内容来源：https://news.microsoft.com/signalmagazine/issue/issue-02/#slam-dunk

内容总结：

微软《信号》杂志深度报道：洞察、创新与人文关怀的科技叙事

微软最新一期《信号》（Signal）杂志以多元视角，探讨了在信息过载时代如何智慧决策、人工智能如何赋能个体创造力、网络安全战线的创新斗争、下一代互联网的变革愿景、硬件设计的人文思考以及科技如何重塑体育体验。这些故事共同勾勒出一幅科技与人文深度交融的图景。

一、智慧沟通：在“全知”与“克制”间寻找平衡
微软首席传播官弗兰克·肖以北欧神话中牺牲一只眼睛换取智慧的神王奥丁为喻，反思现代企业面临的沟通困境：在数字时代试图关注每一条信息并急于回应，反而可能因“史翠珊效应”放大负面声音。真正的智慧并非全知全能，而在于审慎判断——知道何时行动、何时沉默，通过长期建立信任，让发声更具分量。《信号》杂志的使命正是帮助读者在信息噪音中辨别真知，而非追逐每一个热点。

二、AI赋能：从濒临失明的恐惧到珠宝创作的热忱
虚拟现实先驱贾伦·拉尼尔分享了一段个人经历。一次误诊为即将失明的经历，激发他借助AI工具深入珠宝制作。他将AI视为“人类协作的最高效形式”，而非独立实体。通过具体、 grounded（基于事实）的提问方式（例如询问其他工匠的具体经验），他能有效利用AI学习技艺、解决问题，同时避免其“幻觉”。拉尼尔强调，AI的价值在于增强人的自主性与创造力，而非取代。作品的终极意义源于创作者自身的故事与情感。

三、网络安全：以创新法律手段打击“利器”滥用
2021年，爱尔兰公共卫生系统遭勒索软件重创，攻击者使用了经过破解的合法安全测试工具“钴击”。微软数字犯罪部门（DCU）没有止步于传统防御，而是与软件原厂商Fortra及全球健康信息安全组织Health-ISAC携手，发起了一项开创性的法律行动。他们通过民事诉讼，结合商标法、版权法乃至反勒索法，成功获取法院命令，查封用于分发恶意软件的域名与服务器。此次行动使非法“钴击”服务器数量下降72%，存活时间从49天大幅缩短至16天，展现了通过跨部门合作与法律创新主动打击网络犯罪生态的有效路径。

四、未来网络：CTO凯文·斯科特展望“智能体网络”
微软首席技术官凯文·斯科特将当前AI带来的变革，类比为上世纪90年代早期互联网的开放与创新浪潮。他提出了“智能体网络”的愿景：AI智能体将能代表用户理解目标、执行复杂任务、跨平台协作。实现这一愿景需要像早期HTTP协议那样的开放标准，如模型上下文协议（MCP）和微软开发的NLWeb框架，让网站能以自然语言交互。斯科特强调，保持开放、普惠的创新环境至关重要，AI应是增强人类创造力的工具，其最终价值在于服务人与人之间的连接。

五、设计哲学：三十年硬件创新背后的以人为本
微软合作伙伴设计总监卡尔·莱德贝特回顾了其主导的多个标志性产品设计，核心始终是深入理解用户。从为解决Excel用户导航难题而发明滚轮的IntelliMouse，到通过与残疾玩家共同研发、彻底改变游戏可及性的Xbox自适应控制器，莱德贝特践行着“与我们无关的事，不能没有我们参与”的原则。他认为，当前AI技术浪潮是又一次提升人类能力的巨大飞跃，而优秀的设计应让技术无缝融入生活，赋能每个人。

六、科技重塑体验：史蒂夫·鲍尔默的“直觉穹顶”篮球馆
前微软CEO、现洛杉矶快船队老板史蒂夫·鲍尔默，将其对细节和用户体验的极致追求，倾注于价值20亿美元的“直觉穹顶”体育馆。馆内遍布创新：面向客队的“巨墙”看台、人脸识别的无感消费、每个座位的独立温控与充电接口、乃至通过传感器监测观众噪音并以此激励互动。鲍尔默的目标是融合现场观赛的热情与家庭观看的便利舒适，打造全球最佳的篮球现场娱乐体验，并证明科技能够深度增强体育竞技与粉丝文化。

本期《信号》内容显示，微软正从多个维度思考技术的边界与责任：在传播中追求克制与智慧，在安全战中主动构建生态防线，在AI浪潮中坚持工具属性与开放精神，在产品设计中深植人文关怀，并探索科技赋能传统产业的无限可能。这超越了单纯的技术叙事，呈现出一家科技巨头对复杂世界的系统思考与行动。

中文翻译：

有时，全知全能并非如人们所吹嘘的那般美好。

北欧神话中，奥丁以一只眼睛换取智慧，他深知"看见"的代价：洞悉一切意味着背负知晓过多的重担，而有时，知晓过多本身就是一种陷阱。

在现代传播领域，尤其是在微软这样的公司，我们仿佛也在效仿奥丁——注视着每一条新闻标题、每一条推文、每一篇博客文章、数字世界中的每一声低语。既然看到了一切，便极易产生回应一切的冲动：纠正每一个误解，在每一颗火星酿成燎原之势前将其扑灭。但正如迈克·马斯尼克在"史翠珊效应"中令人难忘的描述：有时，回应行为本身反而会放大我们期望消弭的事物。

回应的权衡从来不是简单的算术题。我曾亲眼见证，与有缺陷的报道纠缠有时会适得其反，反而为其吸引更多关注，赋予其本不该有的可信度。我也曾因未对失实报道采取更强硬行动而后悔，眼睁睁看着它在客户和意见领袖间传播，被奉为真相。同样，我也见证过沉默的力量——在某些情况下，不加评论地让事件自然平息，反而能使其悄然消逝。真正的挑战在于辨别何时该行动，何时该静观。这没有算法可依，也没有确保正确决策的清单。

"史翠珊效应"对我们所有从事传播工作的人都是一则警世寓言。2005年，歌手芭芭拉·史翠珊因一张其住宅的照片起诉摄影师，而该照片只是数千张记录海岸侵蚀的图片之一。诉讼提出时，这张照片仅被下载过六次，其中两次下载者还是史翠珊的律师。诉讼及舆论发酵后呢？下载量逼近五十万次。当企业和个人拥有日益强大的数字与社交信号时，我们很容易无意中将已看到但他人尚未察觉的事物推向聚光灯下，确保所有人都来注目。

如何决策？首先从我们最关心的受众出发，审视帖子、视频或文章与这些受众的交集。若关联微弱，或许无需回应。其次考虑时机。新闻与信息瞬息万变。若我们未能在第一轮传播周期内回应，我们的回应将引发第二轮传播。最后，任何回应都必须能彻底驳斥或扭转原有论点。我们或许未能抢占先机，但必须确保掌握最终话语权。

正如奥丁所领悟的：智慧不是全知全能，而是明辨是非。是懂得何时该行动，何时该克制。是经年累月建立信任与关系，从而当我们发声时，话语自有分量。在这个每家公司都置于显微镜下、每个决策都被审视、每个失误都被放大的世界里，真正的挑战不是看见一切，而是懂得如何应对所见。

我们的目标不是追逐每一个热点，而是帮助读者在喧嚣中辨明方向。

《信号》杂志的存在，旨在提供一个反思、对话与洞察的空间，而非仅仅作出反应。我们的目标不是追逐每一个热点，而是帮助读者在喧嚣中辨明方向。我们邀请您加入对话，共同塑造重要议题，以您的智慧应对我们面临的挑战。在无限信号的世界里，真正的艺术在于选择放大哪些，又让哪些悄然流过。

弗兰克·X·肖
微软首席传播官

人工智能新视角

杰伦·拉尼尔从濒临失明的经历——以及对珠宝的狂热痴迷——中领悟到的人工智能真谛

虚拟现实背后的远见者杰伦·拉尼尔，数十年来一直在探索他所助力创造的数字世界的承诺与陷阱。作为计算机科学家、音乐家和艺术家，拉尼尔联合创立了第一家VR公司，并推广了这项突破性技术。他曾入选《时代》杂志"全球百大最具影响力人物"，也是畅销书《你不是个 gadget》和《未来属于谁？》的作者。在这篇为《信号》杂志撰写的独家文章中，拉尼尔分享了他关于如何释放人工智能力量的见解。

2025年初，我收到一份误诊，称我可能在春天前失明。随后我不得不煎熬地度过了一个月，直到纠正性的第二诊疗意见终于到来。恰巧在那一个月里，我萌生了一种新的创作痴迷，这对我变得无比珍贵。而这份痴迷，正是由人工智能促成的。我之所以告诉你们这一切，一方面是希望分享一些关于如何最大限度利用人工智能的想法，另一方面，这番坦白或许也有助于消除一种普遍的误解——认为我是人工智能的怀疑者或反对者。事实上，我是一位人工智能爱好者，只是认为我们许多人对人工智能的思考方式错了。

当我被告知只剩下几个月的光明时，我竭尽全力去看清世界。我凝视植物和动物。散步时痴迷于峡谷与日落。我久久注视人们的眼睛和面容。但我也不得不参加那种会议，除了幻灯片上的方程式，没什么可看的。

我在新墨西哥州长大，自幼喜爱本土银饰。确诊后，我开始购买绿松石戒指，以便在长时间讨论矩阵运算时，能在手指上欣赏它们。很多戒指，数量多到荒谬。

在那个漫长月份的最后，得知第一位视网膜专家误诊时，我是多么欣喜。原因让我有点尴尬。早在1980年代初，我曾与外科医生乔·罗森以及我们初创公司VPL Research（也是第一家VR公司）的工程师安·拉斯科合作，在斯坦福医学院参与了可能是最早的外科手术模拟项目。然后，大约在上世纪末，我因一个小问题接受了视网膜激光手术。原来那位视网膜外科医生熟悉我早期的工作，主动让我短暂操作激光，在我的视网膜上打几个点。（显然，我不会透露这位医生的名字。）

我忍不住想加点变化，一点艺术，在激光点围成的圆圈上添了个小波浪。所以我的视网膜上有个轻微的"纹身"，这还不是所有年轻人都在做的事，但随时可能流行。正是这个偏离常规的做法——我忘了告诉第一位视网膜专家——让他警觉起来。

当我收到好消息时，我不仅对视觉，对珠宝的欣赏也达到了前所未有的高度。某种意义上，这次误诊成了一份祝福。我觉得买过多珠宝有些荒唐，于是升级成了珠宝制作者。起初这只是简单的串珠，但后来我开始接触金属粘土，接着是真正的焊接、切割，以及大量的敲打和打磨——然后是铸造、激光切割机和数控铣床。我在斋浦尔和苏门答腊建立了联系，以获取稀有宝石进行镶嵌。我彻底沉迷了。

串珠，作为珠宝制作的入门"毒品"，简单至极——只是将现成的、带孔的漂亮物件串在金属线上，但它在哲学上却发人深省。这很像人工智能！试想：一串珠子有创造性吗？你所做的不过是组合别人制作的东西。然而，一串珠子可以富有表现力。它可以超越其来源。但价值是难以捉摸的，存在于观者眼中。你需要一个故事，一个背景，才能充分欣赏一串珠子。珠子已存在数万年，每一件珠宝艺术品在它所属的时代都是一个故事，对我们而言，则是一个不断演变的故事。

不了解我曾以为自己即将失明，就无法完全理解我的串珠作品。故事必须成为珠子的一部分，它们才能充分散发光芒。抽象的珠子根本不成其为珠子。

我之所以被冠以"人工智能怀疑者"的名声，是因为我确实经常与同领域的同事进行哲学辩论。谈论人工智能的通常方式是，说它是某种新实体的创造，而它的任何问题，比如幻觉，都是需要在该实体中修复的缺陷。这个实体最终将成为一个完全通用的价值来源，而非具有特定用途的具体事物。这些都是常见观念，常见到我们甚至没有注意到它们是选择，但我并不认同。

我更倾向于将人工智能视为迄今为止最高效的人类协作形式。人工智能中没有"人"存在，没有实体，只有所有为人工智能训练提供数据的人。我更喜欢这个框架，因为它比那种认为人类将在经济上过时的观点，更能清晰地描绘出文明未来的幸福路径，但那是宏观层面的动机。在此，我想聚焦于个人、亲密的层面。

珠宝制作很棘手。当我还是新墨西哥州的一个男孩时，我曾请一些纳瓦霍制作者教我一点技艺，他们很乐意，但警告我需要多年才能学会。但那是过去。数字世界并未遮蔽实体珠宝，反而拉近了距离。

在线视频是物理技能的新晋通用教师，很有帮助。（关注我工作的人知道，我也担心在线平台可能造成的社会和心理损害，但其积极用途同样真实。）遗憾的是，视频在帮助你按需学习细节方面效果不佳。你可以观看别人如何制作戒指的教程，这可能很有启发性，但如果你试图解决一个具体问题，可能需要观看数十个相关视频仍找不到答案。哪种造型粘土在用于金属粘土烧结合成中的镶座空间固定时不会收缩？答案就在那里，但很难找到。于是你求助于论坛。不幸的是，论坛回复可能也很慢，且容易滋生烦恼，因为聊天流经常偏离主题。

人工智能是人类数据的融合体。我们创建大型人工智能模型时所做的，就是将大量数据置于统计分析之下，检测其中的文字、声音、像素——任何可以数字化的东西——的模式。然后调用这些模式来生成回应。人工智能的核心就是如此简单。我们检测模式并加以推断。

尽管通常的思维方式建议用户将程序视为伙伴、另一个实体，但我发现这并不是最有用的心态。如果我以提问方式提示，模型自然会构建一个听起来像他人相互回答方式的回应。这增加了答案略微超出数据支持范围的可能性，因为在数据被用于训练之前，人们就是这样相互交谈的。这就是当人工智能似乎超越已知事实时发生的情况，通常被称为"幻觉"。

相反，如果我这样提示："是否有其他珠宝匠成功地在单次作业中，无需更换工具，就数控铣削了玉石及其镶嵌的银饰？"，那么模型会告诉我其他人成功的案例。这更具体。如果我直接询问模型，仿佛它知道什么，我几乎是在诱使它虚构。对于这种一般形式的问题，没有单一答案，因为每件珠宝都是独特的，每个珠宝工作室也是如此。以一般框架提出的提示，其答案可靠性较低，并且低估了珠宝匠之间奇妙多样的实践。通过询问他人成功的经验，我将提示建立在人工智能模型的创建方式之上，并隐含地要求一组现实世界的答案，而非一个不存在的、虚构的折中单一答案。

以务实的心态使用人工智能，既能带来个人益处，也能提高专注的生产力。当你使用人工智能时，你的个人自主性——你的自我导向程度——应该感觉在增强，并且实际上也确实在增强。当我紧密地引导我的探究，而不是费力地浏览聊天群组或视频时，我更能掌控局面。当我询问人工智能模型，在一个庞大的3D设计程序中如何找到一个晦涩的用户界面选项时，我可以立即开始工作，而当我不得不大海捞针般查阅文档时，可能需要数小时甚至数天。我的方向越明确，获得的价值就越大，也越不容易感到人工智能正在取代我。

那些将人工智能模型视为伙伴的人，必须承担模拟这种关系所带来的额外开销。当我观察人们以这种方式使用人工智能时，在我看来，他们是用一种新的延迟替代了人工智能本应消除的延迟。此外，模拟关系在后端确实需要相当可观的计算量，意味着消耗能源。因此，模拟伙伴而非仅仅作为工具，对用户和计算机都是一种"税负"。但我理解人各有异。无意评判！我只能如实报告对我有效的方法。

另一个对我有效的原则是记住技术总是具体的。有些人可能希望将人工智能视为最终完全通用的、能够解决任何可解问题的存在。我拒绝被卷入关于我是否怀疑人工智能将变得多好的争论。那不是重点。重点是，如果你具体思考人工智能实际是什么——即它今天是什么，今天如何运作——那么你就能更好地使用它。这使得人工智能成为一种具体的事物，而非通用的事物，即使它是一种非常优秀的具体事物。尽管人工智能未来可能变成

英文来源：

Sometimes, being all-seeing is not everything it’s cracked up to be
Odin, the Norse god who traded an eye for wisdom, understood the price of vision: to see everything is to carry the burden of knowing too much, and sometimes, knowing too much is its own kind of trap.
In the world of modern communications, especially at a company like Microsoft, it can feel as if we’re channeling Odin, watching every headline, every tweet, every blog post, every whisper in the digital wind. And having seen everything, the temptation is strong to respond to everything, to correct every misconception, to stamp out every spark before it becomes a wildfire. But as Mike Masnick so memorably described in the Streisand effect, sometimes the very act of responding can amplify what we wish would fade away.
The calculus of response is never simple. I’ve seen firsthand how engaging with a flawed story can sometimes make it worse, drawing more attention and lending it legitimacy it never deserved. I’ve regretted not taking stronger action in the face of an inaccurate story, seeing it shared and accepted as truth by customers and influencers. I’ve also seen the power of silence, how, in some cases, letting a story pass without comment allows it to die a natural death. The challenge is knowing which is which. There’s no algorithm for this, no checklist that guarantees the right call.
The Streisand effect is a cautionary tale for all of us in communications. For context, in 2005 the singer sued a photographer about a photo of her house, collected as one of thousands documenting coastal erosion. At the time of the suit, the photo had been downloaded six times, two of those being by Streisand’s lawyers. After the suit and publicity? Nearly half a million downloads. And as companies and individuals increasingly have strong digital and social signals, it can be easy to accidentally take something we’ve seen but others have not and ensure everyone looks at it.
How to decide? Start with the audience we care most about and look at the intersection between the post or video or article and that audience. If it is a weak connection, then a response might not be needed. Next, consider timing. News and information move fast. If we can’t respond within that first cycle, our response will cause a second cycle. And finally, any response has to conclusively rebut or redirect the thesis. We might not have gotten the first word, but we should for sure have the last.
Wisdom, as Odin learned, is not omniscience, it’s discernment. It’s knowing when to act and when to hold back. It’s building trust and relationships over time, so that when we do speak, our words carry weight. In a world where every company is under a microscope, where every decision is scrutinized and every misstep magnified, the real challenge is not to see everything, but to know what to do with what we see.
Our goal is not to chase every story, but to help our readers make sense of the noise
Signal Magazine exists as a space for reflection, dialogue, and context, not just reaction. Our goal is not to chase every story, but to help our readers make sense of the noise. We invite you to join in the conversation, to help shape what matters, to bring your own wisdom to bear on the challenges we face. In a world of infinite signals, the real art is choosing which ones to amplify and which ones to let pass.
Frank X. Shaw
Chief Communications Officer, Microsoft
A new perspective on AI
What Jaron Lanier learned about AI from a brush with blindness – and a turbo-charged obsession with jewelery
Jaron Lanier, the visionary behind virtual reality, has spent decades exploring the promises and pitfalls of the digital world he helped create. A computer scientist, musician and artist, Lanier co-founded the first VR company and popularized the ground-breaking technology. Named one of Time magazine’s 100 Most Influential People, he’s also the author of bestsellers You Are Not a Gadget and Who Owns the Future? In this exclusive piece for Signal, Lanier offers his tips for how to unlock the power of artificial intelligence.
In early 2025 I received a false diagnosis that I was likely to go blind by spring. I then had to endure a precipitous month until the corrective second opinion finally arrived. As it happens, during that month I developed a new creative obsession that has become precious to me. And that obsession was enabled by AI. One reason I am telling you about all this is that I hope to share some thoughts on how to get the most out of AI, but this confession might also help to undo a common misunderstanding that I am an AI skeptic or opponent. Instead, I am an AI enthusiast who thinks that many of us are thinking about AI wrong.
When I was told I had only a few months left to see, I did all I could to see well. I stared at plants and animals. Went on walks to ogle ravines and sunsets. I lingered on eyes and faces. But I was also obliged to go to the kind of meeting where there isn’t a lot to look at other than equations on a slide.
I grew up in New Mexico and adored native silverwork as a boy, so after my diagnosis I started to buy turquoise rings to adore on my fingers during long discussions about matrix operations. Lots of rings, an absurd number.
What a joy it was to learn at the end of that long month that the first retina guy had gotten it wrong. The reason was a little embarrassing for me. In the early 1980s I had collaborated on what was probably the first surgical simulation, at Stanford Med, with Joe Rosen, a surgeon, and Ann Lasko, an engineer from our startup VPL Research, which was also the first VR company. Then, around the end of the century, I had retinal laser surgery for a minor issue. It turned out that the retina surgeon was familiar with my earlier work and offered to let me briefly operate the laser to put a few of the dots on my retina. (I will not reveal this doctor’s name, obviously.)
I couldn’t help myself, I had to add a little variation, a little art, a wiggle along a circle of laser dots. So there’s a slight tattoo on my retina, which is something all the kids are not doing yet, but any minute. It was this deviation from the norm – which I had neglected to mention to the first retina specialist – that alarmed him.
By the time I got the good news, my appreciation of not only vision, but of jewelry, was supercharged. In a way, the false diagnosis had been a blessing. I was feeling ridiculous buying excessive amounts of jewelry, so I graduated to jewelry maker. Initially this meant simple beading, but then I went on to working with metal clays, then to proper soldering, cutting, and so much hammering and sanding – and then to casting, a laser cutter and a CNC (computerized machining) mill. I developed contacts in Jaipur and Sumatra to source rare gems to set. I got it bad.
Beading, the gateway drug of jewelry making, is as easy as could be, just arranging preexisting, pretty items with holes in them on a wire, but it is philosophically provocative. It is similar to AI! Consider: Is a string of beads creative? All you are doing is combining things other people have made. And yet a string of beads can be expressive. It can be more than the sources. But the value is elusive, in the eye of the beholder. You need a story, a context, to fully appreciate a string of beads. There have been beads for tens of thousands of years, and each work of jewelry art was a story in its time, and an evolving story to us now.
You can’t fully see my bead work without knowing that I thought I was going blind. The story has to be part of the beads for them to radiate light fully. Beads in the abstract are not beads at all.
The way I got my reputation for being an AI skeptic is that I do argue philosophy with my colleagues in our field quite a lot. The usual way of talking about AI is to say it’s the creation of some sort of new entity, and that anything wrong with it, like hallucinations, is a flaw to be fixed in that entity. The entity will eventually become a fully general source of value rather than a specific thing with specific uses. These are common ideas, so common we do not even notice that they are choices, but I do not embrace them.
I prefer to think of AI as the most productive form of human collaboration yet. There isn’t anyone home in an AI, no entity there, just all the people who made data that the AI was trained on. I like this framing better because it makes happy future paths for civilization clearer to imagine than if one thinks of people becoming economically obsolete, but that’s a big picture motivation. Here I want to focus on the personal, intimate level.
Jewelry making is tricky. When I was a boy in New Mexico I asked some Navajo makers if they would show me a little of how to do it, and they were happy to, but warned me it would take years to learn. But that was then. The digital world has not eclipsed physical jewelry but brought it nearer.
Online video is the new universal teacher of physical skills, and was helpful. (Those who follow my work know that I also worry about societal and psychological damage that can be wrought by online platforms, but the positive uses are also real.) Unfortunately, videos are poor at helping you learn details on demand. You can watch someone’s tutorial on how they made a ring, which can be revelatory, but if you are trying to figure out how to solve a specific problem, you might watch dozens of related videos and still not find the answer. Which modeling clays won’t contract if you use them to hold a space in a bezel within a sintering metal clay? The answer is out there, but hard to find. So you ask on forums. Unfortunately, these can also be slow to provide answers, and are prone to annoyances, as chat streams frequently go off path.
AI is an amalgam of people’s data. What we do when we make a large AI model is bring a lot of data under statistical analysis, in which patterns of words, sounds, pixels – really anything that can be digitized – are detected. These patterns are then called on to create responses. AI is that simple in its core. We detect patterns and extrapolate them.
While the usual way of thinking about AI suggests that a user should treat the program as a partner, as another entity, I don’t find that to be the most useful mindset. If I prompt with a question, the model naturally constructs a response that sounds like the way other people have answered each other. This increases the likelihood that the answer will reach a little beyond what the data justifies, because that is how all those people talked to each other in the time before their data was used for training. This is what is happening when AI seems to get ahead of known facts, usually called a hallucination.
If I instead prompt with something like, “Have other jewelers had success CNC milling jade and the silver it is set in in a single job without tool changes?”, then the model tells me stories of what has worked for other people. It is more concrete. If I ask the model directly, as if it knows anything, I am practically begging it to confabulate. There is no single answer to the general form of the question, since each piece of jewelry is unique, as is each jewelry workshop. The answer to a prompt with general framing becomes less reliable, and underplays the wondrous variety found among jewelers. By asking what has worked for others, I have grounded the prompt in how the AI model was created, and asked implicitly for a cluster of real-world answers instead of a constructed middle, single answer that does not exist.
Using AI with a grounded mindset brings personal benefits as well as more focused productivity. As you use AI, your personal autonomy – your degree of self-directedness – should both feel like it is increasing, and it should actually be increasing. When I direct my investigation tightly instead of wading through chat groups or videos, then I am in more control. When I ask an AI model where to find an obscure user interface option in a sprawling 3D design program, then I can work right away, while when I had to dragnet through documentation, it could take hours or days. The more directed I am, the more value I get, and the less I am vulnerable to feeling like the AI is replacing me.
Those who treat an AI model as a partner have to add the overhead of whatever that simulated relationship entails. When I watch people use AI in that way, it seems to me that they are substituting a new kind of delay that AI should have done away with. It is also true that simulating a relationship costs a nontrivial amount of computation, meaning energy, in the backend. So there is a tax on both the user and the computer for simulating a partner instead of just being a tool. But I understand people are different. No judgment! I can really only report what works for me.
Another principle that works for me is to remember that technology is always specific. Some might want to think of AI as eventually being fully general, able to solve any solvable problem. I refuse to be drawn into arguments about whether I am skeptical about how good AI will become. That is not the point. The point is that if you think concretely about what AI actually is – meaning what it is today, how it works today – then you can use it better. That makes AI into a specific thing instead of a general thing, even if it is a very good specific thing. And while AI might be something else in the future, the only tool you can use is the one that already exists.
One thing AI can do is use patterns in training data to extrapolate. Even if there is no treatise out there about a particular esoteric jewelry making method, there’s nothing wrong with the model extrapolating using language patterns. If you prompt the model to solve a problem as if you are confident an ideal solution is already within reach, then it might hallucinate to form an answer. On the other hand, if you prompt for a speculation about a technique that sounds like things that have already worked for specific people – and who are they, and has anyone speculated about this thing I want to try, so that the model can suggest ideas about it, or even why it might or might not work – then these types of grounded prompts eschew hallucinations even while moving a little beyond what is already known.
I don’t think there’s such a thing as a hallucination, really, just a misaligned expectation from a user. A model that is used more concretely and practically can be wrong, but can’t hallucinate, because it wasn’t given a theater for hallucinations. Try this approach, see if it works for you.
Finally, notice that it doesn’t make sense to worry about whether you really did anything, or if the AI did the work for you. Whatever you are doing, it is only meaningful and valuable in the context you created. Combining beads might not be hard, and would be a relatively easy thing for an algorithm to do well enough to pass a bead Turing Test, but that’s a deceptive way to understand the situation. Beads tell a story, and the story in combination with the beads is where the value is. The point of my beading was that it was a response to my fear of imminent blindness. The AI did not live that story, I did. Enduring value is always grounded in reality.
Using AI helped me learn jewelry making with an almost preposterous speed. It was a feeling of freedom and autonomy. I have made at least one piece a day since I started and many of them are pretty good. I like wearing the stuff. I like seeing other people wearing pieces I made. It’s not about the AI, and that is what makes the AI good.
You have to innovate because the cybercriminal is always innovating
How Microsoft’s Digital Crimes Unit, a cybersecurity company and a global health organization came together to take on a new breed of hacker
One of the first signs of sabotage emerged before dawn on May 14, 2021, when doctors and nurses at several hospitals in Ireland found that they couldn’t access patient records. A ransomware attack had been silently unleashed across the country’s public health network in the middle of the night, after someone had prowled inside the IT system for weeks. By the time many employees arrived at work that morning, tens of thousands of devices had been encrypted, prompting staff to cut off internet access altogether to contain the devastation.
The cyberattack on the Ireland Health Service Executive (HSE) threw its large network of hospitals, clinics and services into chaos, endangering patients who faced delayed treatments and canceled appointments. Care providers, already dealing with the COVID-19 pandemic, reverted to pen and paper as they coped with no electronic patient records, networked phone lines or email.
Many ransomware attacks lock people out of their devices and data until a payment is made. This time, the attackers unexpectedly shared a free decryption key a few days after the strike. But it would still take the HSE four months to fully repair its IT system and 18 months to start contacting the 90,000 people affected by a data breach stemming from the attack. The incident and recovery were documented in a public report commissioned by the HSE.
“There is no underestimating the damage that this cyberattack has caused,” Paul Reid, then chief executive officer at the HSE, testified at a parliamentary health committee a month after the attack. “There are financial costs, certainly, but there will, unfortunately, also be human costs.”
Attackers had hijacked the HSE with a favorite tool of cybercriminals, a version of the legitimate security tool Cobalt Strike that had been “cracked” – i.e., stolen, pirated or otherwise manipulated to bypass licensing controls. A cracked version of the tool would also be used to extort the Costa Rican government a year later, triggering a state of emergency.
When used legitimately, Cobalt Strike is a powerful tool for “red teams,” or security testers who simulate cyberattacks in a safe, controlled environment to identify vulnerabilities. The tool can deploy malware (malicious software) to prowl a network, steal credentials, remotely control systems and carry out other harmful activities for testing purposes.
Around the time of the HSE attack, cracked Cobalt Strike was already emerging as a major threat on the radar of Microsoft’s Digital Crimes Unit (DCU), a diverse team of investigators, lawyers and other experts known for tackling cybercrime in groundbreaking ways.
When the team formed in 2008 to confront the growing problem of malware and other online threats, most cybersecurity groups at other companies were focused on more reactive efforts like patching vulnerabilities and improving antivirus software. Microsoft had a twofold interest in fighting malware: It wanted to safeguard its brand and software code – which hackers often exploit to attack Windows devices – and to proactively protect computer users worldwide.
Over the years, the DCU has developed an aggressive strategy of legal actions and global partnerships to lead more than 30 operations against malware systems, criminal groups, crime enablers and government-affiliated hackers. The operations have included disruptions of Waledac, a prolific botnet, in 2010; Forest Blizzard, a Russian-sponsored hacking group that targeted U.S. elections, in 2016; and Lumma Stealer, a fast evolving malware often used in credential theft, in 2025. The work has severed criminal control of millions of infected devices worldwide.
The operations generate valuable threat intelligence that the DCU shares with customers, partners and teams across Microsoft to help strengthen the security of the company’s services and enhance cybersecurity across global industries. The team also analyzes the intelligence and other data to identify evidence for law enforcement investigations, which has resulted in nearly 800 arrests.
“The dynamic nature of cybercrime demands constant vigilance and innovation,” says Steven Masada, Microsoft assistant general counsel and head of the DCU. “Each sector sees different aspects of the cybercrime ecosystem, and when we share our insights, we evolve our strategies to counter emerging threats more effectively.”
For cracked Cobalt Strike, the DCU deployed a novel playbook it has pioneered. Instead of targeting individual hackers, the strategy aimed to shut down hackers’ systems for spreading malware, specifically their elaborate web of internet domains and IP addresses. To do that, the DCU would need to file a lawsuit against alleged attackers and get a court order. But despite the team’s considerable expertise, the path to taking down cracked Cobalt Strike would be far from easy.
Unlike previous operations that had targeted malware directly, the DCU wanted to pursue unauthorized copies of a popular tool owned by another company. And it wanted to focus on many malware groups at once, instead of a single group or botnet (a network of infected computers). This would help drive maximum impact.
The complexities of this ambitious case meant that it would take two years of detailed technical and legal work to build, starting with Microsoft investigator Jason Lyons, who had worked on the DCU disruptions of TrickBot, Necurs and other notorious botnets.
From his home office in Texas, Lyons had been tracking the fallout from cyberattacks around the world, including those in Ireland and Costa Rica, as well as an attack on an essential U.S. fuel pipeline. A former U.S. Army counterintelligence special agent and cybersecurity incident responder, he had spent years working nights and weekends responding to crises in previous roles. Now he wanted to make a bigger impact on crime.
“Instead of me responding to the bad guys and being on call like a firefighter, I wanted to make their lives a little worse and disrupt their business, their networks,” says Lyons.
During the pandemic, he began to suspect that hackers were increasingly using cracked Cobalt Strike to attack businesses that had become distracted and vulnerable in the sudden shift to remote work. He just had to prove it.
For months, he and a coworker sifted through Microsoft data for clues, starting with alerts for all instances of Cobalt Strike use from the company’s antivirus product Defender. They studied forensic analyses from a company team that responds to customers’ cyberthreat incidents. They developed a database of known attacks involving the tool, with the picture becoming clearer. “The ransomware extortion angle was blowing up at the time, and… cracked Cobalt Strike was all over the internet,” says Lyons.
A full picture of how much hackers were using the tool would have to come from the tool’s owner, Fortra. The risk of failure was high for Microsoft to proceed alone – it needed Fortra to join the case and provide evidence and public support, prompting months of trust-building and information-sharing with an essential partner.
“We didn’t know where Fortra was going to land if we said, ‘Oh hey, we’ve got a huge problem, and you’re part of it,’” Lyons says. “Were they going to help? Were they just going to tell us to suck eggs? We just didn’t know.”
Early in the operation, Lyons and his team tried to buy a copy of Cobalt Strike to open it up and understand how it works. Fortra, a 3,000-employee company headquartered in Eden Prairie, Minnesota, said no. “We don’t just sell it to anybody,” says Bob Erdman, associate vice president for Research & Development at Fortra.
“There is a lot of background vetting before somebody can legitimately obtain a copy. We need to know their use case. Are they a real company? Are they going to use it in a manner that we’re OK with and meets the license criteria we have?”
Fortra already knew about the problem – it was seeing around a thousand instances of cracked Cobalt Strike activity every day. It had added more security controls to the software and was already removing unauthorized copies from hacker forums and file-sharing sites.
But Microsoft’s approach was much broader, prompting Fortra to join the case as a co-plaintiff in early 2023. The company shared a list of watermarks linked to unauthorized Cobalt Strike use that turned out to be a crucial piece of evidence. The watermarks are a unique value assigned to every licensed copy of the tool, giving the DCU team and its partners a thorough, precise way to identify unauthorized or compromised copies that needed to be disabled.
“Working with Microsoft allowed us to do what we were doing on a much larger footprint,” Erdman says. “They brought a lot of new data to the table, and we could bring the ability to tear apart the tool and know if it’s a real customer’s copy, or an unauthorized copy that shouldn’t be running.”
When it came to laying out the legal arguments for the case, Richard Boscovich, assistant general counsel for the DCU, knew he would have to present more than a simple intellectual property (IP) case. He had led almost every malware disruption for Microsoft and shaped the company’s legal approach with a knack for using civil laws creatively.
As in previous cases, he accused cracked Cobalt Strike defendants of breaking a copyright law more usually associated with protecting the work of musicians and artists, not the software code of tech companies. He said that defendants had violated a trademark law that’s often used to fight counterfeits like fake designer bags and stolen logos.
The defendants were never expected to show up in court – the lawsuit was just a mechanism to secure a court order for taking down their malware operation.
For the first time in his malware cases, Boscovich leveraged a civil racketeering law, arguing that developers, sellers, hackers, extortionists and money launderers colluded in a lucrative ransomware-as-a-service enterprise. “We look at all the tools that are available, including tools that weren’t meant to address cybercrime,” he says. “You have to innovate because the cybercriminal is always innovating.”
A former federal prosecutor for 17 years, he understood that it wasn’t enough to argue that hackers are simply misusing Fortra’s software and Microsoft’s code to run malware on Windows devices. For a court to allow the companies to take down other people’s digital assets, Boscovich had to show the public devastation of malware. “Judges don’t really care too much about Microsoft as a multinational corporation that’s suffering. They’re like, ‘Why are you in my courtroom?’” he says. “So the case became less about protecting Microsoft’s IP or Fortra’s IP, and more about protecting the ecosystem and our customers.”
Enter Health-ISAC, a global health security organization representing more than 1,000 member institutions. The Florida-headquartered group joined the case as a co-plaintiff to show the vulnerability of healthcare organizations and the human toll of ransomware.
The pandemic and years of underfunded IT security had left many healthcare organizations susceptible to ransomware. Meanwhile, the need to continue patient care and maintain critical systems like electronic medical records and diagnostic equipment forced some hospitals to pay attackers off, making them profitable targets. In the same year as the HSE attack, U.S. healthcare organizations were hit by a staggering wave of more than 400 ransomware assaults, according to the U.S. Office of the Director of National Intelligence, which oversees the country’s intelligence agencies.
“The modern-day hospital is so reliant on IT that when these systems go down, it’s incredibly devastating,” says Errol Weiss, chief security officer for Health-ISAC. “They can’t do patient intake, and ambulances are diverted. Services slow down because they’re relying on paper and manual processes. If you’re with a patient trying to do surgery and need to know their blood type, you’ve got to go to paper backup and hope it’s available and reliable.”
Ransomware often has severe downstream consequences, and Weiss ticks off a few that made headlines. A rural hospital in Illinois closed after spiraling financially from an attack and the pandemic. Hackers stole patient records from a health network in Pennsylvania and published them, including naked photos of cancer patients receiving treatment. The attack led to a class-action lawsuit against the network and a $65 million settlement. In Finland, a patient died by suicide after a hacker stole confidential records from a psychotherapy center, failed to get a ransom, exposed the records and blackmailed patients.
Health-ISAC, Fortra and Microsoft were able to merge their considerable data and expertise to link cracked Cobalt Strike to 68 health-related ransomware attacks in 19 countries. Their investigation connected cracked copies to eight malware families, including LockBit, a fast encryption and denial-of-service attacker, and Conti, the malware used in the HSE and Costa Rican attacks.
“I’m a big advocate for the work that’s being done,” Weiss says. “There’s an ecosystem that criminals can use to their heart’s content, and unless we do something about that, this problem will not go away.”
Anatomy of a takedown
How Microsoft’s Digital Crimes Unit team break up networks that use “cracked” copies of legitimate software to spread malware

1. With the help of the software’s creators, identify online distributors of unauthorized or compromised copies of software that are being used to mount malware attacks.
2. Bring a civil lawsuit against these distributors based on alleged violations of trademark law. The defendants won’t show up in court, but that doesn’t matter.
3. In court, link the use of cracked software to malware attacks, show the public devastation it causes, and demonstrate the necessity of protecting the digital ecosystem.
4. Gain a court order allowing for the seizure of domains hosting cracked software copies and directing hosting providers to remove them.
5. Systematize the takedowns by crawling the web for instances of cracked software and automatically sending out removal notices to hosting providers.
   6.Result: a huge drop in the number of servers hosting unauthorized copies of the cracked software and a reduction in how long unauthorized servers stay active.
   Nearly two years after the HSE attack, a U.S. federal judge issued a court order in 2023 allowing Microsoft to seize domains and direct hosting providers to remove instances of cracked Cobalt Strike. The immediate impact was swift, with all malicious .com and .net domains seized within 24 hours of the order.
   The disruption has since evolved into a collaborative, automated takedown process, with the DCU crawling the internet for instances of cracked Cobalt Strike, Fortra providing a list of unauthorized watermarks and the DCU sending notices to hosting providers and government cybersecurity authorities to remove illegal IP addresses.
   The work has contributed to a 72% drop in the number of servers hosting unauthorized Cobalt Strike and a sharp decline in the lifespan of those servers, which are used to control infected computers. Before the operation, unauthorized servers stayed active for an average of 49 days. By the summer of 2025, the lifespan was a mere 16 days.
   “This is the impact of persistent notifications and the automated framework,” says Zoe Krumm, director of data analytics for the DCU. “It’s not just that unauthorized C2s (command-and-control servers) go down. When they go up, they’re not up as long. That gives me chills.”
   The operation has had a particularly significant impact in the U.S., thanks to the Digital Millennium Copyright Act (DMCA), a federal law that imposes steep fines on hosting providers who fail to quickly remove IP addresses hosting infringing content. “The DMCA is a very big hammer,” Boscovich says. “The order goes out. The sites go down.”
   In response, hackers have moved many cracked Cobalt Strike servers out of the U.S. and into countries with less regulation like China and Russia. Some security experts compare the maneuvering to a game of Whac-A-Mole, with the DCU chasing hackers globally with rapid takedowns customized for different countries, local laws and international IP treaties.
   The DCU is also continuing to seize domains and “sinkhole” them, redirecting malicious traffic to Microsoft servers for threat intelligence analysis. It has shared evidence from the case with law enforcement agencies to support criminal investigations. Fortra has worked with European law enforcement agencies to remove nearly 600 malicious IP addresses. And both companies have shared their expertise in the case at security conferences to help others battle ransomware.
   “This case is a powerful example of our team’s mission in action,” says DCU head Masada, a former federal prosecutor who led cases against major cybercrime groups in that role. “It highlights our commitment to strong partnerships and continual innovation to disrupt cybercriminal operations and protect not just our customers but the broader digital ecosystem.”
   For DCU investigator Lyons, the operation was another opportunity to make the digital world a little safer for large numbers of customers through teamwork with his colleagues, an eclectic group of lawyers, analysts, former law enforcement and government workers, and other experts dedicated to fighting cybercrime. “I’ve been able to do a lot of cool things in my life, protecting national security with the military and counterintelligence and things like that,” Lyons says. “But if I had to look back on my career, the greatest impact I’ve ever had is this job. We are helping millions of people.”
   “I haven’t felt this sort of excitement in a while”
   Microsoft’s chief technology officer and amateur potter Kevin Scott on how the web will be transformed by AI in coming years
   Back in 1993, Kevin Scott saw a demo of the Mosaic browser, the first widely used graphical interface for the nascent World Wide Web. As a technologist more interested in back-end workings than user experiences, he wasn’t impressed.
   “I was like, this is the stupidest thing I’ve ever seen,” recalls Scott, then an intern at the National Center for Supercomputing Applications at the University of Illinois Urbana-Champaign, where Mosaic was developed. “I didn’t understand it at all. Like, why would anyone care about that?”
   But a few years later, Scott’s thinking shifted.
   He built his own HTTP server from scratch, stood it up on a public IP address and realized anyone in the world could access it. Scott saw how easy it was to create and share on the internet – an open platform that offered people the power of permissionless innovation, a place where anyone with imagination could go experiment and try out their ideas. Using a simple set of protocols, people could build what they wanted, how they wanted, no approval needed.
   Now Microsoft’s Chief Technology Officer, Scott sees a similar spirit of openness and innovation around the agentic web, an emerging vision of an internet powered by artificial intelligence (AI). “I haven’t felt this sort of excitement and this amount of creative energy about building brand new things in a while,” he says.
   As Scott explains, the agentic web is an open ecosystem of AI agents that can act on behalf of users. These agents won’t just answer questions. They’ll perform complex tasks, make purchases and interact with services. They can navigate websites and APIs. They will understand users’ goals and preferences, learning from interactions to improve over time.
   “You want to be able to tell an agent to go do arbitrarily complicated things,” he says. “And it should be able to get access to all the resources it needs to do those things relatively autonomously, inside of the parameters you’ve defined for how much you want to be involved in the process.”
   From scroll to solve
   The agentic web represents a radical shift in how we use the internet and what we have come to expect from it. In the 1990s, websites were mostly read-only, static pages of content that users couldn’t interact with. There was no AI involved, and any “intelligence” came from basic algorithms and humans creating and linking content.
   Over the following decades, the web became a more dynamic and engaging experience. Social media platforms allowed people to connect online, and websites evolved from static information hubs to virtual communities. Users became participants and content creators, not just consumers of information.
   As the web evolved through the 2000s, artificial intelligence was advancing in ways that would soon converge with internet services. AI researchers leveraged the massive datasets the web produced to train powerful models. That laid the groundwork for large language models, which began to emerge in the 2010s and transformed how we interact with digital content, enabling machines to understand and respond to human language with unprecedented fluency.
   Large language models, Scott says, shifted web search from typing keywords into a box – “that was revolutionary technology 20 years ago that kind of looks barbaric now by comparison” – to a more interactive, natural way of getting information.
   “You don’t have to think about things in terms of keywords,” he says. “You just say exactly what you want, and to the extent that the system has to guess at all about what it is you’re looking for, it can even ask you to clarify.”
   Microsoft’s launch of Copilot in 2023 further redefined how people use and interact with the internet. Not simply a standalone chatbot, Copilot was designed to enhance productivity and creativity in work and daily life. Integrated across Microsoft applications, the conversational assistant quickly became a valuable tool capable of helping with everything from summarizing meetings and managing inboxes to helping plan vacations and suggesting what to make for dinner.
   In late 2024, Microsoft introduced Copilot Agents, task-specific assistants that can act autonomously, orchestrate workflows and respond to triggers from external systems. While Copilot began as a productivity assistant, it has become a foundational layer for Microsoft’s vision of the agentic web, where AI agents collaborate across systems and websites to handle complex tasks for people.
   Achieving that vision, Scott says, requires a new set of protocols, standards and conventions that allow agents to interact with the web in meaningful ways. And crucially, he says, the agentic web must remain as open as possible to encourage broad participation and not stifle innovation.
   “The thing that worries me most about AI, more than anything else, is that we lose that environment of participation too soon because of commercial pressures,” Scott says. “In the early stages of something like AI, you have no idea whether you’ve discovered the best possible idea yet. So you don’t want anything to get in the way of that discovery of the best possible.” Everyone working in AI right now, Scott says, should strive for more openness, not less.
   “We should want things to evolve more in the direction of how the internet evolved, where it really is simple and permissionless and encourages lots of people being able to do the most creative thing that they can imagine doing – rather than things being more vertically integrated and closed off to people being able to freely participate.”
   Microsoft’s role, Scott believes, is to provide platforms that empower others. Microsoft’s partnership with OpenAI is key to advancing the agentic web, he says, but there is a need for broader collaboration – with AI infrastructure companies, developers and regulators.
   “As a platform company, we’re only as good as our partners are,” he says. “We have to create the conditions for lots of people to have a lot of success.”
   ‘A super simple protocol’
   As an example of that openness, Scott points to the Model Context Protocol (MCP), a new standard introduced by AI company Anthropic that standardizes how AI systems connect to external data sources and tools. Like early internet protocols, MCP is composable – designed to be modular and interoperable – and can be combined with other components or systems to build more complex functionality. Scott likens it to HTTP, the system that lets browsers communicate with websites.
   “It’s a super, super simple protocol – it’s open source, and it’s not that much work to wire a thing you’re already doing or build something from scratch and give it an MCP interface,” he says with enthusiasm. “Anything that can speak to an MCP endpoint can then access the thing you just put out there. It has all of the things that I thought were really interesting about the early web protocols.”
   Another key innovation is NLWeb, an open-source framework developed by Microsoft to bring conversational interfaces to websites. The system lets any site become an AI app by enabling users and AI agents to interact with web content using natural language. Instead of having to rely on site menus or keyword searches, users can just ask questions – for example, “Can you tell me which recipes on this site are gluten-free?” – and the NLWeb-equipped site responds intelligently.
   NLWeb was developed and conceived by R.V. Guha, a technical fellow at Microsoft and the creator of widely used web standards including RSS, RDF and Schema.org. Built on those standards, NLWeb makes it easy to make content and services discoverable by AI agents, Scott says.
   “It’s a low-effort way to participate in the agentic web,” he says. “There are businesses that don’t exist yet that are going to use NLWeb as the way to build their little slice of the agentic web to help agents serve their users better.”
   Scott gives a practical example from his own life: sourcing specialized and sometimes obscure materials for his pottery projects, like sodium hexametaphosphate (the active ingredient in Calgon), which is used to enhance ceramic glazes. With NLWeb-enabled sites, an agent could find suppliers, compare prices and even make purchases – all without Scott needing to do anything.
   “Instead of having to make a list of things that I want to buy and ordering them, I could have had the agent do all of it,” he says.
   Building agentic memory
   One recurring theme in Scott’s agentic web vision is that of memory – specifically, how AI agents remember and use information. Without memory, agent interactions are transactional and limiting. “If you were delegating a task to an employee or colleague who had no memory, it would be very difficult for them to do anything useful,” he says. “Memory will make agents more efficient and useful.”
   Scott envisions standards for memory like those around documents – created, owned and shared by their users. The approach, he says, would allow people to control how their data is used and prevent fragmentation, with different agents having siloed memories and being unable to collaborate on tasks.
   “You don’t want to have to teach every new agent you’re using what your preferences are,” he says. “It would be way easier if those were part of a set of memory preferences you could share.”
   Recent breakthroughs are already improving those capabilities, Scott says. Copilot and other agentic systems are getting better at remembering information from previous interactions and using it in the appropriate context, similar to how human memory works.
   “If you think about biological memory, it has really good recall. You can recall across a huge number of experiences,” Scott says. “The first thing that you remember about something may not be accurate, but you have a whole bunch of tools at your disposal to refine the precision of the recollections. I think that’s going to be an important quality of the memories that agents have.”
   A tool for creativity
   Scott grew up in the small rural town of Gladys, Virginia. His was a family of makers, the sort of folks who were forever tinkering with cars or restoring furniture and couldn’t let their hands be idle, even for a moment. Working on furniture projects with his dad and grandfathers as a kid, Scott developed a deep curiosity about craftsmanship and a fascination with how things are made.
   As someone who is passionate about making things – from digital tools to books, jewelry and ceramics – Scott views the question of the role of artificial intelligence in creativity as “one of the more interesting challenges of our times.”
   In an interesting experiment, he recently used Copilot and other AI tools to reverse-engineer a 17th-century Japanese ceramic firing process called hikidashi, in which pots are pulled from a hot kiln to quickly cool and develop a distinctive glaze. There is little documented about the technique in English, and AI helped Scott find Japanese sources of information, translate them and adapt the process to modern materials.
   “If I didn’t have AI to help, the problem would probably be so daunting that I would just have to give up on it before I got it solved, because I’d have to move onto something else,” he says. “For me, it’s really about accelerating my own creative productivity.”
   But Scott is clear that AI should support creativity, not replace it. “I don’t want AI designing anything I’m making. I’m perfectly happy to use it to help me solve a technical problem with something that I’m doing, but I want to do the work,” he says. “I think the most important thing for a creative person is that they should be able to choose how they want to use AI tools, if at all.”
   That philosophy extends to Scott’s broader view of AI. Tools are only as important as the people who use them, he says. And the best tools empower people to create for each other.
   “You can have a whole universe where AI is making a bunch of shit for other AIs,” he says. “But we as human beings would be profoundly uninterested in that. We do things for each other.”
   The path forward
   Scott is optimistic about the current moment, seeing it as an inflection point that could rival or even surpass the mobile revolution. “We are on an inevitable course right now,” he says. “The technology exists. It’s good enough. The only thing stopping it is cost and diffusion.”
   His advice? Be ambitious. Try things. AI technology is getting better and cheaper all the time, so don’t wait and risk having to play catch-up later. To Scott, the agentic web offers the same exhilarating, limitless possibility he felt standing up that server decades ago.
   “There are a whole bunch of people who are working feverishly using these AI tools to make brand new things that I haven’t even imagined yet,” he says. “And it could be the most amazing thing in the world. And then I get the chance to experience new things and have my mind changed.”
   “To me, it’s just awesome when the world’s in that state.”
   What’s the agentic web all about?
   Key terms to help you understand this new AI-powered version of the internet
   Agent (A)
   An AI-powered helper that can take actions, make decisions and interact with other agents or humans on your behalf. Think of an agent as a digital assistant that’s proactive, not just reactive – able to handle tasks, answer questions and learn as it goes.
   Agentic web (B)
   An open ecosystem in which AI agents act on behalf of users – from handling complex tasks to making purchases and collaborating with other agents across different sites and services. The next evolution of the internet, the agentic web will make online experiences more personalized and efficient.
   Agentic memory (C)
   An agent’s ability to remember things over time, like your preferences, past conversations and tasks you’ve asked it to do. Instead of starting from scratch every time, agentic memory helps agents build up knowledge and get smarter about helping you.
   Copilot Agents (D)
   Specialized AI agents built into Microsoft Copilot that can help with specific tasks like researching, summarizing or organizing information. Designed to work together and with you, Copilot Agents can be customized for different roles and workflows.
   Model Context Protocol (E)
   A new technical standard introduced by AI company Anthropic that helps AI agents connect to external tools, apps and data sources in a smart and consistent way, even if they’re running on different platforms or models. MCP is like a common language that allows AI agents to “talk to” other systems to get things done.
   NLWeb (F)
   Short for “Natural Language Web”, NLWeb is an open-source framework developed by Microsoft that lets humans and AI agents interact with web content using natural language. Any NLWeb-enabled site can become an AI app – instead of clicking through menus or forms, you just ask for what you want using natural language.
   The designer’s notebook
   Carl Ledbetter has been shaping the world for 30 years. As Microsoft’s Partner Director of Design, he is the visionary behind landmarks in hardware including the IntelliMouse, the Xbox and the game-changing Adaptive Controller. He talks us through five influential creations he helped bring to life
   The IntelliMouse (1996)
   “My first day at Microsoft was 30th January 1995, when I was hired to design a new mouse. At the time, Microsoft was very much a software company, so I expected to pick up a few new skills, meet interesting people and create a product or two and be done. I certainly didn’t expect to still be here 30 years later contributing to a legacy of hardware design.
   I soon realized the most important thing when designing products for Microsoft was to understand the customer. With the mouse, the challenge was coming from the Excel team. They were saying that people were producing enormous spreadsheets that were too big to fit on a screen. The only way to navigate around this environment was through scroll bars at the top and bottom and then trying to zoom in and out. My job was to create a mouse that made that easier.
   I quickly learned a lot about spatial mapping. When someone is navigating on a screen, their mind maps forward, back, left and right in a certain way. It’s abstract and subconscious, but you cannot mess with that as an industrial designer. If a product looks good and brings beauty to what you’re doing, that’s great, but it needs to be intuitive, and it must have a functional value.
   With that in mind, I started thinking about how to put control directly in the user’s hands. I created sketches and built prototypes with all these different ways to zoom in and out, to pan, to scroll… Eventually I determined that a wheel was probably the best way of doing this: it was adaptable and flexible and fit naturally within the mouse’s shape. We refined it, shaped the mouse to fit the hand and made the wheel feel as intuitive as possible. The result was the IntelliMouse – which went on to be Microsoft’s most popular and best-selling mouse for years. I’m proud that it set the bar for ergonomics, and it is great to see the wheel still deployed in a lot of mice today. When people ask what I do, my wife always jokes, ‘Yeah, he invented the wheel.’”
   ActiMates Barney (1997)
   “Six months into my role at Microsoft, the hardware division made a bold move, acquiring a company pushing the boundaries of interactive technology. Together, we launched a new generation of toys – starting in 1997 with none other than the beloved purple dinosaur Barney. The reason I’ve included ActiMates Barney in my selection is because it’s another example of where Microsoft was ahead of its time.
   The industrial design aspect was limited – we created intuitive receivers that fit into both the ActiMates ecosystem and the home – but the experience was incredible. Kids could play with Barney on his own – you could cover his eyes and he would say, “I can’t see you” and then you’d pull your hand away and he would say, “there you are” – but the real differentiator was when you connected him to a PC. There was a game which asked you math problems and, as you were going through them, Barney could help you because of the connection between the game, the PC and the toy. If you plugged a receiver into your TV, you could watch the Barney & Friends show with the toy next to you and it would respond to whatever was happening on screen. It was like having a virtual friend there for these kids. That didn’t exist before.
   ActiMates was an ambitious and forward-thinking entry into consumer entertainment and helped Microsoft build momentum in the PC gaming space. It also proved that Microsoft technology could be more than just functional – it could be magical.
   Like pretty much everything I’ve been involved in, it is part of a quest to try to do things that impact people in new ways. Of course there’s a business behind these things, but that’s never the starting point. The beginning is always ‘How do we do something that can really change the way people engage with the world?’ And that’s not a bad way to spend your career.”
   The Xbox (2001-today)
   “The mission behind these consoles echoes everything I’ve learned over 30 years – to create technology that’s powerful, purposeful and beautifully integrated into people’s lives. How did Xbox come to be? For the first-generation version (released in 2001), we had to be super scrappy: we were leveraging off-the-shelf components to get it out. But what is interesting for me is how we refined it with every new iteration.
   One of the first things I did was to work on the controller. The first controller was way too big. It hurt people’s hands, so we used our human factors expertise for the next iteration – it was designed for comfort. We thought about control layouts and worked with female gamers to see what was needed for their hand sizes.
   This human-centric design was at the heart of everything we did with Xbox from then on. With Xbox 360 (2005), we started to push what could be done with wireless technology and online gaming. Xbox 360 S was an exercise in reduction. Instead of having all these plug-in wireless receiver antennas and the hard drive on top that looked a bit like Frankenstein’s forehead, we were able to make the console significantly smaller and still build in everything.
   We made a misstep with 2013’s Xbox One, we got a few things wrong with that, but it’s like soccer, right? You miss, but it’s all about the recovery. How fast did you bounce back? And Xbox One S and Xbox One X were definitely comebacks. These products are incredible.
   I just love the progression. We design for the everchanging landscape of devices and the way people play. Every time we make a new edition, it’s this exercise of refine, refine, refine. So while on the inside we’re adding more and more technological capabilities, on the outside we’re striving to keep it simple. And we’re not done yet. We recently launched Ally X, which is a collaboration with Asus [to create a new line of handheld gaming devices]. This world just keeps getting bigger.”
   Zune (2006)
   “While it wasn’t the commercial success I thought it deserved to be, Zune was, in many ways, the highlight of my career. There were so many ideas crammed into that music player. It was a physical device but also an entire ecosystem that had a bunch of technological advancements you can see in technology today – it has had a real ripple effect. You could share tracks Zune-to-Zune, Airdrop before Airdrop if you will; it had a PC client so you could listen across devices, Zune marketplace where you could buy tracks and set up playlists and a subscription service, offering unlimited access to millions of songs. Looking back on it, I don’t even know how we did it all in the time we had.
   From an industrial design perspective, we were really pushing what you could do with molded resins. If you look at the design of that first device, you can see what’s called a ‘double shot’ plastic casing on it. The first shot was an opaque color, sort of root beer brown, and then over the top of that, we layered a coating that almost made it look like worn beach glass. What that gives you is a depth to the product, thanks to the ways light would come through and reflect off the surfaces. We really wanted to create something that when you held it in your hand, it felt special, not just like a hunk of plastic. We wanted to feel you’re getting a glimpse into this world of music.
   Zune was one of the most collaborative projects I have worked on. Everybody was shipped into this small building down in Bear Creek, which is off Microsoft campus down in Redmond. You had marketing, designers, program managers and engineers all jammed into this building, and it felt like this small community of purpose. Everything was about celebrating the art of creating music. There’s a whole case study on Zune that would show how if you can mobilize people with a clear goal to go do something, you can change the world.
   While we may not have sold millions, it’s awesome to see the ideas we had in that space play out in different ways, in different businesses and different teams. I love that. There are certainly no sour grapes.”
   The Xbox Adaptive Controller (2018)
   “In my 30-plus years in design, the product I am most proud of is the Xbox Adaptive Controller, which was designed to meet the unique needs of gamers with limited mobility. It was one of those grassroots ideas that seems to take on a life of its own. You can trace the origins back to the Xbox Elite controller, which allowed you to personalize the device by remapping buttons and controls for how you play. When we were doing some research on what people think of it, we discovered that people with disabilities were modding it so they could play games one-handed. This started us thinking about what more we could do. After a Microsoft Hackathon, we came up with a design that we thought would be even more adaptable and inclusive, but when we started meeting with these players, we found that it provided little value because many of them couldn’t hold the controller. We were being told, ‘The idea of it is right, but the solution is wrong.’ That’s when I first heard the phrase: ‘Nothing about us without us.’ We were being told, ‘Don’t pretend you know what we need and what we want on your own: Work with us.’ This became something I applied across my professional life from then on – don’t be so bold as to design for people whose needs you don’t understand. So, we started working with hospitals and wounded veterans. They tried prototypes, gave us feedback and helped create the Adaptive Controller we know.
   It’s not a mass market product, but I don’t think I’ve worked on anything with a bigger impact. We created something that unlocked the ability for people to play games that they could not otherwise, and as a designer, that’s a proud moment. I’ll never forget talking to a wounded veteran who told me that this product changed his life. Before he felt like an outcast, like he no longer fit in, that all the things that he used to like to do, he couldn’t do any more because of his disabilities. But through gaming he found a new sense of purpose and a place where the playing field was even. That was very powerful. Since 2014, Microsoft’s core mission statement has been ‘to empower every person and every organization on the planet to achieve more.’ I can’t think of a product I’ve worked on that better embodies this than the Adaptive Controller.
   It’s part of the reason I’m still excited to be at Microsoft 30 years on. There’s always a new challenge. Right now, we are seeing a pivot for the entire industry with AI, and Microsoft is at the heart of that. I, like a lot of people, use AI every day, and it has profoundly changed the way I work. I can get a lot more done. We’re in a constant state of change with technology, and AI is the latest great leap forward. Being in the middle of that, seeing how we work and interact with the world changing, is a pretty cool place to be.”
   “I just thought that there’s a better way to do things”
   What do you do when you’ve run one of the biggest companies in the world? If you’re former Microsoft CEO Steve Ballmer you set about changing the way people watch live sports…
   Surely Kevin Durant couldn’t miss. It was October 2024 and with moments left of the LA Clippers’ first game in its new home, the Intuit Dome, the visiting Phoenix Suns’ star player looked like he was about to change the course of a tight contest.
   Durant, fresh from winning Olympic gold with the US basketball team, had secured two free throws following a foul. He has some of the best stats in the NBA when it comes to shooting – netting 88.4 percent from the line – and seemed likely to hit both. But he had never come up against The Wall before. The jewel in the crown of Steve Ballmer’s $2bn sports arena, The Wall is an imposing grandstand of 51 rows of seats reserved for the 4,500 loudest Clippers superfans that has been designed to change the game – literally. Away teams face the Wall for the second half of games, the theory being that it will provide a sizable home court advantage by intimidating and distracting those who come up against it – and it certainly had the desired effect on Durant. He missed both shots. “It was crazy,” Durant said after the game, which the Suns won in overtime. “I was just staring at it the whole time. You’re not used to that.”
   The Wall is only one of the things at the Intuit Dome that people are not used to. When we take a tour of the Dome in the off season, it feels like we are coming across something unprecedented around every corner. The club store and food stalls are frictionless, using facial recognition via the Dome’s app to allow you to simply grab what you want and walk straight out, with payment taken automatically from your account. On game days the same system is used to allow people quickly to gain entry to the arena without waiting in line. Not only does every one of the 18,000-plus seats have a built-in USB port for charging your cellphone, but also its own climate control, thanks to a ventilation system that draws air directly from the roof. Noise levels aren’t just monitored as a whole, but for every seat. This is how the Clippers’ management builds up The Wall – if you’re in The Wall section and not making enough noise during a game, you’ll be politely encouraged to up the volume levels or directed to a more sedate spot next time around.
   Above the court hangs a 4K Halo board. Covering the better part of an acre, it is the largest doublesided display in any arena in the world, and when staff flip the switch during our tour to turn on its 233 million LEDs we are left blinking as if the sun has just re-emerged after an eclipse.
   The arena, which is planned to be the first in the NBA to be carbon-neutral, bears the hallmarks of a man known for his obsession with detail. Over his 30 years at Microsoft, Steve Ballmer helped steer the company through its most explosive period of growth. As CEO from 2000 to 2014, he tripled revenue and doubled profits, overseeing the launch of products like Xbox, Office 365 and the Surface tablet. Having left the company, he purchased the Clippers in 2014 and has delivered the same precision and focus he applied to software to the world of sports. Over a wide-ranging conversation he tells us how he brought the noise…
   What was it like, watching an event in the IntuitDome for the first time?
   Steve Ballmer: The first event was a Bruno Mars concert and I was mostly just harried as hell, a bundle of nerves going around the building. Bruno played the next night too and I was calmer: we were getting good feedback, people really liked the building. But the first basketball game was very different because I wasn’t worrying about the logistics of the building. I was worried about two things: Firstly, how the basketball-specific stuff, particularly the scoreboard, would perform and, secondly, how the crowd would sound and feel. I sat with my wife, my son and my oldest friends and it was like, ‘Yeah, this feels good. This feels right’. And then I got sad because we lost the game. We actually lost our first four in a row in the building. I’m glad we won our fifth. I’d told the team I’d have to start building a new arena from scratch if they lost that, because this one clearly isn’t working. Our players got a little chuckle out of that.
   I get the impression that you spent years as a fan going to games and feeling frustrated by the experience. Is that right?
   SB: I just observed things that I thought could be better. The only team I ever had season tickets for, before the Clippers, was the Seattle Supersonics. And yeah, I got frustrated waiting in line for bathrooms and drinks, for sure. I just thought that there’s a better way to do things. And so I mostly came at it thinking ‘How do we do the best job anybody’s ever done?’. I wanted something special for our fans. You can watch our games elsewhere [on TV or streaming services] so a lot of it is about the live entertainment experience. And how do our fans help our team win? What does it mean to have fans engaging in a way that can actually help with winning?
   It’s ironic, because wasn’t one of the reasons you bought the Clippers in the first place was because they didn’t need a new arena building – the team was happily sharing an arena with the Lakers and the NHL’s Los Angeles Kings?
   SB: Yes. Originally, when I bought the team, I said, ‘It’s great, I don’t need to buy a new arena!’ I looked at buying Milwaukee, Sacramento and Seattle. In all three instances, they needed new arenas and I said ‘I never want to do this. It’s not my area of expertise.’ Then a friend of mine said: ‘Hey, this is not our home. This is the Lakers’ home. How do we have a place that our fans can feel is special? How do we break out of the shadows?’ This was in 2015 and we knew we needed to be out in 2024 [when the Clippers’ 25 year lease at the Crypto.com Arena expired]. So that put us on a firm timeline.
   How did you approach the development of Intuit? You’ve obviously developed a lot of products at Microsoft. Was it the same process?
   SB: It’s different. It’s not as if I was closely involved in the development of all the products at Microsoft, but with the Intuit Dome I had a very clear vision of what I wanted the thing to be. I’d never been involved in construction projects, even house remodeling, my wife had done all that, I had no experience with architecture, but I knew what I wanted. I wanted a place that would pulse with energy. I wanted a place where the fans would become part of the experience and part of helping us win. I knew we had to have concerts. I knew I didn’t want hockey. Hockey is a great sport, but a hockey building’s never going to be as intimate because the rink just spreads things out. I wanted people to be in their seats. That’s partly how you bring energy. You have to be there. You can’t say, ‘Oh, I’m caught in a line for the bathroom or to go get food.’ So those were design principles. That’s why we have facial recognition in the stores, admission gates and food stands. We put that in place to get people into their seats faster. [Intuit also has three times the number of toilets of the average NBA arena].
   How do you set about keeping people in their seats?
   SB: Leg room is a big issue. I want to give people leg room because if you don’t give people leg room, you’re losing their attention on the game, you’re driving them to stand up, walk around, whatever, but the more leg room people have, the further back people get pushed. So we had to sort of balance that. Same thing with headroom. I didn’t want people to struggle to see over the people in front of them.
   Did you get involved in the discussions surrounding tech in the stadium?
   SB: I told the guys all along the way, ‘I don’t want technology in here for technology’s sake.’ This is about being the greatest live entertainment experience, particularly for basketball, in the world. We decided to connect power to every seat so people can charge their phones, but that opened up other interesting options – we could also put lights in the seat, we built a mini video game controller into the seat so you can interact with the big board, because people like to do that. We can do a lot of different things and that all came for free once we had power in the seats. I had another design principle: we’re going to treat people who sit up top and pay less for their seats as well as we treat people down below. They have to have a good seat, they have to have legroom, they have to have good access.
   One of the major features of the stadium is The Wall. To what extent have you seen that influencing games?
   SB: If you look objectively at the statistics when it comes to free throw shooting percentages by the opposing team when facing The Wall, it’s the lowest in the NBA by an interesting margin. So that worked.
   And technology plays a role in The Wall too, I believe…
   SB: It does! How do you get people to make noise and be in their seats? We have proximity sensors in the seats. We can tell whether you’re in, out, how much time you’re in your seat during the game. We have something we call sound cameras. So we don’t hear what you’re saying, but we can tell how noisy you are, and can tailor a rewards system in order to try to use the infrastructure to get you to be noisy. We have a thing that we copied from airports where two people standing next to each other are looking at the same screen, but see different things. We have one of those at The Wall entrance and at the moment it says say ‘Hey, Joe, welcome’. Now that we have a reward system, we could say ‘You didn’t make much noise that last game! Get it up this game and you get a free hot dog’. We’re still playing with what the rewards infrastructure looks like, but everything’s about getting people into the experience.
   What about players – how have they reacted to the new building?
   SB: I knew I wanted to have the best player spaces in the NBA as a tool to help boost results, but also recruitment. We want to tell the players on our team – and for the word to get out around the league – that we invest in our players, this is a good place to go play. Obviously if players are getting paid a lot more money someplace else, they’ll go, but for those on the margin, the fact we care more is the message. For example, the guys want to have long coats even though it’s LA. That style drove part of the design of a locker we made. You need room for lots of shoes, because a lot of these guys have lots of them, so every one of our lockers holds 32 pairs.
   Have you seen other teams taking note of what you’re doing and then starting to copy it?
   SB: We certainly have had a lot of NBA owners come through and look at our place, particularly if they’re going to build a new arena. It’s my belief that you’ll see other stadiums built over the next five to 10 years that have “Walls” or that kind of intimacy and steepness in the bleachers.
   You’ve said you want to create an experience of watching a game live that has the best elements of watching at home. What do you mean by that?
   SB: You’ve got to have a great view. You have to have leg room, and the ability to run to the bathroom or grab a drink quickly. You want to be able to see some of the statistics that get overlaid on the broadcast. We don’t want people looking at their phones to get these because if people look at their phones, they’re distracted from the game. So that’s why we built the controller into the seats. That’s why we have a huge scoreboard. The goal is to not drive you away from the live game experience. What is it that people crave when they come? It’s the energy, it’s the excitement. And perhaps increasingly so, given that we spend so much of our time just buried in screens.
   You’ve done some incredible things in your life. How does buying the Clippers and then creating this stadium compare?
   SB: Will anything I ever achieve or do match the kind of importance and complexity of growing Microsoft from 30 people and $2.5 million in revenue to 88,000 people and $88 billion revenue? No, of course not. With that said, this [buying a basketball team and building a stadium] is a great opportunity to make a civic contribution. But it’s also, for me, just the joy of watching basketball. I love it and I love being involved. I don’t try to drive too many decisions about who’s on the team and who should play, that’s on the coaches, but I ask a lot of questions. In a sense, it’s a little bit like me managing engineers. I learned to ask a lot of good questions. For pure fun this blows Microsoft away for me.
   How do the worlds of sport and business compare?
   SB: Business people think they’re highly accountable, but compared to sports people? It’s not even close. Every 24 seconds, you either score or you don’t score. You’re getting your performance reviewed in real time. If the coach doesn’t think you’re doing a good job, he pulls you out of the game. That’s a performance review. So everything is more intense, more accountable.
   Are there any similarities?
   SB: One thing that’s like the software business, at least the software business of old, is we do major version upgrades once a year. Every summer we pivot the roster, change anything we’re going to do differently in the arena. That’s the major upgrade. We do minor upgrades at the trade deadline. That’s the next place where people redesign the product, if you will. And I guess in the world of agile development, the coach is continuously making upgrades and changes in the way we play. So yeah, it’s got that notion of rapid change that was so wonderful in the software business.
   You were famous at Microsoft for being very focused on the numbers. What’s the metric in basketball that you track that might surprise people?
   SB: There’s a lot of data on the basketball side, obviously, but what we’re doing now is what I call a user sum. We’ve got the old revenue sums – where the revenue comes from – but because you essentially log in when you come into our building we can chart that. How many times does that fan come back? Why are they coming back? Were they noisy? Did they bring a guest? What did it all look like? And what do we do to improve not just the fan experience, but the fan involvement? At Microsoft you’d have these Windows fans, and not only were they good customers, but they’d help spread the word to others. They would help Windows succeed. Clippers fans need to help our team succeed. And so we came to this notion of having a complete map of their behavior.
   What has been your most memorable moment in the Intuit Dome so far?
   SB: It’s funny, but on this stuff I have more of an emotional memory than I do specifics, but I’ll give you two. Our first victory, after four straight defeats, that’s a top memory for me, no question. The second is the game we lost in the 2025 playoffs [game four, against the Denver Nuggets]. We were tied with 13 seconds left and the Denver Nuggets’ best player, maybe the best player in the world, Nikola Jokic, goes to take a three-point shot. He air balls it, but one of his teammates catches it midflight, dunks it and with less than one second left and they win the game. There used to be a show in the US called Wide World of Sports and they always talked about the joy of victory and the agony of defeat. And I guess I have one of each of those moments etched in my head. I would say those two things, but also opening night. Just having my closest people with me and being able to say, ‘Yeah! We built this’