本文由 AI 分析生成
建立時間： 2026-03-23 來源： https://x.com/Lonely__MH/status/2036651579005194426
Summary
A plain-language explanation of the autoresearch iteration framework derived from Karpathy’s open-source project (630 lines of Python, 50k GitHub stars in days). The author demonstrates applying the framework to Xiaohongshu (Little Red Book) content generation, achieving a measurable quality gap of 62% between skill-guided and un-guided outputs. The core argument: autoresearch is not an ML tool — it is a universal iteration framework applicable to anything that can be scored.
對源自 Karpathy 開源項目（630 行 Python，數天內達到 5 萬 star）的 autoresearch 迭代框架的通俗解釋。作者展示了將框架應用於小紅書內容生成，在有 skill 指導和無指導輸出之間達到 62% 的可測量品質差距。核心論點：autoresearch 不是 ML 工具——它是可應用於任何可打分事物的通用迭代框架。
Key Points
- Core loop: change one thing → score → keep if better, revert if not → repeat (Karpathy applied this to ML training code; it generalizes to anything scoreable)
- Round 1 (6 criteria): skill-guided 100% vs. no-skill 66.7% — +33% gap
- Round 2 (8 criteria, added tag quality and quotability checks): skill-guided 100% vs. no-skill 37.5% — +62% gap
- Binary yes/no beats 1-10 because “does the opening sentence contain a specific number?” is objective; “how good is the opening?” is not
- A community member applied the same logic to media content review with a two-layer scoring system: rule-based checklist first, then “reader perspective” holistic score
- The framework belongs to no tool or platform — tools expire, frameworks don’t
Insights
The two-layer scoring approach from the community example is more sophisticated than the single-layer checklist: combining rule compliance (objective) with reader gestalt (subjective) addresses the “teaching to the test” failure mode where a skill satisfies all criteria but still produces poor output. This dual scoring design is directly applicable to improving vault analysis quality checks.
Connections
Raw Excerpt
“工具会过时，框架不会。Karpathy 今天开源的是 630 行代码，但他真正贡献的，是一个任何人都能拿走的迭代方法论。”