演讲者首先将二战后的建筑热潮与当前的“AI建筑热潮”进行了类比。二战后，迫切的建造需求导致现代主义的简洁几何形状和单色调被广泛复制，但往往缺乏最初的意图或对语境的理解。这导致了“僵尸建筑”的出现——普遍、无差别的结构，对用户或其周围环境漠不关心，这与早期更精心设计的建筑（例如象征信任和安全的银行）形成了鲜明对比。 如今，AI建筑热潮也带来了类似的警示点。首先，大型语言模型（LLMs）擅长提供“最可能答案”，反映的是流行或过去有效的东西，而不是提供原创的或特定于语境的解决方案。一个例子是一家韩国烤肉品牌的网站，它虽然功能齐全，但缺乏特色和语境，显得像软件一样普通。其次，存在着“完成的诱惑”。AI让事情感觉“非常非常快”地完成了，但往往是过早的。演讲者将其比作“微波炉卷饼”困境：执行速度可能导致人们忽视最终产品中的严重缺陷。正如微波炉卷饼“几乎无法食用”，但由于速度快而被迅速消费一样，AI生成的界面可能看起来很精致，但可能无法解决实际问题或提供差异化。演讲者强调，“表面上的精致状态往往具有误导性”。 第三个警示点是，AI生成的工作因为如此轻松和快速，常常让人感觉是“一次性的”，并被如此对待，导致缺乏长期责任感和维护。这三个警示点共同威胁着创造“僵尸UI”——单调、空洞、无人照管的数字体验，类似于战后建筑的失败。鉴于人们“一半的清醒时间都在看屏幕”，强烈需要“真正被精心照料的软件”，展现个性和精细的细节，例如GrokBot动画或显示正确日期的日历选项卡。这些例子表明开发者关心用户并预见他们的需求。 为了应对这些挑战，演讲者提出了四条建议，教你如何用AI打造“有心有魂”的产品： 1. **拥有观点：** 如果一个组织不定义其品牌、用户需求和价值观，AI将提供一个通用、落后的观点。这种观点构成了标准的基础，尤其在分布式构建时至关重要。Stripe将“乐观主义”融入每个细节的承诺就是一个例子。一个关于AI生成广告的设计评审讨论强调了产出质量高于创造方法的重要性，最终产生了17条改进意见，并创造了“百事冒泡”（Pepsi bubbling）一词来形容精益求精的工艺。发展这种观点需要“非常擅长注意到”用户需求、世界信号和更广泛的灵感。 2. **将标准编码到机器中：** 由于在分布式和代理式构建中人类不会参与每一个决策，因此标准必须直接嵌入到系统中。这就是为什么设计系统正当红的原因，但它们必须从仅仅是组件发展到能够扩展“意图”而非仅仅是“一致性”的“完整模板和流程”。古腾堡（Gutenberg）的详细排版系统被用作类比，说明了如何创造出既可扩展又具有足够鲜明观点，让人感觉像是手工制作却由机器生成的东西。Stripe基于其设计系统构建的CLI，通过适时地吸收文档，使AI“更加顺从”，就是例证。 3. **拒绝将“完成”与“优秀”混淆：** 产品开发中传统的“质量过滤器”已基本消失，因为AI允许一周内构建20个想法。这使得关键角色转移给“编辑”，由他们在构建后进行评估，确保产品真正解决问题、符合用户思维且具有连贯性。这在于辨别某个东西是否“完全成形”，并推动其完成，而不仅仅是说“是”或“否”。演讲者引用Nabil Qureshi的话，指出“意想不到的细节”和“更深层次的意义”如何使艺术变得伟大，而AI常常错过这一点。Stefan通过AI开启“可能性空间”实现了动画的56次迭代，这说明了精细编辑的力量。 4. **释放创造力和艺术性：** 尽管AI可能制造单调，但它也是“我们有史以来最伟大的创意催化剂”。在一个拥挤的市场中，差异化是关键，而当前的界面远非“卓越交互的典范”。AI能够发明新的界面和美学，就像多点触控和合成器所做的那样。为了让AI成为更强大的创意伙伴，演讲者建议通过增加特异性和独特兴趣来改进输入，添加原始材料，并通过超越最初的“完成”阶段来“压榨你的输出”，甚至可以使用“对抗性代理”。在文化上，领导者必须提供探索空间并“保护奇特之处”，利用AI低廉的创作成本节约下来的资源来打造“真正特别”的产品。 总而言之，演讲者认为意图性、强烈的观点、编码化的标准、严格的编辑和释放的创造力，对于避免数字领域的“僵尸建筑”至关重要。目标是将这次AI建筑热潮变成一场“创意复兴”，创造出“更强大、更能展现制作者的匠心和关怀”的产品，就像约翰·拉斯金（John Ruskin）对哥特式建筑的品质和工艺的赞美一样。用户被打动的是解决了问题以及预见他们需求的“巧妙之处”，而不仅仅是速度或复杂性。

The speaker opens by drawing a parallel between the post-World War II building boom and the current "AI building boom." After WWII, the urgency to build led to the widespread copying of modernism's simple geometries and monochromatic palettes, often without the original intentionality or understanding of context. This resulted in "zombie buildings" – generic, undifferentiated structures that showed no care for users or their surroundings, contrasting sharply with earlier, more thoughtfully designed architecture like banks signaling trust and security. Today, the AI building boom presents similar watch points. Firstly, Large Language Models (LLMs) excel at providing the "most probable answer," reflecting what is popular or has worked before, rather than offering original or context-specific solutions. An example given is a Korean barbecue brand's website that, while functional, lacks character and context, appearing generic like software. Secondly, there's the "temptation of done." AI makes things feel finished "really, really fast," often prematurely. The speaker likens this to the "microwave burrito" dilemma: speed of execution can lead one to overlook serious flaws in the final product. Just as a microwave burrito is "nearly inedible" but quickly consumed due to speed, AI-generated interfaces can seem polished but may not solve the actual problem or offer differentiation. The speaker emphasizes that an "apparently polished state can often be misleading." The third watch point is that AI-generated work, being so easy and quick, often feels "disposable" and is treated that way, leading to a lack of long-term responsibility and maintenance. These three watch points collectively threaten to create "zombie UI" – monotonous, vacant, and uncared-for digital experiences, akin to the post-war architectural failures. Given that people spend "half our waking lives looking at screens," there's a strong need for "software that feels actually cared for," showing personality and meticulous detail, like the GrokBot animation or a calendar tab displaying the correct date. These examples demonstrate builders caring for their users and anticipating their needs. To navigate these challenges, the speaker offers four recommendations to use AI for building products with "care and soul": 1. **Have a point of view:** If an organization doesn't define its brand, user needs, and values, AI will supply a generic, backward-looking one. This point of view forms the basis of standards, especially crucial when building is distributed. The example of Stripe's commitment to "optimism" infused into every detail is cited. A design crit discussion regarding an AI-generated advertisement highlighted the importance of output quality over the method of creation, leading to 17 bullet points of improvements and the coining of "Pepsi bubbling" for meticulous craft. Developing this point of view requires "getting really good at noticing" user needs, world signals, and broader inspirations. 2. **Encode your standards into the machine:** Since humans won't be involved in every decision with distributed and agentic building, standards must be embedded directly into the systems. This is why design systems are having a moment, but they must evolve from just components to "full templates and flows" that scale "intent" rather than just consistency. Gutenberg's detailed typesetting system is used as an analogy for creating something both extensible and opinionated enough to feel handmade despite being machine-made. Stripe's CLI built on its design system, making AI "far more obedient" by consuming documentation at the right time, exemplifies this. 3. **Refuse to confuse done with good:** The traditional "quality filter" in product development is largely gone, as AI allows 20 ideas to be built in a week. This shifts the crucial role to an "editor" who evaluates post-build, ensuring products actually solve problems, are attuned to user thinking, and are coherent. It's about discerning if something is "fully formed" and pushing for completion, not just saying yes or no. The speaker quotes Nabil Qureshi on how "unexpected details" and "deeper meaning" make art great, which AI often misses. The example of 56 iterations of an animation by Stefan, enabled by AI opening "the possibility space," illustrates the power of meticulous editing. 4. **Unleash creativity and artistry:** While AI can manufacture monotony, it's also "the greatest creative catalyst we've ever had." In a crowded market, differentiation is key, and current interfaces are far from the "epitome of great interactions." AI enables inventing new interfaces and aesthetics, just as multi-touch and synthesizers did. To make AI a stronger creative partner, the speaker recommends improving inputs with specificity and unique interests, adding source material, and "stressing your outputs" by pushing beyond the initial "done" stage, even using "adversarial agents." Culturally, leaders must provide room to explore and "protect the strange," using the savings from AI's low cost of creation to build "truly special" products. In conclusion, the speaker argues that intentionality, a strong point of view, encoded standards, rigorous editing, and unleashed creativity are vital to avoid "zombie buildings" in the digital realm. The goal is to make this AI building boom a "creative renaissance," creating products that are "more powerful and show the hand and care of the maker," much like John Ruskin's admiration for Gothic architecture's quality and craft. Users are impressed by problems solved and "clever touches" that anticipate their needs, not just speed or complexity.