<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Chiffonwind (清凤) · Journal</title><description>Chiffonwind (清凤) — Full-stack &amp; design engineer building AI agent systems, developer tools, and interfaces across design and engineering.</description><link>https://chiffonwind.com/</link><language>zh-CN</language><item><title>GPT-4.5 到底有多大？</title><link>https://chiffonwind.com/zh/journal/2026-05-07-gpt-45-size-estimation/</link><guid isPermaLink="true">https://chiffonwind.com/zh/journal/2026-05-07-gpt-45-size-estimation/</guid><description>一次非正式的估算：用缩减版 IKP 做了 320 道闭卷事实题，得出的等效事实容量在 1.3T–1.8T 之间；再结合定价和生成速度，推测它是活跃参数很大的 MoE。</description><pubDate>Fri, 08 May 2026 00:00:00 GMT</pubDate><category>技术</category><category>LLM</category><category>GPT-4.5</category><category>IKP</category><category>Benchmark</category><author>qff45688@gmail.com (Chiffonwind (清凤))</author></item><item><title>How big is GPT-4.5, really?</title><link>https://chiffonwind.com/journal/2026-05-07-gpt-45-size-estimation/</link><guid isPermaLink="true">https://chiffonwind.com/journal/2026-05-07-gpt-45-size-estimation/</guid><description>An informal estimate: a scaled-down IKP run of 320 closed-book factual questions puts its equivalent factual capacity at 1.3T–1.8T. Combined with pricing and generation speed, it points to an MoE with very large active parameters.</description><pubDate>Fri, 08 May 2026 00:00:00 GMT</pubDate><category>Tech</category><category>LLM</category><category>GPT-4.5</category><category>IKP</category><category>Benchmark</category><author>qff45688@gmail.com (Chiffonwind (清凤))</author></item><item><title>重启之后，你还认识我吗？</title><link>https://chiffonwind.com/zh/journal/2026-03-31-agent-kin/</link><guid isPermaLink="true">https://chiffonwind.com/zh/journal/2026-03-31-agent-kin/</guid><description>kin 与 silt 的设计思考：Agent 只保留一条持续的会话，不再“新建对话”，用 sleep / merge / recall 和上下文自编辑工具代替一刀切的 Compaction。</description><pubDate>Tue, 31 Mar 2026 00:00:00 GMT</pubDate><category>技术</category><category>AI Agent</category><category>Memory</category><category>Context Management</category><author>qff45688@gmail.com (Chiffonwind (清凤))</author></item><item><title>Claude Code 泄露之后，值得学的不是源码</title><link>https://chiffonwind.com/zh/journal/2026-03-31-claudecode-leak/</link><guid isPermaLink="true">https://chiffonwind.com/zh/journal/2026-03-31-claudecode-leak/</guid><description>Prompt 和工具循环早就不是护城河了。从做 Arkloop 的经验看，Agent 交互真正的差别在信息密度和叙事节奏。</description><pubDate>Tue, 31 Mar 2026 00:00:00 GMT</pubDate><category>技术</category><category>AI Agent</category><category>Claude Code</category><category>UI/UX</category><category>TUI</category><author>qff45688@gmail.com (Chiffonwind (清凤))</author></item><item><title>After a restart, do you still know me?</title><link>https://chiffonwind.com/journal/2026-03-31-agent-kin/</link><guid isPermaLink="true">https://chiffonwind.com/journal/2026-03-31-agent-kin/</guid><description>Design notes on kin and silt: an agent that keeps a single continuous session with no &quot;new chat&quot;, using sleep / merge / recall and self-editing context tools instead of one-shot compaction.</description><pubDate>Tue, 31 Mar 2026 00:00:00 GMT</pubDate><category>Tech</category><category>AI Agent</category><category>Memory</category><category>Context Management</category><author>qff45688@gmail.com (Chiffonwind (清凤))</author></item><item><title>After the Claude Code leak, the source isn&apos;t what&apos;s worth learning</title><link>https://chiffonwind.com/journal/2026-03-31-claudecode-leak/</link><guid isPermaLink="true">https://chiffonwind.com/journal/2026-03-31-claudecode-leak/</guid><description>Prompts and tool loops stopped being a moat long ago. From building Arkloop, the real difference in agent interaction is information density and narrative pacing.</description><pubDate>Tue, 31 Mar 2026 00:00:00 GMT</pubDate><category>Tech</category><category>AI Agent</category><category>Claude Code</category><category>UI/UX</category><category>TUI</category><author>qff45688@gmail.com (Chiffonwind (清凤))</author></item><item><title>把 Llama 3 刻进芯片，有什么用？</title><link>https://chiffonwind.com/zh/journal/2026-02-23-taalas-jimmy/</link><guid isPermaLink="true">https://chiffonwind.com/zh/journal/2026-02-23-taalas-jimmy/</guid><description>讨论大多集中在 Taalas 的方案装不下大模型，却忽略了 Agent 系统里大量高频、低延迟的小模型调用，比如路由判断、搜索摘要和数据清洗。</description><pubDate>Mon, 23 Feb 2026 00:00:00 GMT</pubDate><category>技术</category><category>AI Hardware</category><category>Taalas</category><category>LLM Inference</category><category>AI Agent</category><author>qff45688@gmail.com (Chiffonwind (清凤))</author></item><item><title>What&apos;s the point of burning Llama 3 into a chip?</title><link>https://chiffonwind.com/journal/2026-02-23-taalas-jimmy/</link><guid isPermaLink="true">https://chiffonwind.com/journal/2026-02-23-taalas-jimmy/</guid><description>Most of the debate is about Taalas not fitting large models. It misses the huge volume of fast, low-latency small-model calls inside agent systems — routing, search summarization, data cleaning.</description><pubDate>Mon, 23 Feb 2026 00:00:00 GMT</pubDate><category>Tech</category><category>AI Hardware</category><category>Taalas</category><category>LLM Inference</category><category>AI Agent</category><author>qff45688@gmail.com (Chiffonwind (清凤))</author></item></channel></rss>