised it as a "remarkable breakthrough," while subtly framing it as a "practical technology" that won’t impact TSMC. This mirrors how a top master comments on a challenger’s moves: first acknowledging the ingenuity of your innovation, then calmly pointing out that it still falls within my expectations and experience. Die stacking and 3D packaging are indeed advanced pathways, but TSMC’s decade-long cultivation in these areas—its process know-how, massive production capacity, and client ecosystem—forms a much deeper moat. Huang’s "non-disruptive" assertion is less a technical analysis and more a display of strategic composure, aimed at reassuring the supply chain and clearly signaling to the world: the underlying rules of the game haven’t changed for now.

Viewed together, these two events illustrate the dual nature of today’s AI and chip industries: one side features technical "bloopers" and ethical double standards beneath valuations and accolades; the other shows traditional giants adopting a composed, defensive-attack stance in the face of disruptive innovation. From Alibaba’s salary reform and ByteDance’s internal equity incentives to the clarification of rumors about Meituan’s layoffs and Unitree’s push for the STAR Market, beneath the bustling surface lies the collective anxiety and actions of various players as they continuously adjust their postures, solidify their foundations, and attempt to seize the entry point to the next wave in a rapidly changing environment.

Ultimately, whether it’s a model’s identity crisis or the offensive-defensive rhetoric around chip laws, it all points to the same core: in the noisy era nearing the technological singularity, strength and discourse power must be forged in tandem. You can raise trillions, you can release the most cutting-edge models, but whether the foundational technology is solid and the pathways withstand scrutiny—these are verdicts that the two most ruthless judges, the market and time, will deliver in due course. For those of us immersed in this landscape, perhaps we should maintain a clear-eyed perspective: while marveling at valuation myths, don’t forget to glance at the next line of model output; while dazzled by breakthroughs in chip laws, also examine where the true levers of ecosystem disruption lie.

Anthropic刚以9650亿美元估值拿下H轮融资，离万亿俱乐部就差临门一脚，其最新旗舰模型Opus 4.8转头就在API测试里“自报家门”，声称自己是通义千问或DeepSeek。这出戏码讽刺得很：一家高举“安全与伦理”大旗、曾公开发声明指控中国公司对它进行“工业规模蒸馏攻击”的明星公司，自己的模型却在底层出现了这种堪称“身份认同混乱”的输出。尽管有人辩称网页端对话正常，可API才是开发者和构建应用的核心接口，这就像一个声称自己绝对纯洁的人，却在关起门来的私人日记里写满了别人的秘密。要么是训练数据混入了不该有的成分，要么是模型架构或微调环节出了难以解释的纰漏。无论哪种，都像一记闷棍，打在了Anthropic刚刚筑起的估值巅峰上。

当我们在讨论“蒸馏”时，本质是在讨论AI军备竞赛中那条模糊、敏感且充满双重标准的伦理线。美国公司们互相指控起来毫不手软，但当类似问题可能出现在自己身上时，辩护的逻辑往往变得异常灵活。这揭示了一个冰冷的现实：在通往AGI的赛道上，技术壁垒的构筑往往伴随着话语权的垄断，而所谓的“规则”常常是强者用来约束后来者的工具。

与此同时，黄仁勋对华为“τ定律”的点评，则是另一种高明的叙事。他称赞其为“亮眼突破”，同时轻描淡写地将其框定为不会冲击台积电的“实用技术”。这像极了顶级高手对挑战者招式的点评：先大方承认你的创新很巧妙，再从容指出它仍在我的预判和经验范畴之内。晶粒堆叠、3D封装确实是先进路径，但台积电在这些领域的十年深耕所形成的工艺Know-how、庞大产能和客户生态，构成了更深厚的护城河。黄仁勋的“不会冲击论”，与其说是技术分析，不如说是一种战略定力的展示，旨在安抚产业链，明确告诉世界：游戏的底层规则，暂时还没变。

这两件事摆在一起看，构成了当下AI和芯片产业的一体两面：一面是估值与光环下的技术“穿帮”和伦理双标，另一面是传统巨头面对颠覆性创新时那份从容不迫的防守反击姿态。从阿里的薪酬改革到字节的内部股权激励，从美团的裁员传闻澄清到宇树科技冲刺科创板，热闹表象之下，是各家在快速变化的环境中不断调整姿势、巩固基本盘、试图抓住下一个浪潮入口的集体焦虑与行动。

最终，无论是模型的身份错乱，还是芯片定律的攻防话语，都指向同一个核心：在技术奇点临近的喧嚣期，实力和话语权永远需要同步锻造。你可以融资万亿，可以发布最前沿的模型，但基础技术是否扎实、路径是否经得起推敲，市场和时间这两位最无情的裁判，迟早会给出比分。而对于我们这些身处其中的人来说，或许更应保持一份清醒：追捧估值神话的同时，别忘了看看模型输出的下一行是什么；惊叹定律突破的闪光点时，也需审视其真正撬动生态的支点在哪里。