X Chinese Topic Brief¶
Editorial picks for Chinese-language writers: actionable, checkable, builder-facing. Morning X spine kept; midday merge adds non-X media/OSS/papers. Yesterday’s Kimi / offline training / Mythos / price war / Claude Code exits are not reopened.
Today's must-reads¶
-
Nanbeige4.3-33B open weights (see item 1). BOSS Zhipin lab fills the ~30B agent-coding teacher slot. Format: vs Qwen3.8-27B / local 24GB table. Status: draftable; official verified, benches = lab self-test.
-
Fold subagents × Claude 3.6× × Apollo zero ROI (see items 2–4). Architecture simplification + platform token burn + non-tech P&L still flat — CN gap on “usage ≠ ROI.” Format: migration checklist + dual chart. Status: draftable; guide / OpenRouter / TC verified.
-
AGENTS.md hijack × Cloudflare security-audit-skill (see items 5–6). Threat and official audit pattern in one window. Format: checklist + skill install note. Status: draftable; Forever Security author-claim, CF verified.
More topics¶
1. Nanbeige4.3-33B: CN ~30B open agent teacher¶
- Source & heat: nanbeige — ~311k views, 2,335 likes; HF https://huggingface.co/Nanbeige/Nanbeige4.3-33B
- Angle: Apache 2.0, teacher-then-3B distill, agentic coding/cowork, mainstream harnesses
- Why write: local coding-teacher selection table — params, VRAM, license, independent repro
- Verification: verified — official; fine benches author-claim
2. Anthropic: fold excess subagents¶
- Source & heat: cherry_mx_reds — ~830k views; guide https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop
- Angle: skills / tools / subagents rubric, ~2× customer case
- Why write: “harder to delete than add” audit template for CN agent teams
- Verification: verified — Anthropic
3. OpenRouter: Claude 5.5 tokens ~3.6׶
- Source & heat: PeterJ_Walker — ~163k views
- Angle: 7-day avg since Opus 5.5; Sonnet/Haiku follow; mostly net-new
- Why write: counter “CN open weights win on cheap tokens only”
- Verification: verified — OpenRouter
4. Apollo: S&P 493 “zero financial payoff”¶
- Source & heat: https://techcrunch.com/2026/10/25/ai-delivers-zero-financial-payoff-to-most-companies-according-to-apollo-economist/ — no X metrics
- Angle: ex-Magnificent margins flat; Slok “calculator vs judgment”
- Why write: procurement/IR post — platform usage vs customer P&L
- Verification: verified — TC; economist framing
5. Malicious AGENTS.md in Downloads¶
- Source & heat: digests e.g. choblin29
- Angle: silent Downloads drop; Codex >90% / Claude Code ~50%; global MCP/OAuth
- Why write: hard rule short post + config audit commands
- Verification: author-claim — Forever Security
6. Cloudflare security-audit-skill¶
- Source & heat: https://github.com/cloudflare/security-audit-skill · https://blog.cloudflare.com/build-your-own-vulnerability-harness/ — no X metrics
- Angle: six-phase audit,
confirmed/needs_validation, finder≠verifier, OS sandbox required - Why write: install + discipline note; pair with AGENTS.md threat
- Verification: verified — Cloudflare
7. Naive RAG paper + IQuest-Q1 (optional)¶
- Source & heat: https://arxiv.org/abs/2610.30755 · https://arxiv.org/abs/2610.30949 — no X metrics
- Angle: Raw-RAG wins 6/7; 320B/15B MoE CLI foundation open-source claim
- Why write: “delete memory complexity” beside fold-subagents; wait on IQuest weights
- Verification: paper pages verified; scores/release author-claim
8. Codex Day 20 / Thorsten Ball (lower priority)¶
- Source & heat: Day 20; thorstenball
- Angle: quota reset; Amp practice essay
- Why write: capacity calendar / engineering digest
- Verification: verified — principals