When I talk to customers about AI, there’s one question that always seems to come up: “Will AI take my job?” My answer, “Maybe.” It’s an honest response that often catches people off-guard, but the answer is grounded in every technology shift since the Industrial Revolution. Some work will be automated, some job families will disappear, some industries will consolidate, some of the skills we have today will be commodities tomorrow.当我和客户谈论人工智能时，有一个问题似乎总是会被提起：“AI会取代我的工作吗？”我的回答是：“也许会。”这个诚实的答案常常让人措手不及，但它基于自工业革命以来每一次技术变革的逻辑。有些工作会被自动化，有些职业类别会消失，有些行业会整合，我们今天掌握的某些技能明天就会变成寻常之物。

If I reframe the question to “Will AI make me obsolete?”, my answer is, “No. Not if you evolve.” While we’re seeing incredible results from AI, it’s still just a tool. And as we’ve seen throughout history, staying relevant is tied to our ability to adapt, to use the tools at our disposal to build better and faster, and to solve harder problems.如果把问题换成“AI会让我过时吗？”，我的回答是：“不会，只要你能进化。”尽管我们看到了AI带来的惊人成果，但它终究只是一个工具。纵观历史，保持自身价值的关键在于适应能力——利用手头的工具更快更好地构建，并解决更困难的问题。

I evolve by getting deep with new tools. I need time to do real work with them and confirm or disprove any initial skepticism. I need to develop honest opinions about what works and what doesn’t work. Over the past few weeks, I’ve been spending the time I set aside to learn with Kiro Crew, an open-source agent workspace we recently released. I use Kiro Crew agents to run research tasks, monitor my schedule, manage cron jobs, and help prepare me for meetings. One piece of this that has captured my attention is how Kiro Crew manages memory across multiple independent agents. How it decides what to keep and what to throw away. How it implicitly learns from our conversations. How the agents improve themselves over time.我通过深入掌握新工具来进化。我需要时间用它们做实际工作，以证实或推翻最初的怀疑。我需要形成对哪些有效、哪些无效的诚实看法。过去几周，我一直在用自己留出的学习时间研究Kiro Crew——我们最近发布的一个开源智能体工作区。我用Kiro Crew智能体执行研究任务、监控日程、管理定时任务，并帮助我为会议做准备。其中特别吸引我的一点是，Kiro Crew如何在多个独立智能体之间管理记忆：它如何决定保留什么、丢弃什么；如何从我们的对话中隐式学习；智能体如何随着时间自我改进。

This got me thinking about how we – people – store information. Back in 2023, I visited Cortical Labs, an Australian biotech startup that places lab-grown human neurons on silicon chips to use them as a computing substrate. I watched them use a network built on living neurons to play the game of Pong, and I was blown away. A single high-end GPU can draw 700 watts of power and needs hundreds of thousands of training iterations to learn to play Pong. The human brain draws 20 watts and only needs a handful of attempts to do the same.这让我想到我们人类是如何存储信息的。2023年，我参观了Cortical Labs——一家澳大利亚生物技术初创公司，该公司将实验室培养的人类神经元放在硅芯片上，作为计算基质。我看着他们用活体神经元构建的网络玩《Pong》，感到震撼。一块高端GPU功耗可达700瓦，需要数十万次训练迭代才能学会玩Pong。人脑功耗只有20瓦，只需几次尝试就能做到同样的事。

The concept of “memory” ties into some of the work I’ve been doing to prepare for my annual tech predictions. I’ve been reading the research of neuroscientist Jeff Hawkins – who also happens to be the creator of one of my all-time favorite pieces of technology, the PalmPilot. In A Thousand Brains: A New Theory of Intelligence, he suggests that the brain distributes knowledge across thousands of cortical columns, each of which learns its own model through direct sensory experience. Each column predicts what it will sense next and updates when the prediction is wrong. Memories consolidate in the background, during sleep, when raw experience is distilled into structured long-term knowledge. What repeats is strengthened. What does not repeat weakens.“记忆”的概念与我为准备年度科技预测所做的一些工作密切相关。我一直在读神经科学家杰夫·霍金斯的研究——他恰好也是我最喜欢的科技产品之一PalmPilot的创造者。在《千脑智能：一种新的智能理论》中，他提出，大脑将知识分布在数千个皮层柱中，每个皮层柱通过直接的感觉经验学习自己的模型。每个皮层柱会预测接下来会感知到什么，并在预测错误时进行更新。记忆在后台整合，尤其是在睡眠期间，原始经验会浓缩为结构化的长期知识。重复出现的记忆被强化，不重复的则被削弱。

The more I learn about how the brain works, the more people I find working hard at this very moment to find the answers. I’ve talked to computer scientists and chip designers and neuroscientists who are all reverse-engineering the brain from different angles and arriving at the same answers, which has led me to an observation: the brain is the oldest architecture in computing.我对大脑工作机制了解得越多，就越发现此刻正有许多人在努力寻找答案。我和计算机科学家、芯片设计师、神经科学家交流过，他们从不同角度对人脑进行逆向工程，却得出相同的结论。这让我产生了一个观察：大脑是计算领域最古老的架构。

It is older than the Von Neumann architecture, older than the abacus, older than mathematics, older than language itself. Hundreds of millions of years of evolution produced structured memory, background consolidation, graceful decay, prediction, and learning from errors, all done at a fraction of the energy we use to approximate the same in silicon.它比冯·诺依曼架构更古老，比算盘更古老，比数学更古老，比语言本身更古老。数亿年的进化产生了结构化记忆、后台整合、优雅的衰减、预测和从错误中学习，而这一切所消耗的能量，仅为我们用硅基芯片实现类似功能所需能量的一小部分。

My journey into the internals of Kiro Crew gave me a concrete look at these principles in software. A small team of engineers built it by starting from practical engineering constraints, but the architecture they arrived at looks remarkably similar to the one evolution arrived at. It maintains several distinct memory types working together. In this case, memories are stored across markdown files, a local vector database, and a key-value index. Like the brain, some memories decay over time, so recall remains efficient. And, just as Hawkins posits how our brain works, consolidation of these memories runs in the background. Recent memories are kept in full detail, older ones compress, and the insignificant ones are pruned entirely. Without active forgetting, any system that accumulates experience will eventually spend more time sorting through old context than doing useful work.我对Kiro Crew内部的探索让我在软件中真切看到了这些原理。一个小型工程师团队从实际的工程约束出发构建了它，但他们最终得到的架构与进化得出的架构惊人地相似。它维护几种不同的记忆类型协同工作。在这个场景中，记忆存储在Markdown文件、本地向量数据库和键值索引中。像大脑一样，有些记忆会随时间衰减，从而保持高效回忆。与霍金斯提出的大脑工作原理相同，这些记忆的整合在后台运行。近期记忆保持完整细节，较旧的记忆被压缩，无关紧要的记忆被彻底删除。如果没有主动遗忘，任何不断积累经验的系统最终都会花更多时间整理旧上下文，而不是做有用的工作。

Memory enables something else that matters when I think about where agents are headed: trusted autonomy. The agents in Kiro Crew can run remotely around the clock. They can interact with each other, they can coordinate on tasks and generate reusable skills from patterns in your work. While doing my research on the brain, one of my agents suggested building a reading dashboard for me to track what I’d read and what I should read next. Agents that learn from you and remember your rules and preferences earn the trust to do much of the work unattended. I still review what they do and what tools they use to do it, but there’s a growing trust I’m finding with these agents that I have had trouble establishing with other tools that have come before them.记忆还使另一件重要的事成为可能——尤其在思考智能体未来方向时：值得信任的自主性。Kiro Crew中的智能体可以全天候远程运行。它们能相互交互，协调任务，并从你的工作模式中生成可复用的技能。在我做关于大脑的研究时，我的一个智能体建议为我构建一个阅读看板，用来跟踪我读过的内容和下一步该读什么。那些能向你学习、记住你的规则和偏好的智能体，能够赢得信任，在无人看管的情况下完成大量工作。我仍然会审查它们的工作内容和所用工具，但我发现，自己对这类智能体的信任正在增长，而这种信任是此前其他工具很难让我建立的。

The oldest architecture in computing is still the best architecture in computing, and it always will be. Among the people building our next tools, the ones who spend the time thinking deeply about how the brain works will be the ones who make the biggest breakthroughs.计算领域最古老的架构依然是计算领域最好的架构，而且永远如此。在打造未来工具的人中，那些花时间深入思考大脑工作原理的人，将取得最重大的突破。

Now, go build.现在，去动手构建吧。