字号·· | 护眼
theregister

中国瓶盖制造商巧妙戳破AI炒作A Chinese bottle-cap manufacturer brilliantly deflated AI hype

点「原文对照」整页切到原文,或双击某段只看那段的原文。

上海是世界上最大、最繁忙的城市之一,却出奇地安静。电动汽车无声穿行,成群的电动滑板车也悄无声息地跟在后面。11月初我到访上海,参加当地举办的 AI Engineer 大会,发现中国对待 AI 的态度也相当轻松。这并非说大家对该话题兴趣低,只是我没感受到任何“通用人工智能即将到来并将毁灭我们”的焦虑氛围。

Shanghai is one of the world’s largest and busiest cities but is surprisingly quiet. Electric automobiles cruise silently, chased by hushed flocks of electric scooters. When I visited in early November to attend the local edition of the AI Engineer conference, I found China's attitude to AI is also quite relaxed. That's not to say interest in the topic is low, but I didn't perceive any anxious artificial-general-intelligence-is-coming-and-will-destroy-us vibe.

整场大会唯一提及 AGI 的演讲来自杨莹,她是神经科学博士、资深网红,在微博拥有 340 多万粉丝,她热情地谈论为即将到来的变化做好准备的人将迎来的营销与产品开发机遇。对我遇到的中国 AI 从业者而言,AGI 不是 P(doom,毁灭概率),而是一个绝佳的新营销工具。部分原因可能在于,OpenAI、Anthropic 和 Google 等美国主要模型厂商在防火长城的中国一侧实际上无法访问。

The only conference talk that referenced AGI came from Ying Yang, a neuroscience PhD and serious influencer who has over 3.4 million followers on Chinese micro-blogging service Weibo, who enthused about the marketing and product development opportunities opening up to those who prepare for what’s coming. AGI isn't P(doom) for the Chinese AI practitioners I met; it's a great new marketing tool. Some of this may be due to the fact that the major US model makers like OpenAI, Anthropic and Google are effectively inaccessible on the Chinese side of the Great Firewall.

中国工程师使用 DeepSeek、Kimi、GLM、MiniMax、MiMo 或 Qwen,这些模型现在都已足够胜任几乎任何任务,且完全没有提及《终结者》。这并非说中国工程师和设计师没有面临我们在其他地方看到的自我价值危机。陆锐(Ryo Lu)的一场精彩主题演讲发问:“当我们什么都能做时,我们还知道自己想要什么?”当我们无法再通过写了多少行代码、设计了多少框架来证明自己是谁时,我们的自我价值衡量标准是否变得毫无意义?

Engineers in China use DeepSeek or Kimi or GLM or MiniMax or MiMo or Qwen, all of which are now more than good enough for almost any task, and all offered without reference to The Terminator. That's not to say that Chinese engineers and designers aren't facing the same crisis of self-worth we've seen elsewhere. A spectacular keynote from Ryo Lu asked, "When we can do anything, do we still know what we want?" Are our measures of self-worth pointless when we can no longer prove who we are by how many lines of code we've written, or frameworks we've designed?

陆锐给出了一个我近来越来越常听到的观点:“我想构建工具,帮助我越来越深度地参与其中,而不是证明我越来越不被需要。”随后的演示展示了这种工具已变得多么强大,以及 AI 工程师能用它做什么。我看到 AutoArk 令人印象深刻的演示:一个 350 亿参数的模型在 iPhone 16 Pro 上仅占用 2.5GB 手机内存即可运行。

Ryo landed on a line I'm hearing more frequently: "I want to build tools to help me be more and more involved, not prove that I'm needed less and less." What followed showed just how good that tooling has become and what an AI engineer can do with it. I saw an impressive demo from AutoArk of a 35 billion-parameter model running in just 2.5GB of smartphone RAM on an iPhone 16 Pro.

DYNA的沈中展示了该公司如何构建能够与每月更新的AI模型配合工作的机器人硬件。我最喜欢的演讲来自一家瓶盖制造商的朱世辉,他提醒在场所有人:“最后一公里是半个旅程。”朱抨击了那一类从未意识到某些制造设备属于前数字化时代、却坚持认为每个流程都能轻松实现自动化的顾问。别管那些工厂的工人缺乏深厚的数字技能(朱的员工平均年龄在40多岁末),也别管这类企业利润微薄,根本没有资源进行彻底的转型。

DYNA's Zhong Shen showed how it builds robot hardware that will work with AI models updating every month. My favorite presentation came from a bottle cap manufacturer called named Zhu Shihui, who reminded everyone in the audience that “the last mile is half the journey.”Zhu flayed the class of consultants who never realize some manufacturing equipment is pre-digital but still insist every process can be easily automated. Never mind that workers in those factories lack deep digital skills (Zhu's average employee is in their late 40s) and that companies of this sort operate on margins so thin that resources simply aren't available to mount a complete transformation.

朱建议,与其挥舞AI这把“破坏球”强行介入,最好的方法是怀着谦逊的态度,挖掘组织的隐性知识,并以此为基础进行转型。

Instead of swinging in on an AI wrecking ball, Zhu suggested the best approach is to exercise humility by seeking out an organization's tacit knowledge and using that as the basis for transformation.

朱离开讲台不久后,一位同行的澳大利亚人注意到会议上女性的比例很高。我当时没留意,因为在工程会议上男女比例相当对我来说感觉完全正常。然而,这在我参加过的澳大利亚或美国的AI活动中从未发生过。无论中国在AI方面还在做什么,他们在AI工程领域已经把性别政治抛在脑后了。

Not long after Zhu left the podium, one of my Australian fellow travelers noted the significant proportion of women at the conference. I hadn't noticed, because having equal numbers of men and women at an engineering conference felt perfectly normal. Yet it's not something I've ever experienced at AI events in Australia or America. Whatever else the Chinese are doing with AI, they've put gender politics aside in AI engineering.

为了说明背景:离开上海不久后,黄爱丽丝成了全球重大新闻。一个在AI领域工作的上海女性,在一个只见过男性从事该领域的美国人眼中可能显得可疑。李飞飞、米拉·穆拉蒂和丹妮拉·阿莫代伊作为例外证明了这一规则,但尽管她们成果卓著,知名度都远不及男性同行。在中国,情况不同。

To put that in context: not long after leaving Shanghai, Alice Huang became major global news. A Shanghainese woman working in AI might look suspicious to an American who only ever sees men working in the field. Fei-Fei Li, Mira Murati and Daniela Amodei stand out as exceptions that prove the rule, but none of them, despite stellar work, have anywhere near the profiles of their male peers. In China, it's different.

对中国在AI方面的进展嗤之以鼻已变得司空见惯,归因于抄袭、“蒸馏”乃至彻头彻尾的窃取。然而,中国制造出了最好的小模型和最好的机器人,并正越来越接近让它们良好地协同工作。

It's become commonplace to be dismissive of Chinese progress in AI, ascribing it to copying, "distillation" and outright theft. Yet the Chinese make both the best small models and the best robots, and are moving ever-closer to making them work well together.

中国实验室与同行及竞争对手分享研究成果:几乎每一个模型,无论闭源还是开源,都受益于深度求索和阿里巴巴的研究工作。务实、平衡与开放似乎对它们大有裨益。或许,在OpenAI步Anthropic后尘走向公开市场、股东获得话语权之后,稳定、理智与一份宁静将飘洋过海而来。这些正是我期盼的“中国进口货”。®

Chinese labs share their results with their peers and their competitors: nearly every model, closed or open, has benefited from the research done by DeepSeek and Alibaba. Practicality, balance and openness seem to be serving them well. Perhaps, after OpenAI follows Anthropic out to the public markets, and shareholders get a voice, stability, sanity and a bit of quiet will cross the Pacific. Those are some Chinese imports I'm looking forward to. ®

阅读媒体原文 ↗