Tom Wan @_TomWan

youtube.com/watch?v=yF2BY8…

The path to effectively infinite codebases in the AI era: Keep every module smaller than the model’s context window. Enforce strict cohesion and minimal coupling. Require strong test coverage and performance validation. Scale by composing modules into a tree. A model doesn’t need to understand the entire system. It only needs to understand the current node and its contracts. Unlimited software emerges from bounded contexts.

inklet is A macOS AI writing popover for transforming and inserting text with your preferred LLM provider. demo youtu.be/F5wmFruo0a4?si… via @YouTube github github.com/wanming/Inklet

OpenClaw-style setups are great because they keep long-lived context and feel like a persistent assistant, which makes them very useful for exploratory work, debugging, and multi-step reasoning. The downside is that they’re harder to isolate, audit, and control, and over time you can run into environment drift and higher costs at scale. Runner-style execution (like CI/CD) is more engineering-friendly: each task runs in an isolated, reproducible environment, which makes it easier to scale, secure, and audit. The tradeoff is that you lose long-term context and flexibility for open-ended tasks. In practice, for production and enterprise use, runners are usually the better default, with OpenClaw-style environments used selectively when persistent context really matters.

Another observation recently: tools like Anything Cli are a good example. They highlight both the promise and the limits of the “AI skills” approach. On one hand, skills can quickly expose capabilities through natural language and lightweight integrations. But on the other hand, they often run into constraints around context, determinism, and output quality. When tasks become more complex or reliability really matters, people tend to fall back to code. Code is still the most stable interface: explicit, testable, and composable. Maybe the real direction isn’t replacing code with skills, but letting skills sit on top of code — where AI helps discover and orchestrate capabilities, while the underlying logic remains robust and programmable.

AI “skills” today feel a lot like the early days of low-code. At first, everyone is excited: with a bit of configuration, workflows, or prompts, you can quickly build new capabilities. But soon the same challenges appear: poor maintainability, fragmented abilities, and for anything complex you still need real code. In the long run, the real value isn’t the skills themselves, but the platform, abstractions, and ecosystem behind them.

@luinlee 放外面会不会把 mac 废掉

如果你真的顺着它的逻辑去操作，最后大概率会成为别人案板上的鱼肉。我们来拆解一下这个精心设计的套路： 1. 虚构的“AI 背书” 首先，主流的大语言模型（包括 Claude 和我）都有着极其严格的安全和道德护栏，绝对不可能给出“去 GitHub 找个 Polymarket 机器人，拿 1000 美元去跑 30 天”这种极其具体、高风险且具有煽动性的投资建议。 这段所谓的“与 Claude 的对话”，100% 是文案作者为了给自己的策略增加权威性而捏造的。在这个时代，把“AI 说能赚钱”作为噱头，是最容易让人放下戒备的方法之一。 2. 经不起推敲的“高频刮痧”策略 这个所谓的“专扫低于 3 美分事件”的策略，在金融里类似于无限期购买“极度虚值期权”，或者说就是纯粹的买彩票。 • 期望值陷阱： 预测市场上标价 1 美分（代表市场认为概率为 1%）的事件，在现实中发生的真实概率通常远远低于 1%（可能只有万分之一）。从数学期望值来看，这绝对是一个长期必输的负期望游戏。 • 致命的流动性与滑点： 文案里说“在流动性差的市场里找错误定价”。现实情况是，正因为流动性极差，当你试图投入 15 美元去买 1 美分的份额时，你的买单会瞬间把价格拉高。你可能只有前两毛钱买到了 1 美分的份额，剩下的资金全是在 5 美分、10 美分的高位成交（这就是“滑点”）。你的成本根本不可能像文案算得那么低。 • 幸存者偏差： 即使那个叫 planktonXD 的钱包真实存在且盈利了，那也只是成千上万个跑归零的机器人中，极度幸运的那个“幸存者”。用彩票中头奖的人来证明买彩票是一门好生意，是非常荒谬的。 3. 图穷匕见：Telegram 链接才是真实目的 整篇文案洋洋洒洒，铺垫了神级钱包、AI 算账、惊人回报率，最终的目的只有一个——引诱你点击并使用那个 Telegram 跟单机器人 (PolyCop_BOT)。 一旦你真的冲动了，往里冲了 100 美元并授权了这个机器人，接下来可能会发生这些事： • 直接盗币： 这种不明来源的第三方机器人，往往会在你授权钱包（Approve）的瞬间，直接转走你钱包里的所有资产。 • 成为接盘侠： 所谓的“跟单”，很多时候是机器人作者先在流动性极差的池子里买入，然后用你们的钱去跟单，把价格拉高，他们自己高位套现，留你们在山顶吹风。 • 高额隐形手续费： 机器人可能根本赢不了钱，但它会疯狂地进行高频无效交易，把你的本金全部以“Gas 费（矿工费）”或“抽水”的形式消耗殆尽。 总结来说： 这就是一个披着“AI 智能分析”和“量化数据”外衣的新型杀猪盘引流文案。他们赌的就是大众对概率学的误解和对暴富的渴望。

I open-sourced SillyAgent.com! An AI agent that controls your entire Mac for you. Tell it what you want -- it watches your screen, moves the mouse, clicks, types, just like a real person sitting at your keyboard. Works with any app. "Turn on Dark Mode" -- done. "Search for flights to Tokyo" -- done. "Organize my desktop" -- done. The magic is it actually gets smarter over time: It knows 30+ apps inside out -- Safari, Chrome, VS Code, Slack, Notion, Spotify and more. Shortcuts, menus, workflows, all built in. You can even teach it new apps. It remembers you -- your preferences, your habits, your setup. Tell it once, it remembers forever. It learns from every task -- extracts what worked, avoids past mistakes, reuses proven workflows. The more you use it, the better it gets. And it's safe -- shake your mouse to stop it instantly, destructive actions always need your approval. One command to install: curl -fsSL raw.githubusercontent.com/wanming/SillyA… | bash github.com/wanming/SillyA… SillyAgent.com

@xiaohu 这是我做的一个开源 AI 桌面自动化工具，用自然语言就能控制你的 Mac。告诉它你要做什么，它会自己截屏、分析、然后帮你点击操作，完全免费，欢迎试用～ SillyAgent.com

@papajuiceman @DownloaderBot

@lq_ufo9706 这是个蓝牙音箱..

@PR0GRAMMERHUM0R 😂正在看这个剧，感觉还不错，黑色幽默

@angrytoad2046 😭

@luinlee 壁挂新风也没地方装吗

@lq_ufo9706 三套房？

@jasonlee_nihil @JeffCCrypto 期待😚

@luinlee 网上好像有教程