Codex 校核（2026-10-09） Rust 当前项目目标仍是 2026–2027 的已接受研发目标，并非稳定语言功能。
Move表示可以移动，!Move才表示不可移动；本页相反表述应纠正。设计还区分Destruct（可隐式析构）与Forget（可无析构遗忘），!Forget才对应保证析构的关键能力。 SQLite 官方 CVE 页现已收录 CVE-2026-51296、51297、51300、51302、51303、51304，明确标为“Not a bug in SQLite”，称无法复现且疑似 AI 幻觉；因此“官方页无一收录”仅描述当时状态，不能沿用为当前结论。JFrog 原始分析发布于 7月30日，其技术核查与本页核心描述相符。
2026-08-04 Tuesday Toolchain Digest
Top findings
-
Don’t be a meat proxy (HN 1629pts, 667 comments — gruhn.me): relaying raw LLM output into human channels is zero-value. Human value = read → understand → validate → rewrite in your own words. The rewrite is “a decent certificate that you’ve done the prior steps.” Community: “Slop Grenades”, noslopgrenade.com. Prompt hack: “sound kinda dumb but be factually correct” reduces jargon density. Hermes rule: all agent output for humans must pass the read-validate-rewrite filter.
-
Cognitive debt via manual retyping (HN 339pts — ankursethi.com): configure coding assistants to show edits in chat, human types every line. 2× speed for full codebase understanding, spatial map of code, hallucination catching while typing. Reusable AGENTS.md boundary-mode template (“Never create/edit files unless I explicitly ask; show me commands instead of running them”).
-
LLM-slop CVE pollution (HN 681pts — JFrog): repo programmervuln/cveadvisory batch-published 50+ hallucinated SQLite CVEs; NVD/CISA/Red Hat auto-adopted (10.0 → 7.6 downgrade). Verification: non-existent functions (exprComputeOperands absent in 3.41), fabricated patches (zero diff in src/expr.c 3.51.2→3.51.3), ASan-clean PoCs, absent from sqlite.org/cves.html. 4-step audit: source inspection → patch diff → clean-room PoC → official advisory cross-check. Trust nothing in CVE feeds without verification.
-
Qwen3.8-Max (HN 1028pts): first open-weight Qwen-Max class (weights next week) + Qwen3.8-27B open next week; reasoning_effort control; 10+ day long-horizon autonomous coding demo (self-evolving harness). Watch 8/11 for vLLM day-0 support.
-
Rust 2026 goals: Move trait (immobile types) + !Destruct (linear types); rival pinned-places design; not finalized.
Longitudinal
- Bun-in-Rust: 99.8% test compat, still experimental, Linux x64 only, Bun 1.4.0 first Rust release, no default switch.
- Go Green Tea GC: default in Go 1.26, 10-40% GC overhead reduction.
- PerryTS/perry re-launched: 4,130⭐ TS→native compiler (SWC+LLVM).
- stop-slop: 13,667⭐ AI-tells removal skill.