About AI Anxiety
Facing the information anxiety of the AI era, shift your focus from external standards back to what you are doing: fight anxiety with action, and go to the bottom rather than trying to keep up with everything.
Jason's Notes
Some things I'm researching and exploring: Engineer · AI · Lifestyle · Health · ……
AI Coding: From Chat to an Agent-Driven Workflow
The point of AI coding is not to treat AI as a chat window, but to let an agent work inside a real project, accumulate compounding assets, and use a test-and-feedback loop to guarantee quality.
Can an Investing Agent Actually Make Me Money?
Live-money experiments show AI agents occasionally win, but are far from a stable strategy: the ones already making money are brokers, advisors, and course sellers, while an ordinary user relying on it for sustained gains is still an open question.
Embodied AI 2026: Robots Can Outrun Humans, So Why Can't They Fold a Shirt?
A robot that can run doesn't mean it can work: the half-marathon record has been broken, but cooking, laundry, and folding clothes are still limited by complex environments and long-flow reliability; what can really scale in the near term is more likely factories and warehouses.
具身智能 2026：机器人已经能跑赢人类，为什么还叠不好一件衣服？
机器人会跑不等于会干活：半马纪录已经被刷新，做饭、洗衣和叠衣仍受复杂环境与长流程可靠性限制，近期真正能规模化落地的更可能是工厂和仓库。
GraphRAG Two Years Later: Still Maintained, but Useless for Most Scenarios
Two years after GraphRAG was open-sourced it's still maintained, but most RAG scenarios don't need it in practice: for simple fact retrieval it's a negative asset; only multi-hop reasoning, cross-document synthesis, and relational queries are worth it — and it's more expensive and harder to maintain.
GraphRAG 两年后：还在维护，但大多数场景用不上
GraphRAG 开源两年后仍在维护，但大多数 RAG 场景实测用不上它：简单事实检索它是负资产，只有多跳推理、跨文档综述和关系型查询值得上，而且它更贵、更难维护。
How Prompt Caching Works: Hit Rate Is Decided by Prefix Stability
Under the same model, cache hit rate can differ by 4x — and it's not the model: prefix byte stability decides the hit rate. Explains the cache mechanism, seven real cache-breaking incidents, and how to avoid the pitfalls.
Mem0 Raised $24M and Stored 97.8% Garbage
While Mem0 announced a $24M raise, an independent audit found 97.8% of the 10,134 memories it stored over 32 days were garbage: agent-memory commercialization is running ahead of verifiable quality.
Mem0 融了 2400 万美元，也存下了 97.8% 的垃圾
Mem0 宣布完成 2400 万美元融资的同时，一份独立审计显示它 32 天存下的 10134 条记忆里 97.8% 是垃圾：agent 记忆的商业化跑在了可验证的质量前面。
The One Without a Sandbox Wrote "No Sandbox" Into Its Docs
Claude Code, Cursor, Codex, and OpenCode compared on sandboxing and security interception: where the boundary is placed, and who can see it when it fails.
唯一没有沙箱的那一家，把"没有沙箱"写进了官方文档
Claude Code、Cursor、Codex、OpenCode 四款 coding agent 在沙箱与安全拦截上的对照：边界放在哪一层，失效时谁能看见。