Tech blog by Eric Cao — Python, Distributed Systems, AI Applicationshttps://eric.run.place/zh-CNhttps://eric.run.place/blog/en/can-you-teach-an-ai-intern/https://eric.run.place/blog/en/can-you-teach-an-ai-intern/Hermes Agent's pitch is 'the agent that grows with you' — an AI that gets smarter on its own, like an intern who learns. I took it literally: with the model frozen, can the agent make itself smarter just by writing its own skills? The answer is counter-intuitive — a human-written skill takes the same model from 10% to 74%, but the skill the agent writes itself closes less than half that gap.Fri, 12 Jun 2026 00:00:00 GMThttps://eric.run.place/blog/zh/can-you-teach-an-ai-intern/https://eric.run.place/blog/zh/can-you-teach-an-ai-intern/Hermes Agent 主打「the agent that grows with you」——像个会自己成长的实习生。我把这句话当真了：模型一个权重都不动，全靠它自己攒 skill，真能越用越聪明吗？答案有点反直觉——一条人写的好 skill 能让同一个模型从 10% 的做对率涨到 74%，但让它自己写一条，连一半差距都补不上。Fri, 12 Jun 2026 00:00:00 GMThttps://eric.run.place/blog/en/lakefs-graveler/https://eric.run.place/blog/en/lakefs-graveler/An ETL job corrupts tens of thousands of parquet files, and when you open the S3 bucket you realize — there are no commits, no rollbacks, and the overwritten objects are gone forever. This article tears apart the design of lakeFS Graveler: how a two-level Merkle tree combined with hash-based chunking allows every commit at billion-file scale to write just 2 new files and reuse 99% of content in place — and how it shares the same mathematics as blockchain while pursuing the exact opposite goal.Mon, 08 Jun 2026 00:00:00 GMThttps://eric.run.place/blog/zh/lakefs-graveler/https://eric.run.place/blog/zh/lakefs-graveler/ETL 写坏了几万个 parquet 文件，你打开 S3 桶却发现——没有 commit、没有回滚，被覆盖的对象永远找不回来了。这篇文章拆开 lakeFS Graveler 的设计：一棵两层 Merkle 树加上基于哈希的分块，如何让十亿文件规模的每次提交只写 2 个新文件、99% 的内容原地复用——以及它和区块链共享了同一套数学，目标却截然相反。Mon, 08 Jun 2026 00:00:00 GMThttps://eric.run.place/blog/en/openclaw-not-the-answer/https://eric.run.place/blog/en/openclaw-not-the-answer/After a week of deep use and 300 million tokens burned, the verdict on OpenClaw: the problem isn't that AI isn't smart enough — it's that pure natural language interaction has fundamental flaws for execution-based tasks. An analysis of four interaction modes through the lens of one core question: who absorbs the ambiguity?Tue, 10 Mar 2026 00:00:00 GMThttps://eric.run.place/blog/zh/openclaw-not-the-answer/https://eric.run.place/blog/zh/openclaw-not-the-answer/深度体验 OpenClaw 一周、烧完 3 亿 Token 后的结论：问题不是 AI 不够聪明，而是纯自然语言交互在执行型任务中存在根本缺陷。用四种交互模式分析「谁来消化歧义」这个核心问题。Tue, 10 Mar 2026 00:00:00 GMT