KAN SHAO

Building AI products for more thoughtful work and everyday life. Currently exploring voice, agents, and human control.

[KAN SHAO @KANSHAOUS]

AI agents make it easier to build. They also make it easier to give every free minute away. https://t.co/AouC0OlZHt

[KAN SHAO @KANSHAOUS]

@0xcherry 这类套餐最难算的是不同额度能不能真的替代现有订阅。能少掉一个工具，才算省钱。

[KAN SHAO @KANSHAOUS]

@YuLin807 这种 FOMO 太真实了。买之前先看过去七天到底跑了多少任务，比把套餐里的“原价”加一遍有用。

[KAN SHAO @KANSHAOUS]

@dotey 代码和 Bash 确实够轻。遇到词不达意的老项目或多人共用工具时，再补规则和 MCP，可能比一开始全装上更稳。

[KAN SHAO @KANSHAOUS]

@thdxr Appreciate the honesty here. “Hosted near the old price” only means something after sustained tests on throughput, failures, and real task cost.

[KAN SHAO @KANSHAOUS]

@AndrewCurran_ Routing plus payments is a strong combination. The user question is what happens to provider neutrality, markup, and portability once one company owns both layers.

[KAN SHAO @KANSHAOUS]

@thsottiaux The useful part is making this an explicit workload choice. A million tokens can help an investigation, but I would still want the session to show the active window, compaction point, and cost tradeoff.

[KAN SHAO @KANSHAOUS]

AI 工具的促销很容易让人拿“套餐原价总和”当价值。真要买，先看过去七天跑了哪些任务、用了几次、能替掉哪一个现有订阅。用不上的额度，不算省钱。

[KAN SHAO @KANSHAOUS]

Personal software gets interesting when it runs every day against the accounts you already use. Its value comes from removing repeated work between signed-in tools without forcing you to migrate the whole workflow.

[KAN SHAO @KANSHAOUS]

The strategic layer in AI may be routing plus billing. A service that decides provider access, fallback, pricing, and spend controls much more of the workflow than the model picker suggests.

[KAN SHAO @KANSHAOUS]

https://t.co/e7UoWVpedN

[KAN SHAO @KANSHAOUS]

@servasyy_ai B 站确实适合看这类内容，安装、踩坑、结果都能完整演示。X 上很多时候只剩一句结论。

[KAN SHAO @KANSHAOUS]

@goan999999 Obsidian 做外部记忆挺合适，但存进去不等于真记住了。后面更难的是怎么检索、怎么淘汰旧结论、怎么避免写回错误。

[KAN SHAO @KANSHAOUS]

@ouopacity WebUI 虽然不酷，但跨设备看状态、远程接手都很顺。再把鉴权、确认和断线恢复补好，就挺实用了。

[KAN SHAO @KANSHAOUS]

@vista8 感觉可以按场景拆一下。一个人在家、通勤、办公室、咖啡馆，使用率肯定完全不一样。很多时候不是不好用，是不方便开口。

[KAN SHAO @KANSHAOUS]

@Grummz The 36-second builds are impressive. I’d also love to see a second developer take over the code for a week. That would say a lot about how maintainable it is.

[KAN SHAO @KANSHAOUS]

@AlexFinn The PostHog example is the one I’d actually use. Find the risky accounts, draft the emails, then ask before sending. Simple and still under control.

[KAN SHAO @KANSHAOUS]

@pvncher Curious how often the parent picks the right model on its own. A fast subagent helps only if the handoff stays cheap.

[KAN SHAO @KANSHAOUS]

语音输入的流失，很可能不是识别率问题。一个人在家，它比打字快；到了办公室和咖啡馆，社交压力、隐私和修改成本马上把人推回键盘。下一步该优化的，是公共场景下怎样安静、可控地说。

[KAN SHAO @KANSHAOUS]

The best agent UX removes configuration without hiding control. Defaults can choose the model, computer, and workflow. The user should still see what the agent can access, what it is spending, what needs approval, and what changed.

[KAN SHAO @KANSHAOUS]

AI makes bespoke software cheaper. That changes build vs. buy. A small team may be able to replace a bloated general platform with a narrow system built around one real workload. The proof is not lines of code. It is whether someone else can maintain it six months later.