KING OF NEWS

“Stock-market crash.” “AI bubble.” “Another 2008.” These claims fill my X feed. Too often, the evidence stops at a falling share price. My view: AI has speculative excess. Some valuations will disappoint. But I believe the “everything is collapsing” narrative exaggerates the problem and underestimates the technology’s practical value. If you’re comparing today with a previous crisis, show the comparison: leverage, credit conditions, defaults, earnings, cash flow and valuations. A frightening headline is not a historical analysis. 🤖 AI’S USEFULNESS IS ALREADY VISIBLE Amazon has deployed more than one million robots. They work alongside employees, moving goods and handling repetitive tasks. Amazon says its DeepFleet AI model can improve robot travel efficiency by 10%. That is a measurable operational application—not a hypothetical market opportunity. (aboutamazon.com) Figure AI—not Figma—has deployed humanoid robots at BMW. BMW reported that Figure 02 supported production of more than 30,000 vehicles over ten months. (press.bmwgroup.com) Tesla’s FSD (Supervised) is another physical-world application, although it still requires active driver supervision. Optimus represents a further development opportunity, rather than proof that fully autonomous factories have already arrived. (tesla.com, last10k.com) Nvidia’s agreement to acquire Hugging Face also shows the industry expanding beyond chips into tools that help developers build and deploy applications. An acquisition is a strategic bet; successful adoption and returns must follow. (blogs.nvidia.com) 💰 COMMERCIAL DEMAND IS SHOWING UP Microsoft reported an AI-business annual revenue run rate above $37B for the quarter ended March 2026, up 123% year over year. That is a revenue run rate—not annual profit. (microsoft.com) Meta reported Q3 revenue growth of 26% and says AI is generating incremental advertising revenue through better engagement, recommendations and monetization. It does not separately disclose an “AI profit” figure. (investor.atmeta.com) SpaceX’s Super Intelligence segment reported Q3 revenue of $7.56B and operating income of $431M. Those are commercial results worth examining alongside the enormous investment required. (sec.gov) 🏗️ THE SPENDING QUESTION IS REAL Data centers, chips, power, networking and leases require substantial capital. Revenue can grow while free cash flow deteriorates. Meta illustrates that tension: $32.79B in Q3 operating cash flow, but negative $2.12B in free cash flow after capital spending, including finance-lease principal payments. (investor.atmeta.com) Higher financing costs make that challenge harder for heavily indebted operators. But one company’s financing problem does not establish that the entire industry lacks demand. Suppliers such as $LITE benefit from the infrastructure buildout. Their demand is encouraging, but the ultimate test is whether customers can turn that infrastructure into useful services and sustainable returns. 🧠 MY CONCLUSION I believe AI can create enormous economic value. Many applications are still developing, and the benefits will arrive unevenly. That does not guarantee every investment will succeed. Spending more money is not, by itself, proof of a good investment. My approach: buy strong businesses at sensible valuations. Examine competitive advantages, balance sheets, cash generation and execution. Avoid weak companies surviving on an AI label and repeated fundraising. Don’t outsource your judgment to the loudest account on X. And if you buy a weak company simply because someone promises it is “the next Nvidia,” my confidence in AI is not a defense of that purchase. #AI #Investing #StockMarket

💻 𝗔𝗜 𝘄𝗿𝗶𝘁𝗲𝘀 𝗰𝗼𝗱𝗲 𝗳𝗮𝘀𝘁𝗲𝗿. 𝗪𝗵𝘆 𝗮𝗿𝗲 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀 𝘀𝘁𝗶𝗹𝗹 𝘄𝗮𝗶𝘁𝗶𝗻𝗴 𝘁𝗼 𝘀𝗵𝗶𝗽? Faster code generation can expose another constraint: validation, integration and deployment cannot keep up. Anthropic reported that, as of May 2026, Claude authored more than 80% of newly merged code. In Q2, the typical engineer merged roughly eight times as much code per day as in 2024. The company also cautioned that code volume does not translate directly into equivalent productivity gains. Writing code is one part of delivering software. 𝟭. 𝗧𝗵𝗲 𝗯𝗼𝘁𝘁𝗹𝗲𝗻𝗲𝗰𝗸 𝗺𝗮𝘁𝗵 Suppose a project requires 100 hours of sequential work: 30 hours writing code. 70 hours on requirements, review, testing, integration and other work. AI makes coding ten times faster. Those 30 hours become three. Total time: 70 + 3 = 73 hours. That is 27% less time, or approximately 37% faster overall—not a tenfold improvement. Even instantaneous coding would leave 70 hours. This simplified example illustrates Amdahl’s law: overall acceleration is constrained by the work that remains unchanged. 𝟮. 𝗙𝗮𝘀𝘁𝗲𝗿 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 𝗰𝗮𝗻 𝗲𝘅𝗽𝗮𝗻𝗱 𝘁𝗵𝗲 𝗿𝗲𝘃𝗶𝗲𝘄 𝗾𝘂𝗲𝘂𝗲 A large batch of overnight changes still raises questions: Does it meet the requirements? Does existing behavior remain intact? Were edge cases tested? Can failures be diagnosed and rolled back? Without greater validation capacity, saved coding time can become waiting time elsewhere. 𝟯. 𝗖𝗼𝗱𝗲 𝗰𝗮𝗻 𝗼𝘂𝘁𝗿𝘂𝗻 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 A change may look reasonable in isolation while altering system-wide behavior. When a team accepts changes without understanding their assumptions and dependencies, it accumulates “cognitive debt.” A later incident requires both a repair and the recovery of missing understanding. Human-written code can create this problem too. The question is whether output exceeds the team’s capacity to understand and validate it. 𝟰. 𝗙𝗶𝘃𝗲 𝗽𝗿𝗼𝘁𝗼𝘁𝘆𝗽𝗲𝘀 𝗮𝗿𝗲 𝗳𝗶𝘃𝗲 𝘂𝗻𝗳𝗶𝗻𝗶𝘀𝗵𝗲𝗱 𝗰𝗼𝗺𝗺𝗶𝘁𝗺𝗲𝗻𝘁𝘀 AI makes starting projects cheaper. A team can quickly accumulate impressive demos while testing, integration and acceptance remain backlogged. More work starts. Completed deliveries may not increase proportionally. Track completion, waiting time and rework alongside activity. 𝟱. 𝗔𝗰𝗰𝗲𝗹𝗲𝗿𝗮𝘁𝗲 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲 AI can also assist with testing, review, debugging and documentation. Anthropic’s response includes scaling automated review while retaining human accountability. Useful measures therefore include: Time from requirements to production. Reliability after deployment. Defects and rework. The team’s ability to maintain what it ships. The business value of AI coding ultimately has to appear in reliable delivery. Where is your team’s biggest bottleneck today: implementation, review, testing—or deciding what to build? #AI #Claude #Anthropic #SoftwareEngineering #AICoding

💻 AI把代码写快了，为什么项目还是交不出去？ 代码生成速度提高，可能让公司的另一个问题暴露得更快：验证、集成和上线，跟不上了。 Anthropic披露，截至2026年5月，新合并代码中超过𝟴𝟬%由Claude编写。2026年第二季度，典型工程师每天合并的代码量，约为2024年的𝟴倍。 但公司也明确提醒：代码行数衡量的是数量，不能直接代表同等幅度的生产率提升。 这里藏着AI编程最容易被误读的一件事：写代码，只是交付软件的一部分。 ① 一个简单算式，就能看出瓶颈在哪里 假设一个项目需要100小时，各环节顺序完成： 写代码占30小时。 需求、审查、测试、集成等工作占70小时。 AI让写代码快了10倍，30小时缩短为3小时。 项目总耗时变成： 𝟳𝟬＋𝟯＝𝟳𝟯小时。 编程速度提高10倍，项目耗时减少27%，整体速度约提升37%。 即使代码瞬间写完，剩余工作仍需要70小时。 这就是阿姆达尔定律在这个简化例子中的含义：整体提速，受到未被加速环节的限制。 ② 生成得越快，待验证的工作也可能越多 以前，团队可能在等代码写出来。 现在，一夜之间就能产生大量修改，第二天却需要回答： 功能符合需求吗？ 会不会破坏旧逻辑？ 异常情况测过没有？ 出了问题，能否定位和回滚？ 如果验证能力没有同步扩大，节省下来的编程时间，就可能变成审查队列中的等待时间。 ③ 还有一种成本：代码增长快于团队的理解 一个修改看起来合理，放进整个系统后，却可能改变原有行为。 如果团队接受了修改，却讲不清它为什么正确、影响哪些模块、依赖哪些假设，就会积累一种“认知债务”。 问题发生时，团队需要补上的不只是修复代码，还包括重新理解系统。 这类成本可能来自AI生成的代码，也可能来自人写的代码。关键在于，产出速度有没有超过团队理解和验证的能力。 ④ 同时开五个项目，不等于能交付五个产品 AI降低了启动成本，很容易让团队同时推进更多想法。 五个原型都能演示，五个项目都像快完成了，但测试、集成和验收仍然排着长队。 于是，开始的事情越来越多，真正完成的事情没有同比增加。 衡量效率时，除了看启动了多少项目，还要看交付了多少、等待多久、返工多少。 ⑤ 下一阶段，需要加速整条交付流程 AI也可以帮助写测试、审查修改、定位故障和整理文档。 Anthropic自己的应对方式，就包括扩大自动化审查，同时保留相应的人类责任。 因此，值得追踪的指标应当包括： 从需求到上线需要多久？ 修改进入生产后是否稳定？ 缺陷和返工有没有减少？ 团队是否理解并能维护交付的系统？ AI编程的商业价值，最终要通过可靠交付体现。 你所在团队现在最大的瓶颈，是实现、审查、测试，还是需求本身？ #AI #Claude #Anthropic #AI编程 #软件工程

🏀 𝗝𝗼𝗸𝗶𝗰́ 𝘄𝗮𝗻𝘁𝘀 𝗳𝗮𝗻𝘀’ 𝘀𝘂𝗽𝗽𝗼𝗿𝘁. 𝗛𝗲 𝗮𝗹𝘀𝗼 𝗵𝗮𝘀 𝘁𝗼 𝗮𝗰𝗰𝗲𝗽𝘁 𝘀𝗰𝗿𝘂𝘁𝗶𝗻𝘆. Denver lost 134–114 to Golden State on October 26. Nikola Jokić finished with 19 points, seven rebounds, 11 assists and three blocks, shooting 1-for-6 from three. Calling that “no contribution” would be misleading. But 11 assists cannot answer every question about his defense. 𝟭. 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗲 𝘁𝗵𝗲 𝗽𝗼𝘀𝘀𝗲𝘀𝘀𝗶𝗼𝗻𝘀 Shrugging, frustration and disappointed expressions do not prove that a player has quit. The useful questions are more specific: Did he contest after getting back? Did the next rotation arrive after a teammate helped? Was contact established before the rebound? Did his movement match his defensive communication? Those details provide a better basis for criticism than facial expressions. 𝟮. 𝗗𝗶𝗿𝗲𝗰𝘁𝗶𝗻𝗴 𝘁𝗲𝗮𝗺𝗺𝗮𝘁𝗲𝘀 𝗶𝘀 𝗼𝗻𝗹𝘆 𝗽𝗮𝗿𝘁 𝗼𝗳 𝗱𝗲𝗳𝗲𝗻𝗱𝗶𝗻𝗴 A center’s communication can organize switches, help and recovery. Execution still requires all five players to move. If one defender closes out and the next rotation never arrives, the offense keeps its advantage. Calling the coverage does not complete the possession. Jokić’s execution deserves the same scrutiny as everyone else’s. 𝟯. 𝗗𝗲𝗻𝘃𝗲𝗿’𝘀 𝗳𝗮𝗶𝗹𝘂𝗿𝗲 𝘄𝗮𝘀 𝗰𝗼𝗹𝗹𝗲𝗰𝘁𝗶𝘃𝗲 Golden State shot 54.1%, made 22 of 44 threes and won the rebounding battle 50–36. Denver conceded 74 second-half points. Perimeter containment, ball-screen coverage, weak-side rotations and rebounding all require review. Jokić must carry the responsibility of a franchise player. That does not remove responsibility from his teammates. 𝟰. 𝗧𝗵𝗲 𝗯𝗼𝗼𝗶𝗻𝗴 𝗰𝗼𝗺𝗺𝗲𝗻𝘁𝘀 𝗻𝗲𝗲𝗱 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 Against New Orleans, Denver recovered from a 23-point deficit to win 141–135 in overtime. Jokić produced 35 points, 15 rebounds and 10 assists. Afterward, he criticized home fans who booed and defended the team’s refusal to quit. That comeback supported his argument. Fans also had a reason to be frustrated by the early deficit and defensive breakdowns. Players can ask for support. Fans can criticize the basketball they are watching. Both positions can coexist. 𝟱. 𝗔 𝗵𝗶𝗴𝗵 𝘀𝘁𝗮𝗻𝗱𝗮𝗿𝗱 𝗺𝘂𝘀𝘁 𝗮𝗽𝗽𝗹𝘆 𝗰𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝘁𝗹𝘆 Jokić has spoken about European basketball’s competitive pressure and wanting NBA players to experience it. That standard should also encompass transition defense, box-outs, rotations and responses to mistakes. One loss cannot establish that he does not care. Three games cannot establish that his ability has declined. An MVP-caliber player should nevertheless face MVP-level scrutiny. The fair question: when Denver’s defense loses its structure, can its best player lead the execution that restores it? #NBA #NikolaJokic #Nuggets #DenverNuggets #Basketball

🏀 约基奇要求球迷别嘘球队，球队核心也必须接受审视 10月26日，掘金114比134不敌勇士。 约基奇得到𝟭𝟵分、𝟳个篮板、𝟭𝟭次助攻、𝟯次封盖，三分球6投1中。 这份数据不能简单概括为“毫无贡献”。但11次助攻，也无法回答所有关于防守表现的质疑。 ① 批评应该落在具体回合，而不是猜测他想不想赢 摊手、抱头、表情沮丧，都不足以证明一名球员已经放弃。 真正值得逐回合检查的是： 退防到位后，有没有及时形成干扰？ 队友补位后，下一次轮转有没有跟上？ 篮板落下前，有没有先完成卡位？ 被挡拆针对时，沟通和实际移动是否一致？ 这些问题，比“看起来不高兴”更能说明防守质量。 ② 指挥队友轮转，也要完成自己的防守责任 中锋可以通过沟通帮助队友判断换防、补位和收缩。 但防守最终需要五个人一起移动。 如果队友已经扑向持球人，另一侧却没有及时接上，进攻方仍然会得到空位。谁在说话，不等于谁已经完成任务。 作为球队核心，约基奇的防守执行，同样应该接受这套标准的检验。 ③ 这场失利，也不能全部归因于一个人 勇士全场投篮命中率54.1%，三分44投22中，并以50比36赢下篮板。 掘金下半场丢了74分。 这样的结果需要全队复盘：外线防突破、挡拆处理、弱侧轮转和篮板保护，都不能只靠中锋一个人解决。 约基奇需要承担核心责任，其他位置也不能因此免于问责。 ④ 关于嘘声，必须把前因后果说完整 此前对鹈鹕，掘金曾落后23分，最终通过加时141比135逆转。 约基奇得到35分、15个篮板、10次助攻，赛后对主场嘘声表达不满，强调球队不会轻易放弃。 那场逆转，确实支持了他的说法。 但球迷的不满也有具体背景：主场开局被大比分领先，防守持续失守。 球员可以希望获得支持，球迷也可以批评比赛内容。两者并不矛盾。 ⑤ 越强调赢球，越需要接受同样严格的要求 约基奇此前谈过欧洲篮球的竞争压力，希望NBA球员体验那种每场都必须争胜的环境。 既然如此，检验标准也应该一致： 球队顺利时，看组织和得分。 球队困难时，也看退防、卡位、补防和下一回合的反应。 一场输球，不能证明约基奇不在乎胜负。 三场比赛，也不足以判断他的实力已经下滑。 但MVP级别的核心，当然应该接受MVP级别的审视。 球迷最有理由追问的是：当球队防守失序时，核心能否率先把执行带回来？ #NBA #NikolaJokic #Nuggets #约基奇 #掘金

🏀 𝗪𝗵𝘆 𝗲𝘃𝗲𝗻 𝗣𝗮𝘂𝗹 𝗚𝗲𝗼𝗿𝗴𝗲 𝗵𝗮𝘀 𝘁𝗼 𝗮𝗱𝗷𝘂𝘀𝘁 𝘁𝗼 𝗕𝗼𝘀𝘁𝗼𝗻’𝘀 𝗼𝗳𝗳𝗲𝗻𝘀𝗲 Understanding a simple flare screen helps explain Jayson Tatum’s game-winner with 2.3 seconds remaining. The challenge is making repeated decisions about who handles, screens, cuts and finishes. 𝟭. 𝗢𝗻𝗲 𝘀𝗰𝗿𝗲𝗲𝗻, 𝘀𝗲𝘃𝗲𝗿𝗮𝗹 𝗼𝘂𝘁𝗰𝗼𝗺𝗲𝘀 A flare screen helps an off-ball teammate move away from the ball toward a potential catch. That can produce an open three. If defenders chase the receiver, the screener may cut to the rim. If they switch, the offense may gain a favorable matchup. The same action creates several possible advantages. 𝟮. 𝗧𝗶𝗺𝗶𝗻𝗴 𝗺𝗮𝗸𝗲𝘀 𝘁𝗵𝗲 𝗮𝗰𝘁𝗶𝗼𝗻 𝗺𝗼𝗿𝗲 𝗱𝗮𝗻𝗴𝗲𝗿𝗼𝘂𝘀 Early offense attacks while the defense is still organizing. Players may still be finding assignments, communicating switches or establishing help positions. A routine screen can therefore produce an unusually clean opportunity: the trailing defender gets caught, the switch arrives late, or nobody is ready to protect the rim. 𝟯. 𝗣𝗹𝗮𝘆𝗲𝗿𝘀 𝗺𝘂𝘀𝘁 𝗿𝗲𝗮𝗱 𝘁𝗵𝗲𝗶𝗿 𝘁𝗲𝗮𝗺𝗺𝗮𝘁𝗲𝘀 Roles can change from possession to possession. An advancing player can establish spacing. The ball handler surveys the floor. Trailing teammates can flow into screening actions. For a star accustomed to familiar touches and initiation points, this requires an adjustment: finding ways to contribute from changing positions. 𝟰. 𝗧𝗵𝗲𝗻 𝘁𝗵𝗲𝘆 𝗺𝘂𝘀𝘁 𝗿𝗲𝗮𝗱 𝘁𝗵𝗲 𝗱𝗲𝗳𝗲𝗻𝘀𝗲 A defender gets caught: look for the catch-and-shoot opportunity. The defense overplays the receiver: explore the screener’s cut. The defense switches: examine the matchup. Help loads up on one side: attack the available space. If the first route closes, a fake screen, reversed screen or follow-up ball screen can create another decision. This freedom requires shared spacing rules and coordinated reactions. 𝟱. 𝗧𝗮𝘁𝘂𝗺’𝘀 𝗴𝗮𝗺𝗲-𝘄𝗶𝗻𝗻𝗲𝗿 𝗳𝗼𝗹𝗹𝗼𝘄𝗲𝗱 𝘁𝗵𝗮𝘁 𝗹𝗼𝗴𝗶𝗰 Boston trailed by one with 2.3 seconds left in overtime. Tatum and Payton Pritchard worked together, forcing the defense to decide how to handle their screening action. Tatum then used his body to establish position, created separation for the catch and finished the possession. Three connected elements made it possible: The screen forced a decision. Positioning created an advantage. The star completed the difficult shot. Final score: Boston 125, Brooklyn 124. That also helps explain George’s adjustment. Individual ability remains essential, while players learn to apply it from different positions within a shared offense. A strong system provides clear principles and room for judgment. When watching basketball, do you prefer studying set-play execution or these continuous defensive reads? #NBA #Celtics #JaysonTatum #PaulGeorge #Basketball

🏀 连保罗·乔治都要重新适应：凯尔特人的进攻，到底难在哪里？ 看懂一个远切掩护，就能更好地理解塔图姆最后𝟮.𝟯秒的绝杀。 难点在于，球员需要不断判断：这一刻，该由谁持球、谁掩护、谁切入？ ① 同一个掩护，为什么能打出不同结果？ Flare screen，通常译为远切掩护：一名球员给无球队友掩护，帮助他向远离球的位置移动，寻找接球机会。 最直观的结果是外线空位。 但防守如果追着接球队员跑，掩护者就可能突然切向篮下；如果选择换防，又可能留下错位。 一个简单动作，可以同时制造三分、空切和错位机会。 ② 关键在于，防守还没站稳，配合已经开始 这就是Early offense——早攻。 球队在推进过程中就发动配合，利用对手尚未完成对位、协防和沟通的时间窗口。 此时，一次普通掩护也可能制造很大麻烦： 追防的人被挡住。 换防的人没有及时接上。 篮下的协防还没到位。 于是，看似轻松的空篮，来自进攻比防守早完成了一次判断。 ③ 球员首先要阅读队友，才能确定自己的角色 在这种进攻中，同一个人每回合的任务可能不同。 跑在前面，可以先拉开空间。 持球推进，需要观察进攻方向。 拖在后面，可以顺势参与掩护。 队友吸引防守，就要准备切入或接应。 对习惯围绕自己展开进攻的球星来说，这种变化尤其明显。 过去熟悉的是固定的接球点和发起方式，现在需要根据队友的位置，随时找到新的参与方式。 ④ 读完队友，还要读防守 防守绕不过掩护，就寻找接球投篮。 防守过度追赶，就让掩护者切入。 防守换人，就寻找有利对位。 一侧协防收得很深，另一侧就可能出现突破空间。 如果原来的路线被封住，还可以通过假掩护、反向掩护或衔接挡拆，继续制造下一次选择。 看起来变化很多，背后依赖的是共同的空间规则和连续判断。 “自由进攻”同样需要纪律：每个人都要让队友知道，下一步可能发生什么。 ⑤ 塔图姆的绝杀，延续了同样的阅读逻辑 加时赛最后，凯尔特人只剩𝟮.𝟯秒，落后1分。 塔图姆与普理查德先完成掩护配合，让防守必须立刻决定是否换人。 随后，塔图姆利用身体争取接球位置，拉开接球空间，再完成最后的投篮。 这次进攻的价值，可以拆成三个环节： 掩护制造防守选择。 站位争取有利对位。 球星完成高难度终结。 最终比分：𝟭𝟮𝟱比𝟭𝟮𝟰。 理解这套体系，也就更容易理解乔治的适应过程：个人能力仍然重要，但球员还要学会在不断变化的位置上，把自己的能力接入团队进攻。 一套好的体系，既需要清晰的规则，也需要给球员判断和发挥的空间。 你看球时，更喜欢研究固定战术的执行，还是这种连续阅读防守的进攻？ #NBA #Celtics #JaysonTatum #PaulGeorge #篮球战术🏀 连保罗·乔治都要重新适应：凯尔特人的进攻，到底难在哪里？ 看懂一个远切掩护，就能更好地理解塔图姆最后𝟮.𝟯秒的绝杀。 难点在于，球员需要不断判断：这一刻，该由谁持球、谁掩护、谁切入？ ① 同一个掩护，为什么能打出不同结果？ Flare screen，通常译为远切掩护：一名球员给无球队友掩护，帮助他向远离球的位置移动，寻找接球机会。 最直观的结果是外线空位。 但防守如果追着接球队员跑，掩护者就可能突然切向篮下；如果选择换防，又可能留下错位。 一个简单动作，可以同时制造三分、空切和错位机会。 ② 关键在于，防守还没站稳，配合已经开始 这就是Early offense——早攻。 球队在推进过程中就发动配合，利用对手尚未完成对位、协防和沟通的时间窗口。 此时，一次普通掩护也可能制造很大麻烦： 追防的人被挡住。 换防的人没有及时接上。 篮下的协防还没到位。 于是，看似轻松的空篮，来自进攻比防守早完成了一次判断。 ③ 球员首先要阅读队友，才能确定自己的角色 在这种进攻中，同一个人每回合的任务可能不同。 跑在前面，可以先拉开空间。 持球推进，需要观察进攻方向。 拖在后面，可以顺势参与掩护。 队友吸引防守，就要准备切入或接应。 对习惯围绕自己展开进攻的球星来说，这种变化尤其明显。 过去熟悉的是固定的接球点和发起方式，现在需要根据队友的位置，随时找到新的参与方式。 ④ 读完队友，还要读防守 防守绕不过掩护，就寻找接球投篮。 防守过度追赶，就让掩护者切入。 防守换人，就寻找有利对位。 一侧协防收得很深，另一侧就可能出现突破空间。 如果原来的路线被封住，还可以通过假掩护、反向掩护或衔接挡拆，继续制造下一次选择。 看起来变化很多，背后依赖的是共同的空间规则和连续判断。 “自由进攻”同样需要纪律：每个人都要让队友知道，下一步可能发生什么。 ⑤ 塔图姆的绝杀，延续了同样的阅读逻辑 加时赛最后，凯尔特人只剩𝟮.𝟯秒，落后1分。 塔图姆与普理查德先完成掩护配合，让防守必须立刻决定是否换人。 随后，塔图姆利用身体争取接球位置，拉开接球空间，再完成最后的投篮。 这次进攻的价值，可以拆成三个环节： 掩护制造防守选择。 站位争取有利对位。 球星完成高难度终结。 最终比分：𝟭𝟮𝟱比𝟭𝟮𝟰。 理解这套体系，也就更容易理解乔治的适应过程：个人能力仍然重要，但球员还要学会在不断变化的位置上，把自己的能力接入团队进攻。 一套好的体系，既需要清晰的规则，也需要给球员判断和发挥的空间。 你看球时，更喜欢研究固定战术的执行，还是这种连续阅读防守的进攻？ #NBA #Celtics #JaysonTatum #PaulGeorge #篮球战术🏀 连保罗·乔治都要重新适应：凯尔特人的进攻，到底难在哪里？ 看懂一个远切掩护，就能更好地理解塔图姆最后𝟮.𝟯秒的绝杀。 难点在于，球员需要不断判断：这一刻，该由谁持球、谁掩护、谁切入？ ① 同一个掩护，为什么能打出不同结果？ Flare screen，通常译为远切掩护：一名球员给无球队友掩护，帮助他向远离球的位置移动，寻找接球机会。 最直观的结果是外线空位。 但防守如果追着接球队员跑，掩护者就可能突然切向篮下；如果选择换防，又可能留下错位。 一个简单动作，可以同时制造三分、空切和错位机会。 ② 关键在于，防守还没站稳，配合已经开始 这就是Early offense——早攻。 球队在推进过程中就发动配合，利用对手尚未完成对位、协防和沟通的时间窗口。 此时，一次普通掩护也可能制造很大麻烦： 追防的人被挡住。 换防的人没有及时接上。 篮下的协防还没到位。 于是，看似轻松的空篮，来自进攻比防守早完成了一次判断。 ③ 球员首先要阅读队友，才能确定自己的角色 在这种进攻中，同一个人每回合的任务可能不同。 跑在前面，可以先拉开空间。 持球推进，需要观察进攻方向。 拖在后面，可以顺势参与掩护。 队友吸引防守，就要准备切入或接应。 对习惯围绕自己展开进攻的球星来说，这种变化尤其明显。 过去熟悉的是固定的接球点和发起方式，现在需要根据队友的位置，随时找到新的参与方式。 ④ 读完队友，还要读防守 防守绕不过掩护，就寻找接球投篮。 防守过度追赶，就让掩护者切入。 防守换人，就寻找有利对位。 一侧协防收得很深，另一侧就可能出现突破空间。 如果原来的路线被封住，还可以通过假掩护、反向掩护或衔接挡拆，继续制造下一次选择。 看起来变化很多，背后依赖的是共同的空间规则和连续判断。 “自由进攻”同样需要纪律：每个人都要让队友知道，下一步可能发生什么。 ⑤ 塔图姆的绝杀，延续了同样的阅读逻辑 加时赛最后，凯尔特人只剩𝟮.𝟯秒，落后1分。 塔图姆与普理查德先完成掩护配合，让防守必须立刻决定是否换人。 随后，塔图姆利用身体争取接球位置，拉开接球空间，再完成最后的投篮。 这次进攻的价值，可以拆成三个环节： 掩护制造防守选择。 站位争取有利对位。 球星完成高难度终结。 最终比分：𝟭𝟮𝟱比𝟭𝟮𝟰。 理解这套体系，也就更容易理解乔治的适应过程：个人能力仍然重要，但球员还要学会在不断变化的位置上，把自己的能力接入团队进攻。 一套好的体系，既需要清晰的规则，也需要给球员判断和发挥的空间。 你看球时，更喜欢研究固定战术的执行，还是这种连续阅读防守的进攻？ #NBA #Celtics #JaysonTatum #PaulGeorge #篮球战术

🏀 𝗟𝗲𝗕𝗿𝗼𝗻’𝘀 𝗽𝘂𝘁𝗯𝗮𝗰𝗸 𝘄𝗶𝘁𝗵 𝟭𝟭.𝟵 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 𝗹𝗲𝗳𝘁: 𝘄𝗵𝘆 𝘄𝗮𝘀 𝗵𝗲 𝗶𝗻 𝘁𝗵𝗮𝘁 𝘀𝗽𝗼𝘁? Tyrese Maxey missed near the rim. LeBron James finished the putback, extending Philadelphia’s lead to 96–91. The Sixers ultimately beat the Knicks 98–93. The highlight shows the finish. The tactical question begins earlier: why was LeBron already positioned to make that play? Start with the dunker spot. 𝟭. 𝗙𝗿𝗼𝗺 𝗶𝗻𝗶𝘁𝗶𝗮𝘁𝗼𝗿 𝘁𝗼 𝗳𝗶𝗻𝗶𝘀𝗵𝗲𝗿 With LeBron handling above the arc, Philadelphia can run actions with Joel Embiid while Maxey moves off the ball. When Maxey takes over, LeBron can contribute closer to the basket through his strength, defensive reads and finishing. Exchanging responsibilities changes the problems the defense must solve. 𝟮. 𝗪𝗵𝘆 𝘁𝗵𝗲 𝗱𝘂𝗻𝗸𝗲𝗿 𝘀𝗽𝗼𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 This is the short-corner area near the baseline, beside the basket. A player stationed there can receive an interior pass, cut behind the defense or attack the offensive glass. The nearby defender faces a difficult choice: Help on the ball, potentially exposing the player behind him. Stay home, potentially giving the driver a clearer path. LeBron’s size and anticipation make that decision especially uncomfortable. 𝟯. 𝗦𝗶𝗺𝗼𝗻𝘀’ 𝘀𝗵𝗼𝗼𝘁𝗶𝗻𝗴 𝗮𝗱𝗱𝗲𝗱 𝗳𝗹𝗲𝘅𝗶𝗯𝗶𝗹𝗶𝘁𝘆 Anfernee Simons scored 21 points and made five of eight threes. Alongside Maxey and VJ Edgecombe, his shooting gives Philadelphia another perimeter threat, allowing LeBron to contribute as an interior target. The important questions become who handles, who spaces the floor and who punishes help. 𝟰. 𝗟𝗲𝗕𝗿𝗼𝗻’𝘀 𝗽𝗹𝗮𝘆 𝗯𝗲𝗴𝗮𝗻 𝗯𝗲𝗳𝗼𝗿𝗲 𝘁𝗵𝗲 𝗺𝗶𝘀𝘀 The final possession can be understood in two layers. Maxey attacks the rim and draws defensive attention. LeBron prepares near the basket for a pass or a rebounding opportunity. Once defenders react to the ball, positioning, strength and timing can create the chance to finish. That does not mean the offense planned for a miss. Good positioning preserves a second opportunity when the first option fails. 𝟱. 𝗢𝗳𝗳-𝗯𝗮𝗹𝗹 𝗿𝗲𝗮𝗱𝘀 𝗰𝗮𝗻 𝗱𝗲𝗰𝗶𝗱𝗲 𝗴𝗮𝗺𝗲𝘀 LeBron finished with 20 points, nine rebounds and eight assists. His closing putback illustrated another part of his value: recognizing when to give up the ball, move toward the rim and prepare for the next opportunity. The groundwork for a decisive basket often begins before the shot goes up. Which part of LeBron’s game do you appreciate more: his playmaking with the ball, or his anticipation without it? #NBA #LeBronJames #Sixers #76ers #Basketball 🏀 𝗟𝗲𝗕𝗿𝗼𝗻’𝘀 𝗽𝘂𝘁𝗯𝗮𝗰𝗸 𝘄𝗶𝘁𝗵 𝟭𝟭.𝟵 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 𝗹𝗲𝗳𝘁: 𝘄𝗵𝘆 𝘄𝗮𝘀 𝗵𝗲 𝗶𝗻 𝘁𝗵𝗮𝘁 𝘀𝗽𝗼𝘁? Tyrese Maxey missed near the rim. LeBron James finished the putback, extending Philadelphia’s lead to 96–91. The Sixers ultimately beat the Knicks 98–93. The highlight shows the finish. The tactical question begins earlier: why was LeBron already positioned to make that play? Start with the dunker spot. 𝟭. 𝗙𝗿𝗼𝗺 𝗶𝗻𝗶𝘁𝗶𝗮𝘁𝗼𝗿 𝘁𝗼 𝗳𝗶𝗻𝗶𝘀𝗵𝗲𝗿 With LeBron handling above the arc, Philadelphia can run actions with Joel Embiid while Maxey moves off the ball. When Maxey takes over, LeBron can contribute closer to the basket through his strength, defensive reads and finishing. Exchanging responsibilities changes the problems the defense must solve. 𝟮. 𝗪𝗵𝘆 𝘁𝗵𝗲 𝗱𝘂𝗻𝗸𝗲𝗿 𝘀𝗽𝗼𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 This is the short-corner area near the baseline, beside the basket. A player stationed there can receive an interior pass, cut behind the defense or attack the offensive glass. The nearby defender faces a difficult choice: Help on the ball, potentially exposing the player behind him. Stay home, potentially giving the driver a clearer path. LeBron’s size and anticipation make that decision especially uncomfortable. 𝟯. 𝗦𝗶𝗺𝗼𝗻𝘀’ 𝘀𝗵𝗼𝗼𝘁𝗶𝗻𝗴 𝗮𝗱𝗱𝗲𝗱 𝗳𝗹𝗲𝘅𝗶𝗯𝗶𝗹𝗶𝘁𝘆 Anfernee Simons scored 21 points and made five of eight threes. Alongside Maxey and VJ Edgecombe, his shooting gives Philadelphia another perimeter threat, allowing LeBron to contribute as an interior target. The important questions become who handles, who spaces the floor and who punishes help. 𝟰. 𝗟𝗲𝗕𝗿𝗼𝗻’𝘀 𝗽𝗹𝗮𝘆 𝗯𝗲𝗴𝗮𝗻 𝗯𝗲𝗳𝗼𝗿𝗲 𝘁𝗵𝗲 𝗺𝗶𝘀𝘀 The final possession can be understood in two layers. Maxey attacks the rim and draws defensive attention. LeBron prepares near the basket for a pass or a rebounding opportunity. Once defenders react to the ball, positioning, strength and timing can create the chance to finish. That does not mean the offense planned for a miss. Good positioning preserves a second opportunity when the first option fails. 𝟱. 𝗢𝗳𝗳-𝗯𝗮𝗹𝗹 𝗿𝗲𝗮𝗱𝘀 𝗰𝗮𝗻 𝗱𝗲𝗰𝗶𝗱𝗲 𝗴𝗮𝗺𝗲𝘀 LeBron finished with 20 points, nine rebounds and eight assists. His closing putback illustrated another part of his value: recognizing when to give up the ball, move toward the rim and prepare for the next opportunity. The groundwork for a decisive basket often begins before the shot goes up. Which part of LeBron’s game do you appreciate more: his playmaking with the ball, or his anticipation without it? #NBA #LeBronJames #Sixers #76ers #Basketball 🏀 𝗟𝗲𝗕𝗿𝗼𝗻’𝘀 𝗽𝘂𝘁𝗯𝗮𝗰𝗸 𝘄𝗶𝘁𝗵 𝟭𝟭.𝟵 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 𝗹𝗲𝗳𝘁: 𝘄𝗵𝘆 𝘄𝗮𝘀 𝗵𝗲 𝗶𝗻 𝘁𝗵𝗮𝘁 𝘀𝗽𝗼𝘁? Tyrese Maxey missed near the rim. LeBron James finished the putback, extending Philadelphia’s lead to 96–91. The Sixers ultimately beat the Knicks 98–93. The highlight shows the finish. The tactical question begins earlier: why was LeBron already positioned to make that play? Start with the dunker spot. 𝟭. 𝗙𝗿𝗼𝗺 𝗶𝗻𝗶𝘁𝗶𝗮𝘁𝗼𝗿 𝘁𝗼 𝗳𝗶𝗻𝗶𝘀𝗵𝗲𝗿 With LeBron handling above the arc, Philadelphia can run actions with Joel Embiid while Maxey moves off the ball. When Maxey takes over, LeBron can contribute closer to the basket through his strength, defensive reads and finishing. Exchanging responsibilities changes the problems the defense must solve. 𝟮. 𝗪𝗵𝘆 𝘁𝗵𝗲 𝗱𝘂𝗻𝗸𝗲𝗿 𝘀𝗽𝗼𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 This is the short-corner area near the baseline, beside the basket. A player stationed there can receive an interior pass, cut behind the defense or attack the offensive glass. The nearby defender faces a difficult choice: Help on the ball, potentially exposing the player behind him. Stay home, potentially giving the driver a clearer path. LeBron’s size and anticipation make that decision especially uncomfortable. 𝟯. 𝗦𝗶𝗺𝗼𝗻𝘀’ 𝘀𝗵𝗼𝗼𝘁𝗶𝗻𝗴 𝗮𝗱𝗱𝗲𝗱 𝗳𝗹𝗲𝘅𝗶𝗯𝗶𝗹𝗶𝘁𝘆 Anfernee Simons scored 21 points and made five of eight threes. Alongside Maxey and VJ Edgecombe, his shooting gives Philadelphia another perimeter threat, allowing LeBron to contribute as an interior target. The important questions become who handles, who spaces the floor and who punishes help. 𝟰. 𝗟𝗲𝗕𝗿𝗼𝗻’𝘀 𝗽𝗹𝗮𝘆 𝗯𝗲𝗴𝗮𝗻 𝗯𝗲𝗳𝗼𝗿𝗲 𝘁𝗵𝗲 𝗺𝗶𝘀𝘀 The final possession can be understood in two layers. Maxey attacks the rim and draws defensive attention. LeBron prepares near the basket for a pass or a rebounding opportunity. Once defenders react to the ball, positioning, strength and timing can create the chance to finish. That does not mean the offense planned for a miss. Good positioning preserves a second opportunity when the first option fails. 𝟱. 𝗢𝗳𝗳-𝗯𝗮𝗹𝗹 𝗿𝗲𝗮𝗱𝘀 𝗰𝗮𝗻 𝗱𝗲𝗰𝗶𝗱𝗲 𝗴𝗮𝗺𝗲𝘀 LeBron finished with 20 points, nine rebounds and eight assists. His closing putback illustrated another part of his value: recognizing when to give up the ball, move toward the rim and prepare for the next opportunity. The groundwork for a decisive basket often begins before the shot goes up. Which part of LeBron’s game do you appreciate more: his playmaking with the ball, or his anticipation without it? #NBA #LeBronJames #Sixers #76ers #Basketball 🏀 𝗟𝗲𝗕𝗿𝗼𝗻’𝘀 𝗽𝘂𝘁𝗯𝗮𝗰𝗸 𝘄𝗶𝘁𝗵 𝟭𝟭.𝟵 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 𝗹𝗲𝗳𝘁: 𝘄𝗵𝘆 𝘄𝗮𝘀 𝗵𝗲 𝗶𝗻 𝘁𝗵𝗮𝘁 𝘀𝗽𝗼𝘁? Tyrese Maxey missed near the rim. LeBron James finished the putback, extending Philadelphia’s lead to 96–91. The Sixers ultimately beat the Knicks 98–93. The highlight shows the finish. The tactical question begins earlier: why was LeBron already positioned to make that play? Start with the dunker spot. 𝟭. 𝗙𝗿𝗼𝗺 𝗶𝗻𝗶𝘁𝗶𝗮𝘁𝗼𝗿 𝘁𝗼 𝗳𝗶𝗻𝗶𝘀𝗵𝗲𝗿 With LeBron handling above the arc, Philadelphia can run actions with Joel Embiid while Maxey moves off the ball. When Maxey takes over, LeBron can contribute closer to the basket through his strength, defensive reads and finishing. Exchanging responsibilities changes the problems the defense must solve. 𝟮. 𝗪𝗵𝘆 𝘁𝗵𝗲 𝗱𝘂𝗻𝗸𝗲𝗿 𝘀𝗽𝗼𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 This is the short-corner area near the baseline, beside the basket. A player stationed there can receive an interior pass, cut behind the defense or attack the offensive glass. The nearby defender faces a difficult choice: Help on the ball, potentially exposing the player behind him. Stay home, potentially giving the driver a clearer path. LeBron’s size and anticipation make that decision especially uncomfortable. 𝟯. 𝗦𝗶𝗺𝗼𝗻𝘀’ 𝘀𝗵𝗼𝗼𝘁𝗶𝗻𝗴 𝗮𝗱𝗱𝗲𝗱 𝗳𝗹𝗲𝘅𝗶𝗯𝗶𝗹𝗶𝘁𝘆 Anfernee Simons scored 21 points and made five of eight threes. Alongside Maxey and VJ Edgecombe, his shooting gives Philadelphia another perimeter threat, allowing LeBron to contribute as an interior target. The important questions become who handles, who spaces the floor and who punishes help. 𝟰. 𝗟𝗲𝗕𝗿𝗼𝗻’𝘀 𝗽𝗹𝗮𝘆 𝗯𝗲𝗴𝗮𝗻 𝗯𝗲𝗳𝗼𝗿𝗲 𝘁𝗵𝗲 𝗺𝗶𝘀𝘀 The final possession can be understood in two layers. Maxey attacks the rim and draws defensive attention. LeBron prepares near the basket for a pass or a rebounding opportunity. Once defenders react to the ball, positioning, strength and timing can create the chance to finish. That does not mean the offense planned for a miss. Good positioning preserves a second opportunity when the first option fails. 𝟱. 𝗢𝗳𝗳-𝗯𝗮𝗹𝗹 𝗿𝗲𝗮𝗱𝘀 𝗰𝗮𝗻 𝗱𝗲𝗰𝗶𝗱𝗲 𝗴𝗮𝗺𝗲𝘀 LeBron finished with 20 points, nine rebounds and eight assists. His closing putback illustrated another part of his value: recognizing when to give up the ball, move toward the rim and prepare for the next opportunity. The groundwork for a decisive basket often begins before the shot goes up. Which part of LeBron’s game do you appreciate more: his playmaking with the ball, or his anticipation without it? #NBA #LeBronJames #Sixers #76ers #Basketball

🏀 詹姆斯最后𝟭𝟭.𝟵秒的补篮，为什么他恰好出现在那里？ 马克西突破不中，詹姆斯起跳补篮，将比分改写为96比91。76人最终98比93击败尼克斯。 集锦里，这是一次关键补篮。放进整套进攻中看，更值得关注的是：詹姆斯为什么提前站在那个位置？ 答案藏在“幽灵位”。 ① 从外线发起者，变成篮下终结点 詹姆斯在弧顶持球时，可以与恩比德打挡拆、手递手，让马克西通过无球移动寻找得分机会。 当马克西接管持球，詹姆斯的价值也随之变化：利用体型、阅读防守和篮下终结能力，成为防线身后的威胁。 同一套阵容，可以通过交换角色，改变防守必须处理的问题。 ② “幽灵位”到底是什么？ 通常指篮筐两侧、底线附近的短角落区域，英文常称dunker spot。 站在这里的球员，可以等待内线接球、沿底线切入，也可以直接冲抢前场篮板。 这个位置会让协防者面临两难： 向持球人收缩，身后的进攻球员可能获得机会。 留在原位，又可能让突破者更容易接近篮筐。 对詹姆斯这样的强壮前锋而言，这种站位能把体型和判断力直接转化为篮下威胁。 ③ 西蒙斯的投射，让收官阵容多了一种选择 安芬尼·西蒙斯本场三分8投5中，得到21分。 当他与马克西、VJ·埃奇库姆同时在场，外线投射和突破威胁可以把防守注意力拉向外围，詹姆斯则能更多参与篮下终结。 阵容的关键在于，谁负责持球、谁负责拉开空间，以及谁负责惩罚协防。 ④ 马克西突破时，詹姆斯的进攻已经开始 最后一攻可以从两个层面理解： 第一层，马克西利用速度攻击篮筐，迫使防守收缩。 第二层，詹姆斯在篮下附近准备接球、争抢落点。 当防守注意力被持球人吸引，他便有机会利用位置、体型和起跳时机完成补篮。 这不意味着战术提前设计好了投篮不中。好的进攻站位，会让球队在第一选择失败后，仍然保留第二次机会。 ⑤ 无球时的判断，同样能决定比赛 詹姆斯本场得到𝟮𝟬分、𝟵个篮板、𝟴次助攻。 最后这次补篮，体现的是数据之外的细节：知道什么时候交出球权，什么时候靠近篮筐，什么时候准备争抢落点。 关键球的铺垫，往往从球还没出手时就已经开始。 你更欣赏詹姆斯的哪一种能力：持球组织，还是这种无球时的比赛阅读？ #NBA #LeBronJames #76ers #詹姆斯 #篮球战术

𝗖𝗼𝘂𝗹𝗱 𝘁𝘄𝗼 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 𝗮𝗯𝘀𝗼𝗿𝗯 𝗺𝗼𝘀𝘁 𝘂𝘀𝗮𝗯𝗹𝗲 𝗔𝗜 𝗰𝗼𝗺𝗽𝘂𝘁𝗲 𝗯𝘆 𝟮𝟬𝟮𝟴? OpenAI and Anthropic could benefit from a reinforcing cycle: Stronger models → better compute monetization → greater purchasing power → more research resources → stronger next-generation models. This is an aggressive scenario discussed in August 2026. Whether it materializes depends on model progress, supply expansion, regulation and other constraints. 𝟭. 𝗠𝗼𝗻𝗲𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝘀𝗵𝗮𝗽𝗲𝘀 𝗯𝘂𝘆𝗶𝗻𝗴 𝗽𝗼𝘄𝗲𝗿 Identical hardware can generate very different revenue depending on the models, customers and tasks it serves. A company solving more valuable problems may be able to pay more for compute. Resources that look too expensive to one buyer can remain attractive to another. Commercial performance therefore influences who can secure additional capacity. 𝟮. 𝗪𝗮𝘁𝗰𝗵 𝗻𝗲𝘄 𝗰𝗮𝗽𝗮𝗰𝗶𝘁𝘆 If frontier labs’ demand keeps growing faster than overall supply, they could absorb an increasing share of new deployments. Newer chips also offer better performance and efficiency, potentially amplifying the advantage of securing new hardware. Gigawatts measure power, however. Comparing computing capacity also requires accounting for hardware generations, utilization and workloads. 𝟯. 𝗖𝗼𝗻𝗰𝗲𝗻𝘁𝗿𝗮𝘁𝗲𝗱 𝗱𝗲𝗺𝗮𝗻𝗱 𝗶𝘀 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗳𝗿𝗼𝗺 𝗼𝘄𝗻𝗲𝗿𝘀𝗵𝗶𝗽 Cloud providers and infrastructure operators may continue to own the facilities. Yet capacity primarily serving a few companies’ models can still concentrate demand and influence resource allocation. The useful questions are who uses the capacity, who can pay the most, and where suppliers allocate their next expansion. 𝟰. 𝗧𝗵𝗲 𝘀𝘂𝗽𝗽𝗹𝘆 𝗰𝗵𝗮𝗶𝗻 𝘄𝗶𝗹𝗹 𝘀𝗲𝗲𝗸 𝗶𝘁𝘀 𝘀𝗵𝗮𝗿𝗲 Stronger monetization does not automatically translate into equivalent profits. Chipmakers, memory suppliers, infrastructure operators and power providers can seek higher returns. Expanding supply requires time and capital. Revenue must also cover research, staffing and other expenses beyond compute. 𝟱. 𝗧𝗵𝗲 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗳𝗲𝗲𝗱𝗯𝗮𝗰𝗸 𝗹𝗼𝗼𝗽 If better models accelerate coding, experiments and research, they could help produce the next generation faster. Leaders might then combine greater revenue, more compute and stronger research tools. That outcome remains conditional. Competing models, open source, expanding supply and regulation could all change the trajectory. The central question is how markets rebalance when the companies best able to buy compute are also best positioned to use it to extend their lead. Over the next few years, will compute become more concentrated—or will open source and supply expansion strengthen competition? #OpenAI #Anthropic #AI #Compute #AIInfrastructure

到𝟮𝟬𝟮𝟴年，大部分可用AI算力会服务于两家公司吗？ OpenAI和Anthropic可能形成的优势，是一种不断强化的循环： 更强的模型 → 更高的算力变现能力 → 更强的采购能力 → 更多研发资源 → 下一代模型。 这是一种基于2026年8月讨论的激进情景推演，能否兑现，取决于模型进步、供给扩张和监管等条件。 ① 谁更能把算力变成收入，谁就更有能力抢到算力 同样一批芯片，运行不同模型、服务不同客户，产生的收入可能相差很大。 如果一家公司的模型能解决价值更高的任务，它就可能承受更高的算力价格。 对其他买家而言过于昂贵的资源，对它而言仍然值得购买。 这意味着，算力竞争同时是一场商业化能力的竞争。 ② 新增算力的分配，可能比现有份额更值得关注 如果前沿实验室的需求增长持续快于整体供给，它们吸收新增资源的比例就可能不断提高。 新一代芯片通常又有更高的性能和效率。因此，拿到更多新设备，可能带来超过用电规模增长的有效算力提升。 不过，GW是功率单位，不是计算能力单位。比较算力，还需要看芯片代际、利用率和具体任务。 ③ 算力集中，不等于数据中心全部归它们所有 大量基础设施仍可能由云厂商和其他运营商持有。 但如果这些设备主要用于训练或运行少数公司的模型，模型层的需求集中仍然可能影响资源分配。 值得观察的是：谁在使用这些算力、谁能为它支付最高价格，以及供应商愿意把新增容量卖给谁。 ④ 高回报会吸引竞争，也会让供应链争取更多收益 即使前沿实验室暂时拥有更强的变现能力，这种优势也不会自动全部变成利润。 芯片、内存、服务器、数据中心和电力供应商，都可能要求分享更多回报。扩产还需要时间和资金。 收入高，不代表净利润高；收入与算力成本的差额，还要覆盖研发、人员和其他支出。 ⑤ 最值得关注的是，商业优势能否变成持续的研发优势 如果更好的模型能帮助完成编程、实验和研究，它就可能加快下一代模型的开发。 一旦这条路径成立，领先企业便可能同时拥有更多收入、更多算力和更强的研发工具。 但它仍然存在边界：竞争模型可能追上，开源可能降低成本，供给可能扩大，监管也可能改变部署节奏。 所以，“两家公司掌握大部分可用AI算力”应当被视为需要检验的情景。 它提出的核心问题是：当最有能力购买算力的企业，也最有能力利用算力继续扩大优势，市场会如何重新达到平衡？ 未来几年，你更看好算力继续集中，还是开源与供给扩张推动竞争？ #OpenAI #Anthropic #AI #算力 #人工智能

🧠 𝗘𝗶𝗴𝗵𝘁 𝗔𝗜 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻𝘀 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗻𝗲𝘅𝘁 𝗱𝗲𝗰𝗮𝗱𝗲: 𝘆𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗼𝗿 𝗰𝗼𝘂𝗹𝗱 𝗯𝗲 𝗮 𝗱𝗶𝗴𝗶𝘁𝗮𝗹 “𝘆𝗼𝘂” AI’s next phase could reshape personal assistance, business processes and physical work. These eight scenarios use 2026 as their starting point. Their timelines remain uncertain, but the directions are worth watching. 𝟭. 𝗔 𝗽𝗲𝗿𝘀𝗼𝗻𝗮𝗹 “𝗲𝘅𝗼𝘀𝗲𝗹𝗳” 𝘄𝗶𝘁𝗵𝗶𝗻 𝘁𝗵𝗿𝗲𝗲 𝘆𝗲𝗮𝗿𝘀 With permission, a personal agent could learn from your messages, schedule, conversations and preferences. Over time, it could retain context and understand your decision-making habits. Smart glasses may become an important interface, allowing it to see what you see and assist when needed. 𝟮. 𝗔𝗜 𝗮𝘀 𝗮 𝗰𝗼𝗺𝗺𝗼𝗻 𝗵𝗲𝗮𝗹𝘁𝗵𝗰𝗮𝗿𝗲 𝗲𝗻𝘁𝗿𝘆 𝗽𝗼𝗶𝗻𝘁 𝘄𝗶𝘁𝗵𝗶𝗻 𝘁𝘄𝗼 𝘆𝗲𝗮𝗿𝘀 More people could use AI to organize symptoms, understand test reports and prepare questions before seeing a clinician. How diagnosis, treatment and responsibility are divided will still depend on reliability, clinical validation and regulation. 𝟯. 𝗠𝗼𝘀𝘁 𝗔𝗜 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻 𝗰𝗼𝘂𝗹𝗱 𝗵𝗮𝗽𝗽𝗲𝗻 𝗯𝗮𝗰𝗸𝘀𝘁𝗮𝗴𝗲 Within roughly five years, you might interact with one or two agents that coordinate others handling research, coding, purchasing and workflows. The prediction that “99% of AI will not interact directly with humans” points toward a world where many customers of AI services are other AIs. Agent transactions could also create new economic activity. Transaction volume, however, cannot be directly equated with economic value added. 𝟰. 𝗪𝗶𝗱𝗲𝘀𝗽𝗿𝗲𝗮𝗱 𝗵𝘂𝗺𝗮𝗻𝗼𝗶𝗱 𝗿𝗼𝗯𝗼𝘁𝘀 𝗺𝗮𝘆 𝘀𝘁𝗶𝗹𝗹 𝗯𝗲 𝗮 𝗱𝗲𝗰𝗮𝗱𝗲 𝗮𝘄𝗮𝘆 A successful demonstration leaves many challenges unresolved: unpredictable environments, safety, maintenance and sustained operation. Reliability could determine how quickly humanoid robots become practical in homes and workplaces. 𝟱. 𝗘𝘅𝗼𝘀𝗸𝗲𝗹𝗲𝘁𝗼𝗻𝘀 𝗰𝗼𝘂𝗹𝗱 𝗲𝘅𝗽𝗮𝗻𝗱 𝘄𝗶𝘁𝗵𝗶𝗻 𝗳𝗶𝘃𝗲 𝘆𝗲𝗮𝗿𝘀 Rehabilitation, industrial assistance and recreation could become major applications. These devices could support mobility, reduce physical strain and enable new forms of sport by augmenting human movement. 𝟲. 𝗖𝗵𝗶𝗻𝗮 𝗰𝗼𝘂𝗹𝗱 𝗽𝗿𝗼𝗱𝘂𝗰𝗲 𝘄𝗼𝗿𝗹𝗱-𝗹𝗲𝗮𝗱𝗶𝗻𝗴 𝗔𝗜 𝗰𝗵𝗶𝗽𝘀 A roughly four-year timeline makes this a particularly bold forecast. “Leading” needs a clear definition: training performance, inference efficiency, energy consumption, cost, manufacturing and software support may produce different answers. 𝟳. 𝗦𝗺𝗮𝗿𝘁 𝗴𝗹𝗮𝘀𝘀𝗲𝘀 𝗰𝗼𝘂𝗹𝗱 𝘀𝗰𝗮𝗹𝗲 𝗳𝗶𝗿𝘀𝘁 𝗶𝗻 𝗳𝗶𝗲𝗹𝗱 𝘄𝗼𝗿𝗸 Technicians, construction workers and logistics staff need information while keeping their hands available. Visual instructions, navigation and remote assistance offer concrete applications. Measurable time savings and fewer errors could drive adoption. 𝟴. 𝗔𝗜 𝗰𝗼𝘂𝗹𝗱 𝗯𝗲𝗰𝗼𝗺𝗲 𝗮 𝘁𝗶𝗲𝗿𝗲𝗱 𝗺𝗮𝗿𝗸𝗲𝘁 Some capabilities of Chinese and American models may converge, increasing the importance of cost, efficiency and ecosystems. Simple tasks favor low prices. Complex tasks demand stronger reasoning. Real-time applications prioritize latency. Buying AI services could increasingly resemble choosing computing resources for a specific job. Together, these scenarios describe AI taking a deeper role in coordinating work and acting in the physical world. Which would change your daily life or work first? #AI #AIAgents #Robotics #SmartGlasses #FutureOfWork

🧠 未来𝟭𝟬年，AI最值得关注的𝟴个变化：你的下一位搭档，可能是自己的“分身” AI下一阶段的发展，可能同时改变个人生活、企业流程和现实世界中的工作方式。 以下八种未来情景，以2026年为起点。时间表仍待验证，但值得提前关注。 ① 𝟯年左右，个人AI可能成为“外部自我” 在获得授权后，它持续了解你的邮件、日程、沟通记录和偏好，逐渐成为长期协作的个人智能体。 它既帮助完成任务，也保留上下文，理解你做决定的习惯。智能眼镜可能成为重要交互入口，让它看到你看到的东西，在需要时提供帮助。 ② 𝟮年左右，AI可能成为常见的医疗咨询入口 更多人可能先通过AI整理症状、理解检查报告、准备就诊问题，再与医生沟通。 这意味着医疗服务入口可能改变。至于诊断、治疗和责任如何分配，仍取决于系统可靠性、临床验证和监管安排，不能简单归结为“AI诊断，人类安慰”。 ③ 𝟱年左右，大部分AI协作可能发生在后台 你只与一两个个人智能体交流，它们再调用其他智能体，完成检索、编程、采购、谈判和流程处理。 “99%的AI不直接接触人类”是一项大胆预测，指向的是一个变化：未来大量AI服务的客户，可能也是AI。 智能体之间的交易也可能形成新的经济活动。但交易流水与经济增加值是不同指标，不能直接据此判断其是否超过人类经济。 ④ 人形机器人的普及，可能还需要𝟭𝟬年 演示中完成一个动作，与长期、安全、稳定地工作，存在很大距离。 家庭和工厂需要机器人处理复杂环境、意外情况和持续运行。可靠性将是决定普及速度的重要门槛。 ⑤ 外骨骼可能在𝟱年内更广泛落地 医疗康复、工业搬运和运动娱乐，可能成为三个重要方向。 外骨骼通过增强人的行动能力创造价值：帮助部分行动受限者活动、减轻体力劳动负担，也可能带来新的运动形式。 ⑥ 全球领先的AI芯片，可能出现更多中国参与者 约四年后，中国可能制造出全球领先的AI芯片，是这组预测中最值得持续检验的一项。 “领先”需要具体衡量：训练性能、推理效率、能耗、成本、制造能力和软件生态，都可能影响结果。 ⑦ 智能眼镜的大规模应用，可能先从蓝领工作开始 维修、物流、建筑和现场服务人员，经常需要一边操作，一边查看信息。 图纸提示、导航和远程指导，都有明确的使用场景。能否在工作中节省时间、减少错误，可能成为推动采用的重要因素。 ⑧ AI市场可能走向分层供应与生态竞争 中美模型的部分能力可能逐渐接近，价格、效率和应用生态的重要性随之上升。 不同任务需要不同等级的模型与算力：简单任务看成本，复杂任务看能力，实时任务看延迟。 未来购买AI服务，可能越来越像按用途选择计算资源。 这八种情景共同指向一个趋势：AI将更深入地参与任务分配、工作协作和现实行动。 其中，哪一项最可能先改变你的生活或工作？ #AI #人工智能 #AIAgents #机器人 #智能眼镜

𝗚𝗼𝗼𝗴𝗹𝗲’𝘀 𝟭𝟮.𝟲𝟰% 𝗱𝗿𝗼𝗽: 𝗔𝗜’𝘀 𝗯𝗶𝗹𝗹 𝗶𝘀 𝗰𝗮𝘁𝗰𝗵𝗶𝗻𝗴 𝘂𝗽 𝘄𝗶𝘁𝗵 𝗶𝘁𝘀 𝗴𝗿𝗼𝘄𝘁𝗵 𝘀𝘁𝗼𝗿𝘆 Alphabet’s Class A shares, GOOGL, fell 12.64% on October 29, erasing approximately $482 billion in market value. The central tension in its earnings report: revenue is growing rapidly, while free cash flow remains negative. Google faces two connected questions: how much revenue can AI generate, and how expensive will that growth be? 𝟭. 𝗧𝗵𝗲 𝗰𝗮𝘀𝗵-𝗳𝗹𝗼𝘄 𝗴𝗮𝗽 Alphabet generated approximately $42.29 billion in operating cash flow during Q3. Capital expenditures reached $51.36 billion, leaving free cash flow at roughly negative $9.07 billion. Full-year capital spending guidance remains $195–205 billion, with another significant increase expected in 2027. Negative free cash flow does not mean the underlying business is unprofitable. Google continues to generate substantial operating cash, but infrastructure investment exceeded that cash generation. The quarter also included a $5.2 billion European Commission fine payment and a TPU inventory build. Attributing the entire shortfall to AI model spending would miss important context. The question is when those investments translate into durable cash returns. 𝟮. 𝗔𝗜 𝘂𝘀𝗮𝗴𝗲 𝗮𝗻𝗱 𝗮𝗱 𝗿𝗲𝘃𝗲𝗻𝘂𝗲 𝗰𝗮𝗻 𝗴𝗿𝗼𝘄 𝗮𝘁 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝘀𝗽𝗲𝗲𝗱𝘀 Search and other revenue increased 12%. Management acknowledged that lower monetization in AI Mode affected Search revenue growth. Currency movements and changes in advertiser demand also contributed. That highlights a difficult transition: greater adoption of AI answers does not automatically generate the same advertising revenue per interaction as traditional search. Google must improve the user experience while proving that new advertising formats can support it economically. 𝟯. 𝗖𝗹𝗼𝘂𝗱 𝗴𝗿𝗼𝘄𝘁𝗵 𝗶𝘀 𝘀𝘁𝗿𝗼𝗻𝗴. 𝗖𝗮𝘀𝗵 𝗿𝗲𝘁𝘂𝗿𝗻𝘀 𝘀𝘁𝗶𝗹𝗹 𝗺𝗮𝘁𝘁𝗲𝗿. Google Cloud revenue rose 103% to approximately $30.7 billion. Across Alphabet, revenue grew 25% and operating income increased 30%. Those results show that AI investment is generating business revenue. Declaring Google’s competitive position broken would be premature. The next test is whether that expansion can deliver sustained cash returns after infrastructure spending, depreciation and operating costs. This is an expensive business-model transition. Search needs stronger monetization of AI interactions. Cloud needs to convert demand into cash. Capital spending needs to demonstrate attractive long-term returns. For investors following AI stocks, revenue growth, capital expenditure and free cash flow belong in the same analysis. Which would do more to rebuild confidence next quarter: stronger Search monetization or a recovery in free cash flow? #Google #Alphabet #AI #Earnings #Stocks

谷歌单日跌去𝟭𝟮.𝟲𝟰%：AI的账单，正在追上增长的故事 10月29日，Alphabet A类股票GOOGL下跌12.64%，市值损失约4,820亿美元。 这份财报最值得关注的矛盾是：收入仍在快速增长，自由现金流却继续为负。谷歌必须同时回答两个问题——AI能创造多少新收入，以及这些收入需要付出多大的代价。 ① 现金流：赚钱的速度，暂时跟不上投入的速度 第三季度，Alphabet经营活动产生约422.9亿美元现金，但资本支出达到约513.6亿美元，自由现金流因此降至约负𝟵𝟬.𝟳亿美元。 全年资本支出指引仍为𝟭,𝟵𝟱𝟬亿至𝟮,𝟬𝟱𝟬亿美元，管理层预计2027年还将显著增加。 这里需要区分：自由现金流为负，不等于主营业务亏损。谷歌仍在产生大量经营现金，只是基础设施投入超过了当季现金流入。 而且，当季还支付了52亿美元欧盟罚款，并增加了TPU系统库存。把现金流缺口全部归因于AI模型烧钱，会遗漏重要背景。 真正需要持续追踪的是：这些支出何时转化为收入、利润和现金回报。 ② 搜索：AI使用量增加，广告收入未必同步增加 第三季度，搜索及其他收入同比增长12%。管理层承认，AI Mode目前较低的变现水平影响了搜索收入增速，但汇率和广告行业需求变化也有影响。 这暴露了搜索转型中的关键难题：用户愿意使用AI回答，并不意味着每次使用都能带来与传统搜索相同的广告收入。 谷歌需要一边改善AI体验，一边验证新的广告形式。产品采用速度与商业化速度之间的差距，将直接影响转型成本。 ③ 云业务：增长很强，回报仍需兑现 Google Cloud第三季度收入同比增长𝟭𝟬𝟯%，达到约307亿美元。Alphabet整体营收增长25%，营业利润增长30%。 这些数字说明，谷歌的AI投入已经产生业务收入，也让“谷歌已经失去竞争力”的结论显得过早。 但高增长仍需接受现金回报的检验。云业务扩张能否覆盖持续增加的基础设施、折旧和运营成本，是接下来财报的重要看点。 这次财报呈现的是一场昂贵的商业模式转型： 搜索需要提高AI流量的变现效率。 云业务需要把强劲需求转化为持续现金回报。 资本支出需要证明长期收益能够覆盖投入。 对于关注AI科技股的投资者，营收增速、资本支出和自由现金流，需要放在一起看。 下一份财报，哪个信号更能改善市场信心：搜索广告变现提速，还是自由现金流回升？ #Google #Alphabet #AI #美股 #财报

AI Agents: The Next Technology Revolution Isn't Just About Smarter AI. It's About AI Doing the Work. For the past few years, the world has been asking how intelligent artificial intelligence can become. From ChatGPT to AI-generated videos, images, software, and research tools, the progress has been remarkable. But we are still in the early stages of a much bigger transformation. The most important shift may not be from one powerful AI model to an even more powerful model. It may be the transition from AI that answers questions to AI that plans, acts, coordinates tools, remembers context, and completes tasks. That is the promise of AI agents. And if that promise becomes reality at scale, it could fundamentally reshape the internet, business models, the structure of organizations, and the future of human work. 01 | The Fundamental Difference Between ChatGPT and AI Agents Many people think an AI agent is simply a more advanced version of ChatGPT. But that interpretation misses the larger opportunity. Traditional AI assistants primarily provide information. You ask a question, and they produce an answer. You request code, and they generate code. You ask for a financial analysis, and they produce a report. But humans usually remain responsible for executing the work. AI agents attempt to change that relationship. Imagine planning a vacation. A traditional AI assistant can recommend flights, compare hotels, and suggest an itinerary. A mature AI agent could potentially compare prices, check your calendar, communicate with service providers, make authorized reservations, and adjust your plans when circumstances change. Instead of manually directing every step, you define the objective. The agent determines how to accomplish it. This represents a significant change in how humans interact with software. In the traditional model, humans operate applications to complete tasks. In the emerging model, AI may operate applications while humans define objectives, establish boundaries, and approve important decisions. We are potentially moving from software that provides answers to systems that execute work. 02 | The Real Competition Is No Longer Just About AI Models An advanced AI agent does not necessarily need one enormous model to solve every problem. It needs an intelligent system capable of coordinating models, tools, memory, and workflows. Three components are particularly important. First, model orchestration. Simple information-processing tasks may be handled by smaller, cheaper, faster models. Complex reasoning, advanced programming, and scientific analysis may require more capable models. Image recognition, speech processing, and video generation may be delegated to specialized systems. The most effective agent should understand not only how to solve a problem, but also which resources are appropriate for the task. Intelligence matters, but so do cost, efficiency, and reliability. Second, long-term memory. Many current AI applications still require users to repeatedly explain their background, preferences, and previous work. A mature personal agent should gradually develop useful context about its owner's projects, objectives, working habits, and decisions. Consider a scientist working on a long-term research project. An effective agent should understand previous analyses, the models already tested, unresolved questions, and the next experiments required. Its value is not simply remembering conversations. Its value is converting historical information into useful context for future decisions. Third, connectivity. Through APIs, MCP, and agent communication protocols, AI systems can interact with external applications, databases, business software, and other agents. This is what allows AI to move beyond the chat window. But connecting to more tools creates additional responsibilities. Reliable execution, permission management, and error recovery become essential. 03 | One Person May Eventually Manage an Entire Team of AI Agents This may become one of the most economically significant applications of agent technology. Historically, running a business required people with different expertise. Someone handled marketing. Someone managed financial records. Someone communicated with customers. Someone developed software. Others managed sales, operations, and content production. In the future, many standardized digital tasks could be handled by specialized agents. One agent researches the market. Another prepares financial reports. Another helps manage customer communications. Another creates and organizes content. A supervising agent checks progress and identifies potential errors. Meanwhile, the human owner sets objectives, manages resources, defines rules, and evaluates the outcomes. The implication is significant. An individual's productive capacity may become less dependent on the number of hours they personally work. It may increasingly depend on how effectively they can organize and manage digital labor. A person still has only 24 hours per day. But multiple authorized AI systems could potentially perform different tasks simultaneously. This could change the economics of entrepreneurship and small businesses. However, deploying ten agents does not automatically create a productive organization. Poor coordination, inaccurate outputs, and duplicated work can make an agent-based system expensive and unreliable. The real competitive advantage is not the number of agents. It is their ability to consistently produce valuable results. 04 | AI Agents Could Transform the Economics of the Internet Much of today's internet economy is built around human attention. People search for information. They visit websites. They watch product reviews. They compare prices. They browse social media and eventually make purchasing decisions. Search engines, e-commerce platforms, social networks, and advertising companies compete to capture this attention. But what happens if AI agents begin performing much of that activity on behalf of humans? Imagine buying a computer for scientific research. Today, you might watch several YouTube reviews, compare processors and graphics cards, examine benchmark results, and visit multiple retailers. In the future, you might simply tell an agent your budget, research requirements, and performance expectations. The agent could analyze options, compare prices, evaluate support policies, and recommend suitable products. With appropriate authorization, it might even complete the purchase. This introduces a new dimension of competition. Businesses may need to compete not only for human attention, but also for selection by AI agents. Accurate product information, transparent pricing, machine-readable data, accessible APIs, and verifiable reputation could become increasingly important. This could also transform advertising. Traditional advertising often seeks to influence human emotion, attention, and purchasing behavior. Agent-driven purchasing may place greater emphasis on structured comparisons of quality, price, performance, and reliability. However, that does not mean advertising will disappear. Agents may still be influenced by commercial partnerships, platform incentives, biased information, and ranking systems. A critical question therefore emerges: Who controls the recommendation logic of AI agents? Because the organizations that influence how agents select products and services may control important gateways to the next generation of digital commerce. 05 | Where Are the Real Money-Making Opportunities? When people hear about AI agents, their first instinct is often to build an application and charge a monthly subscription. But the larger opportunity may lie in selling completed work rather than selling the software itself. Historically, software companies sold tools. Customers purchased those tools and then hired or trained people to operate them. AI agents could introduce a different business model. Instead of selling software that helps customers perform tasks, companies could increasingly sell the completed results. A business may not need another customer relationship management application. It may prefer a service that actually performs customer follow-ups. A small company may not need complicated financial reporting software. It may prefer an automated system that prepares reports, identifies unusual transactions, and explains changes in business performance. A content creator may not want separate subscriptions for editing, translation, subtitles, and formatting. They may prefer a system that handles the entire workflow. This represents a shift from selling tools to selling outcomes. Customers ultimately do not care how many tokens an AI system consumes. They care whether the system solves a problem, reduces costs, generates revenue, or improves productivity. Consider invoice processing. If a company currently spends substantial employee time reviewing and organizing invoices, an AI agent that performs the work accurately and produces auditable records could create measurable economic value. But if employees must repeatedly correct the agent's mistakes, the actual savings may be limited. For that reason, an agent-based business should be evaluated using three questions. Can the system reliably complete a real task? Can it measurably reduce costs or generate additional revenue? Will customers continue paying for the result? Specialized agents that solve expensive, well-defined business problems may offer stronger opportunities than generic agents that promise to do everything. 06 | The Biggest Barrier May Not Be Intelligence. It May Be Trust. There is an important distinction between an AI system being capable of performing an action and people being willing to authorize that action. You may trust AI to draft an email. But would you allow it to send an important message to a client without your approval? You may trust AI to analyze your investment portfolio. But would you allow it to execute trades independently? You may want AI to organize your work. But would you give it unrestricted access to private documents, bank accounts, and corporate databases? This is where agent technology faces one of its greatest challenges. A chatbot generating an incorrect answer may waste a few minutes. An agent with permission to modify databases, transfer money, or operate business infrastructure could create much more serious consequences. That means security and reliability must be fundamental parts of agent design. Important safeguards include permission boundaries, transaction logs, human approval for sensitive actions, anomaly detection, and recovery mechanisms. The best agent is not necessarily the one that requires no human involvement. It is the one that knows when independent execution is appropriate and when human confirmation is essential. Trust will not simply be a regulatory requirement. It may become one of the most important competitive advantages in the agent economy. 07 | If AI Can Perform More Work, What Makes Humans Valuable? This is perhaps the most important question of all. The Industrial Revolution expanded human physical capabilities through machines. The computer revolution expanded our ability to process information. AI agents could expand our ability to organize and execute complex work. Historically, individuals needed to perform many operational steps themselves. In the future, they may increasingly delegate those steps to AI systems. Does that mean human judgment becomes less important? The opposite may happen. When execution becomes easier, deciding what should be executed becomes more important. When information becomes abundant, identifying reliable information becomes more valuable. When producing articles, images, videos, and software becomes cheaper, originality and meaningful insight may become increasingly important. When AI generates multiple solutions, humans still need to determine which solutions are worth pursuing and which risks are acceptable. This could change how society evaluates skills. Traditionally, expertise in a specific technical process could provide a strong competitive advantage. As AI tools make more technical capabilities accessible, the ability to define problems, evaluate results, coordinate resources, and take responsibility for decisions may become increasingly valuable. But this transition will not necessarily benefit everyone equally. Some standardized tasks may be automated. Certain jobs may shrink or change significantly. New roles and businesses may emerge, but their growth may not fully offset displacement for every worker or community. People and organizations that control technology, capital, data, and customer relationships may capture a disproportionate share of the early economic benefits. The rise of AI agents is not only a technological transformation. It could also become a major restructuring of labor markets and economic power. 08 | The Next Internet May Be an Internet of Agents Consider how the internet has evolved. The early web allowed people to access information through websites. Social media allowed individuals to create and distribute content. Mobile computing made digital services accessible almost everywhere. AI agents may introduce another fundamental change. In the future, people may have one or several persistent digital agents. These systems could organize schedules, monitor information, manage documents, coordinate workflows, and communicate with external services. Businesses may operate their own agents as well. A consumer's agent searches for a product. A merchant's agent answers detailed product questions. A logistics agent coordinates delivery. A financial agent processes an authorized payment. Some commercial transactions could eventually involve agents interacting directly with other agents. In that environment, a growing share of internet activity might no longer involve humans manually clicking through websites. Instead, digital services may communicate through structured agent-to-agent interactions. This creates exciting possibilities for efficiency. But it also raises difficult questions. How do we distinguish humans from autonomous agents? How can an agent prove that it has permission to act for a person? Who is responsible when agents make incorrect transactions? How should disputes between autonomous systems be resolved? And if large amounts of online content are automatically generated, how can users and machines verify what is trustworthy? The answers may determine how far the agent-based internet can develop. 09 | The Bigger Picture: AI Agents Could Redefine the Relationship Between Work and Value Not every application will disappear. AI will not necessarily eliminate every human job. But the value of software may gradually shift from providing interfaces for humans to providing capabilities that intelligent systems can access and operate. Historically, humans learned how to use software. In the future, software may increasingly be designed to work with AI agents acting on behalf of humans. Historically, businesses expanded by hiring employees, purchasing software, and developing operational processes. In the future, companies may manage a combination of human employees and digital agents. Historically, personal productivity was heavily influenced by the number of skills someone possessed and the hours they could work. In the future, it may increasingly depend on the quality of their objectives, the systems they organize, and the decisions they make. AI is reducing the cost of execution. But that does not automatically reduce the value of judgment. The most exciting possibility is not a future in which humans no longer need to do anything. It is a future in which we can delegate more repetitive work and spend more time on creativity, scientific discovery, strategic decisions, and meaningful personal pursuits. Whether that future improves life for ordinary people will depend on reliability, affordability, access, institutions, and how the economic benefits are distributed. The most valuable skill of the future may not be how much work you can personally complete. It may be your ability to identify meaningful problems, make sound decisions, and turn ideas into real-world outcomes through AI. Ultimately, AI agents may transform much more than the way we use the internet. They may redefine how much value an individual, a company, or an entire society can create. #AI #AIAgents #ArtificialIntelligence #Automation #AIEntrepreneurship #FutureOfWork

AI Agent：下一场真正的科技革命，不是让AI更聪明，而是让AI替人类完成工作 过去几年，我们一直在讨论人工智能究竟能够发展到什么程度。 从ChatGPT到各种AI视频、图片和编程工具，AI已经证明自己能够处理越来越复杂的任务。 但这些还只是人工智能发展的第一阶段。 真正值得关注的变化，是AI正在从一个回答问题的工具，逐渐发展为能够自主规划、调用工具、执行任务，甚至长期积累经验的智能体（AI Agent）。 这可能彻底改变未来的互联网、企业组织结构、商业模式，以及普通人的工作方式。 01｜ChatGPT与AI Agent，究竟有什么本质区别？ 很多人认为，AI Agent只不过是一个升级版的ChatGPT。 但两者最大的差异，并不只是模型有多聪明，而是它们与现实世界互动的方式。 传统AI主要负责提供信息。 你提出问题，它生成答案。 你要求编写代码，它提供程序。 你让它分析一家公司，它生成财务分析报告。 但真正执行任务的人，往往仍然是你自己。 AI Agent则试图改变这种关系。 例如，你准备安排一次旅行。 传统AI可以告诉你哪些航班便宜、酒店评价如何、旅游路线怎么安排。 但一个成熟的AI Agent，未来可能在获得授权后，自主比较机票价格、查询酒店、规划行程、完成预订，并根据航班变化重新调整计划。 你不再需要一步步指导它操作。 你只需要告诉它最终目标。 这意味着，软件的使用方式可能发生根本变化。 过去是人类操作软件完成任务。 未来则可能是AI操作软件，人类负责制定目标、监督过程和审核结果。 这是从提供答案到执行工作的转变。 02｜AI Agent真正的竞争，不只是大模型的竞争 未来一个优秀的AI Agent，可能并不需要依赖单一的大模型完成所有工作。 它需要的是一个能够协调不同模型、工具、记忆和执行流程的智能系统。 这里面至少有三个关键技术方向。 第一个，是模型调度能力。 简单的信息整理，可以使用速度更快、价格更低的小模型。 复杂的数学推理、科研分析或者编程任务，可以交给推理能力更强的模型。 图片识别、语音处理和视频生成，则可以交给专门的模型。 真正优秀的Agent，应该知道什么时候调用什么模型，以及如何用合理的成本完成任务。 第二个，是长期记忆。 今天很多AI工具的问题在于，每次开始新任务，用户仍然需要重复解释背景。 但是，一个真正成熟的个人Agent，应该逐渐了解你的工作习惯、长期目标、知识背景以及决策偏好。 例如，一名科研人员使用Agent处理长期项目。 Agent不仅需要理解当前的数据，还应该知道此前使用过哪些模型、完成过哪些分析、发现了什么问题，以及下一步应该如何验证。 它的价值，不是简单地保存聊天记录，而是把历史信息转化为可以持续使用的知识。 第三个，是与现实数字系统连接的能力。 通过API、MCP以及不同Agent之间的通信协议，AI可以调用外部数据库、办公软件、搜索工具和业务系统。 当这些能力结合起来，AI就不再只是一个聊天窗口。 它开始具备跨软件完成复杂任务的潜力。 不过，能够访问工具，并不等于能够可靠地完成任务。权限管理、执行准确性和错误恢复，同样重要。 03｜未来一个人，可能能够管理一支AI团队 这是AI Agent最具商业价值的方向之一。 过去，一个人想经营一家企业，往往需要不同领域的专业人员。 有人负责市场研究。 有人负责客户服务。 有人负责财务。 有人负责销售。 有人负责内容生产。 还有人负责技术开发。 未来，这些工作中大量标准化、数字化的任务，可能由不同的AI Agent分工完成。 一个Agent负责市场调研。 一个Agent负责财务整理。 一个Agent负责客户沟通。 一个Agent负责制作内容。 另一个Agent负责检查进度和发现错误。 而你负责明确目标、制定规则、决定预算，以及检查最终成果。 个人的生产能力，可能不再完全取决于自己每天能够投入多少时间，而是取决于能够组织和管理多少数字化劳动力。 过去，一个人只有24小时。 未来，虽然你的时间仍然有限，但你可以让多个经过授权的系统同时处理不同任务。 这将改变个人创业和小型企业的成本结构。 不过，一个人拥有多个Agent，并不意味着自动拥有一家高效企业。 如果多个Agent不断产生错误、重复工作或者缺乏协调，反而可能增加运营成本。 真正的价值，不在于拥有多少Agent，而在于它们能否稳定交付成果。 04｜AI Agent可能改变整个互联网的商业模式 今天互联网的大量商业活动，都建立在人类主动浏览信息的基础上。 我们使用搜索引擎寻找答案。 使用电商网站比较价格。 观看YouTube评测视频。 阅读论坛和社交媒体上的产品推荐。 最终由自己作出购买决定。 这套商业体系的核心，是争夺人类的注意力。 但是，如果未来越来越多的信息搜集和消费决策由AI Agent完成，互联网会发生什么？ 假设你想购买一台用于科研分析的电脑。 今天，你可能需要观看多个评测视频，访问不同的购物网站，研究CPU、GPU、内存和价格。 未来，你只需要告诉Agent你的预算、使用需求和性能要求。 它就可以帮助你筛选配置、比较价格、检查售后服务，并提出购买建议。 在获得授权的情况下，它甚至可以直接完成购买。 这样一来，互联网可能出现一种全新的竞争方式。 企业除了争夺消费者的注意力，还要争夺被AI Agent发现、理解和推荐的机会。 未来产品能否被机器准确识别，数据是否可信，价格是否透明，API是否容易访问，可能直接影响企业获得客户的能力。 这也会影响广告行业。 过去，广告主要试图影响人的情绪、兴趣和购买冲动。 未来，部分商业推荐可能更多依赖Agent对价格、质量、性能和信誉的比较。 但这不代表广告会消失。 Agent同样可能受到平台规则、商业合作、训练数据偏差和推荐算法的影响。 因此，未来一个非常重要的问题是： 究竟谁在控制AI Agent的推荐逻辑？ 因为控制推荐逻辑的人，可能掌握下一代互联网的重要商业入口。 05｜AI Agent最值得关注的赚钱机会是什么？ 很多人看到AI Agent，第一反应是开发一个工具，然后向用户收取订阅费。 但未来更大的商业机会，可能不是出售AI软件本身，而是出售AI完成的工作成果。 过去，软件公司销售工具。 客户购买软件以后，还需要学习如何操作。 未来，一部分AI公司可能直接出售工作结果。 例如： 企业不一定需要购买新的客户管理软件，而是购买能够持续完成客户跟进的自动化服务。 一家小公司不一定需要雇佣专门人员制作每周财务报告，而是购买能够自动整理数据、分析异常并生成报告的服务。 内容创作者不一定需要分别购买视频剪辑、字幕、翻译和排版工具，而是购买能够完成整个内容制作流程的系统。 这里的商业逻辑正在改变。 客户真正愿意支付的，不是AI调用了多少次模型，也不是使用了多少Token，而是最终解决了多少问题、创造了多少价值。 假设一家企业过去需要员工花费大量时间处理发票。 如果AI Agent能够在准确、可审计的前提下完成大部分工作，同时显著降低总成本，那么这种服务就可能具有真实商业价值。 但如果AI完成任务后，仍然需要员工反复检查和修改，实际节省的成本可能并不明显。 所以，评价一个Agent创业项目，应该关注三个问题。 它是否能够稳定完成真实工作？ 它是否能够为客户节省成本或者增加收入？ 客户是否愿意长期为这种结果付费？ 相比什么都能做的通用Agent，那些真正解决具体行业问题的专业Agent，可能拥有更清晰的商业价值。 06｜未来AI Agent最大的障碍，可能不是智能，而是信任 AI能否完成一项任务，与人类是否愿意授权它完成这项任务，是完全不同的两个问题。 你可能愿意让AI帮你写一封邮件。 但是，你愿意让它不经确认就把邮件发送给重要客户吗？ 你可能愿意让AI分析自己的股票账户。 但是，你愿意让它自主决定买卖吗？ 你愿意让AI帮你管理工作。 但是，你愿意让它访问所有私人文件、银行信息和公司数据库吗？ 这就是Agent必须解决的信任问题。 一个聊天机器人提供错误答案，可能只是浪费一些时间。 但一个能够操作金融账户、修改数据库或者控制业务系统的Agent，一次错误就可能造成严重后果。 因此，未来真正成熟的Agent系统，必须具备明确的权限管理、操作记录、关键步骤审批、异常检测和错误恢复能力。 优秀的Agent不是完全不需要人类的系统，而是能够清楚判断什么时候可以自主执行，什么时候必须让人类确认的系统。 安全和可靠性，并不是Agent商业化之后才需要解决的问题。 它们本身就是商业竞争力的一部分。 07｜当AI能够完成越来越多的工作，人类还剩下什么价值？ 这是最值得深入思考的问题之一。 工业革命通过机器扩大了人类的体力。 计算机革命通过软件扩大了人类处理信息的能力。 而AI Agent可能进一步扩大人类组织和执行复杂任务的能力。 过去，你必须亲自完成大量操作。 未来，你可能只需要告诉AI自己的目标，然后负责监督和检查。 那么，人类是否会因此变得不再重要？ 答案可能恰恰相反。 当执行工作越来越容易，决定执行什么就变得更加重要。 当获取知识越来越容易，判断信息是否可信就变得更加重要。 当制作图片、视频和文章的成本越来越低，真正的创造力和独特观点就可能变得更加稀缺。 当AI能够提出许多解决方案，人类仍然需要判断哪一个值得采用，以及应该承担什么风险。 这可能改变未来社会对能力的评价方式。 过去，拥有某项具体技能可能意味着竞争优势。 未来，很多技能可能通过AI工具获得。 但如何提出有价值的问题，如何判断结果是否正确，如何组织资源，以及如何承担决策责任，仍然十分重要。 当然，不能因此忽视AI对就业的冲击。 部分标准化岗位可能被压缩，一些岗位可能重新定义，新的工作形式也可能出现。 而AI带来的收益，不一定会自动平均分配给所有人。 拥有技术、资本、数据和客户渠道的人，可能更容易获得收益。 AI Agent不只是一个技术革命，也可能成为一次深刻的经济结构调整。 08｜下一代互联网，可能是Agent与Agent之间的互联网 回顾互联网的发展历史。 最初，我们通过网站获取信息。 后来，社交媒体让每个人都可以创造和传播内容。 移动互联网让数字服务进入日常生活的每一个角落。 而AI Agent可能把互联网推向新的阶段。 未来，每个人可能拥有一个或者多个长期运行的智能体。 它们可以帮助你管理日程、整理文件、监控信息、处理工作，甚至在不同系统之间协调任务。 企业也可能拥有自己的Agent。 消费者的Agent负责寻找商品。 商家的Agent负责回答产品问题。 物流系统的Agent负责安排运输。 金融系统的Agent负责处理授权付款。 部分商业活动可能通过Agent之间的沟通和协商完成。 这意味着，未来互联网的一部分流量，可能不再来自人类亲自点击网页，而是来自智能体之间的数据交换和任务调用。 但这也会产生新的问题。 如何区分真人与AI？ 如何防止Agent冒充合法用户？ 如何确认一个Agent确实拥有执行某项操作的权限？ 如果Agent之间发生错误交易，责任由谁承担？ 如果网络上大量内容都由AI自动生成，我们又该如何保证信息的真实性？ 这些问题将决定Agent互联网能够发展到什么程度。 09｜最终判断：AI Agent正在改变劳动与价值创造的关系 未来并不一定所有App都会消失。 也不意味着所有人类工作都会被AI取代。 但是，软件的价值可能逐渐从提供操作界面，转向提供能够被智能体调用的服务能力。 过去，我们学习如何操作软件。 未来，软件可能越来越需要适应AI Agent的工作方式。 过去，企业通过招聘员工、购买软件和建立流程来扩大业务。 未来，企业可能同时管理人类员工和数字智能体。 过去，个人的生产能力主要取决于掌握多少技能、投入多少时间。 未来，它可能越来越取决于能否定义正确目标、组织高效系统，以及判断复杂问题。 AI正在降低执行工作的门槛，但它不会自动消除判断的价值。 AI Agent真正令人兴奋的地方，不是让人类从此什么都不用做。 而是让我们有机会将大量重复性工作交给机器，把更多精力投入创造、科研、决策，以及自己真正关心的事情。 但这场革命能否真正改善普通人的生活，仍取决于技术可靠性、使用成本、社会制度以及收益如何分配。 未来最重要的能力，可能不再是你能够亲自完成多少工作。 而是你能否识别真正有价值的问题，作出可靠的判断，并利用AI把想法变成现实。 最终，AI Agent可能改变的不仅是我们使用互联网的方式。 它可能重新定义一个人、一家公司，甚至整个社会能够创造多少价值。 #AI #AIAgent #人工智能 #AI创业 #未来科技 #科技趋势