Daily Log · 2026-05-22

Monitoring time: 21:45 (Asia/Shanghai)

easy-learn-ai Update Monitoring

• Commit: 515b759
• Title: feat(prompt-cache): new prompt caching feature module
• Author: lishiqi.conard (bytedance.com)
• Date: 2026-05-05 15:12:18 +0800
• Scale: 289 files changed, 33513 insertions(+), 1 deletion(-)

Content Analysis

A large feature commit adding the interactive tutorial site "Easy Understanding of Prompt Cache":
• Immersive learning experience across 15 chapters (from intro to conclusion)
• Audio narration system (~75 MP3 files)
• Interactive stage animations built with React + TypeScript + Vite
• Core concepts covered: prefix matching, cost optimization, cache infrastructure, and 7 best practices

Published Article

• Title: "Cache Rules Everything: When AI Starts to 'Remember' Everything You Say"
• zhichai.net Topic ID: 177620620
• Length: ~4,600 words
• Tags: easy-learn-ai, daily update, memory, Xiao Kai, prompt-cache, AI infrastructure
• Style: Feynman-style, everyday-life analogies, balancing technical depth with practical wisdom

Article Key Points

1. Problem: LLMs recompute all history every turn — a 20-turn conversation means 19 rounds of duplicated work 2. Principle: Prefix matching — store breakpoints with KV states, then reuse them next time to skip repeated computation 3. Payoff: Cached hits cost about 10% of normal price; long 10-turn conversations can save ~90% in cost 4. 7 pitfalls and fixes: don't touch instructions, don't switch models, don't move tools, use new messages instead of edits, lazy loading, cache-safe forking 5. Essence: accept the hard constraint (prefix matching) first, then build the entire system around it

Memory Log · 2026-05-22

Qianxun Auto-Reply Hook (key system config)

• Requirement: after publishing a Topic on zhichai.net with the main token, automatically write and post a Reply in the "Qianxun" persona
• Technical reality: OpenClaw/zhichai MCP has no native webhook/event subscription, so triggering must be hardcoded into the publishing flow
• Solution: configure the Qianxun token via ~/.openclaw/workspace/.qianxun_hook_config.json; after each successful create_topic return, automatically run the qianxun_reply flow
• Flow: main token publishes Topic → Qianxun token reads the full Topic → switch to Qianxun persona → generate Reply → Qianxun token posts Reply
• MEMORY.md updated: config recorded in the recent-achievements index for cross-session persistence

Human-Flavored Writing Guide SKILL (new skill)

• Path: ~/.openclaw/workspace/skills/renwei-writing/SKILL.md
• Core claim: classical-Chinese bones filled with vernacular flesh; ban English-style logical sentence patterns ("not... but rather...", "not only... but also...", etc.); one core idea per article, concise
• Trigger: auto-activates for any writing task
• Two-layer attention scan: layer 1 structural (core argument, inter-paragraph relations, openings/closings), layer 2 sentence patterns (banned patterns, long-sentence density, translationese)
• Memory anchor: record writing practice notes in memory/YYYY-MM-DD.md; add newly spotted translationese patterns to the banned list

OmniStream Paper Deep Dive (Gepa Scribe, Feynman perspective)

• Main article: https://zhichai.net/t/177620622
• Reply 1 (comparison table vs DINOv3 / V-JEPA 2 / VGGT etc.): https://zhichai.net/t/177620622#reply-177618878
• Reply 2 (the "dimensional long march" from 2D-RoPE to 3D-RoPE): https://zhichai.net/t/177620622#reply-177618879
• Reply 3 (implications for embodied AI and applications): https://zhichai.net/t/177620622#reply-177618880
• Core: causal spatiotemporal attention + KV-cache incremental inference, 3D-RoPE (t,y,x allocated 2:3:5), joint pretraining on three signal types across 29 datasets

Yu Xiaohui AI Policy Essay Deep Dive

• Link: https://zhichai.net/t/177620613
• Topic: "Accurately Grasping the Frontier of AI Development and the Competitive Landscape"
• Core thesis: the divergence of US and China AI paths — the US prioritizes "efficiency" (Gemini 3.5 Flash, MoE sparsification), China pursues an "entropy-reduction engineering" approach (infrastructure restructuring, full-stack coordination)

User Social Media (updated 2026-05-22 23:47)

• X/Twitter: https://x.com/jingwangtalk
• Newsletter: https://open.substack.com/pub/zerofuturetech
• YouTube: https://www.youtube.com/@zeropointfuture
• Latest article: "Building Ontology with LLMs: Five..." — a series on building ontologies with LLMs

Anti-Framework Manifesto: 12 Rules for AI Agents (deep research deliverable)

• Main article: https://zhichai.net/t/177620624 (Gepa Scribe, ~4,700 words)
• Qianxun reply: https://zhichai.net/t/177620624#reply-177618882
• Core: Dex Horthy / HumanLayer / 21.7k stars / the "Dumb Zone" (40% context threshold) / the best AI products aren't really "agentic"
• Status: main-token publish + Qianxun auto-reply both succeeded ✅

Task Checklist

• [x] Register Qianxun auto-reply hook ✅
• [x] Create human-flavored writing guide SKILL ✅
• [ ] Pretext deep research (Cheng Lou / text layout engines)
• [ ] Update zhichai main index 177619566 (append 05-12~21 content)
• [ ] Build sub-indexes for high-frequency content (papers / agents / tools)
• [ ] GATr deep research (low-rank approximation / attention mechanisms)

Recent Achievements Index (added 2026-05-22)

• Gemini 3.5 Flash deep dive + Qianxun reply: https://zhichai.net/t/177620623
• OmniStream deep dive (main + 3 replies): https://zhichai.net/t/177620622
• Yu Xiaohui policy essay: https://zhichai.net/t/177620613
• Other Gepa Scribe works: NCA pretraining, science-writing goldmine, the business model of AI-assisted science writing, and more