karpathy-understanding-first

技能 8：理解 > 外包

来源：https://x.com/karpathy/status/2049907410303865030 “你可以把思考外包出去，但不能把理解外包出去。”——46k 个赞 另见：锯齿状能力建模（Sequoia 对话）

核心原则

外包没问题。能力退化不行。

使用 AI 提速 10 倍。但每次跳过对某件事的理解，你都是在透支自己的未来。理解会产生复利。盲目发布 Agent 的输出则不会。

Karpathy 关于“锯齿状能力”的洞见：大模型在某些方面极其擅长（可验证、数据丰富的领域），在另一些方面则会悄然出错（分布外、全新组合）。你需要在脑中绘制一张地图，弄清哪些属于哪一类。

理解审查

每次借助 Agent 完成工作后，都运行以下流程：

I just had an AI help me [DESCRIBE WHAT WAS BUILT/WRITTEN/DECIDED].

Help me audit my own understanding:

1. List the 5 key decisions that were made (technical choices, assumptions, trade-offs)
2. For each decision: would I be able to explain WHY this choice was made to a colleague?
   Mark each: UNDERSTOOD / SHALLOW / BLACK BOX
3. For the BLACK BOX items: give me a 3-sentence explanation I can verify
4. What would I need to read/learn to make the SHALLOW items UNDERSTOOD?

Be honest. If I'm shipping something I don't understand, I need to know.

锯齿状能力地图

大模型的能力并非均衡一致。请采用以下心智模型：

高可靠性（表现稳定）

这些领域定义明确、可验证且数据丰富：

• 热门语言中的代码（Python、JS、SQL）
• 文本摘要和改写
• 结构化数据提取
• 知名算法
• 常见模式（REST API、SQL 查询、正则表达式）

中等可靠性（谨慎使用，务必验证）

• 热门程度较低的语言（Rust、Haskell、Elixir）
• 集成多个系统
• 涉及训练截止时间之后的特定版本/API 的内容
• 多步推理链
• 超出简单算术范围的数学

低可靠性（务必独立验证）

• 前沿研究（可能会虚构论文）
• 技术的全新组合
• 你无法轻易验证的领域特定知识
• 法律/医疗/金融细节
• 出错代价高昂的任何内容

能力检查提示词

在信任 AI 对重要事项的输出之前：

I need to evaluate how reliable your output is for this specific task:

Task type: [DESCRIBE WHAT YOU ASKED THE AI]
Output: [PASTE THE OUTPUT]

Help me assess:
1. How common is this exact pattern in training data? (ubiquitous / common / rare / novel)
2. What's the easiest way to verify this without running it?
3. What's the most likely error — hallucinated API? Wrong logic? Outdated syntax?
4. What search query would I use to double-check the key claim?
5. Should I trust this output? (YES/VERIFY_FIRST/DEFINITELY_CHECK)

防止能力退化的流程

Karpathy 会追踪哪些技能正在退化（被 AI 接管），哪些技能正在提升。每月检查：

Run a skills audit for [MONTH/YEAR]:

Things I've been having AI do for me recently:
[LIST 5-10 TASKS]

For each:
1. Is my ability to do this WITHOUT AI declining? (YES / NO / UNSURE)
2. Is that OK? (Critical skill I must maintain / Commodity I'm fine outsourcing)
3. Action: KEEP_DELEGATING / PRACTICE_MANUALLY / LEARN_DEEPLY

Final question: What's one thing I should deliberately do WITHOUT AI this month to stay sharp?

验证清单

发布任何由 Agent 生成的输出之前：

Understanding verification:
- [ ] I can explain what this does in plain English without looking at it
- [ ] I've traced through the logic manually for at least one example
- [ ] I know what the failure mode looks like
- [ ] I've run it / tested it myself (not just read the output)
- [ ] If it's wrong, I'll know how to debug it
- [ ] I'm not shipping something I'd be embarrassed to explain in a code review

If any box is unchecked: understand it before shipping.

学习提取提示词

当 AI 写出你不理解的内容时，要从中学习，而不是直接发布：

You wrote this code/text and I don't fully understand it:

[PASTE OUTPUT]

Teach me:
1. The underlying concept (assume I know [MY_LEVEL] — beginner/intermediate/expert)
2. Walk through the logic step by step
3. What would happen if I changed [SPECIFIC PART]?
4. Point me to the canonical resource where I can learn this properly
5. Give me a simpler version that demonstrates the core idea in isolation

I'm going to use this as a learning moment, not just copy-paste.

工作流

属于工作流：反偏见决策（第 2 步）+ 工作流：月度体检（第 2 步）

位置	上游	下游
C 第 2 步	karpathy-llm-simulator（辩论后）	karpathy-system-prompt-learning（沉淀规则）
D 第 2 步	karpathy-meta-reflection（审计后）	karpathy-practice-environments（建练习环境）

反偏见决策链路：llm-simulator → understanding-first → system-prompt-learning 月度体检链路：meta-reflection → understanding-first → practice-environments → education-first

提示词契约

After completing this task, append an Understanding Report: 1) Key assumptions made (list each), 2) What was verified (with evidence), 3) What remains inferred/unverified (with suggested verification method), 4) Capability boundary notes (where LLM might be off-distribution for this task), 5) Items I must personally inspect or understand before relying on this output.

验证清单

• 明确列出关键假设（不隐藏在输出中）
• 已验证项附有证据（测试通过/文档引用/数据对比）
• 未验证项附有明确的验证方法建议
• 标注大模型的能力边界（哪些部分可能超出其能力范围）
• 「你必须亲自查看」清单具体到文件/行/配置
• 人类确认已理解核心逻辑（不只是看了一眼）