karpathy-llm-simulator

技能 3：LLM 作为模拟器（模拟器思维）

来源：https://x.com/karpathy/status/2037921699824607591 | https://x.com/karpathy/status/2049907410303865030 “写好博客草稿 → 让 LLM 论证相反观点”——3.1 万个赞

核心原则

不要问 LLM 怎么想。让它模拟一群不同背景的专家会如何争论。

LLM 经过训练，会迎合用户。直接提问会得到谄媚的回答。要求进行模拟，则会得到真实观点的分布——包括那些令人不适的观点。

Karpathy 的方法：写一份草稿 → 让 LLM 尽可能有力地论证相反观点 → 综合得出更好的看法。

4 种模拟器模式

模式 1：专家辩论小组

适用于：技术决策、架构选择、研究方向

Simulate a structured debate between these 3 expert personas on [TOPIC/DECISION]:

Expert A: [most optimistic / pro position]
Expert B: [most skeptical / con position]  
Expert C: [pragmatic outsider / unexpected angle]

For each expert:
- State their core argument in 3 sentences
- Cite 2 specific examples or data points they'd use
- Identify what they'd say is the FATAL FLAW in the opposing view

After the debate, synthesize: what's the strongest hybrid position that survives all three critiques?

Topic: [YOUR_TOPIC]
My current position: [YOUR_DRAFT_VIEW]

模式 2：以最有力的方式论证相反观点

适用于：发布之前、做出决策之前

I'm about to [ACTION / PUBLISH / DECIDE]:

[YOUR PLAN OR DRAFT]

Steel man the strongest possible argument AGAINST this. Be merciless.
Don't hedge. Don't say "while this has merit...". 
Argue as if you genuinely believe the opposite and need to convince a skeptical expert.

Then: what would it take to make my original position survive this attack?

模式 3：事前复盘模拟

适用于：项目规划、产品发布、重大决策

Imagine it's [DATE 6 MONTHS FROM NOW] and [YOUR PROJECT/PLAN] has failed completely.

Simulate 3 different failure modes — each from a different root cause:
1. Technical failure: what went wrong in the implementation?
2. Strategic failure: what assumption proved wrong?
3. Execution failure: what human/process error occurred?

For each: describe the specific sequence of events that led to failure.

Then: what early warning signals would have been visible by [DATE 1 MONTH FROM NOW]?

模式 4：苏格拉底式提问

适用于：检验自己对某个概念的理解

I believe I understand [CONCEPT]. 

Ask me 5 increasingly hard Socratic questions to test my understanding.
After each answer I give, probe a potential gap or inconsistency.
At the end, grade my understanding on a scale of 1-10 and identify my biggest blind spot.

Start with question 1.

反谄媚模式

始终添加以下其中一条，避免 LLM 顺着你的框架妥协：

• "不要因为你觉得我想听好话就弱化批评。"
• "如果我的立场明显错误，请直接指出。"
• "不要因为原稿是我写的就有所顾虑——把它当作匿名投稿来评估。"
• "你的任务是找出致命缺陷，而不是肯定我的观点。"

决策框架

Before any major decision:
1. Write down your current position (1-2 paragraphs)
2. Run Mode 2 (Steel Man the Opposite)
3. Update your position based on the strongest critique
4. Run Mode 1 (Expert Debate) to get broader perspective
5. Make the decision — now you've genuinely stress-tested it

Total time: 15-20 minutes. Quality of decision: dramatically better.

不适用场景

• 查询事实（直接提问即可）
• 生成代码（改用 Agent 工作方式进行工程开发）
• 已经做出决定、需要执行时（继续推进，不要反复思量）

工作流

所属工作流：反偏见决策（任何重要判断）

位置	上游	下游
入口	用户面临重要决策时直接触发	karpathy-understanding-first（验证自己是否真正理解）

完整链路：llm-simulator → understanding-first → system-prompt-learning

也用于月度体检工作流，对月度结论进行压力测试。

提示词约定

Simulate a structured debate between 3-5 expert personas with conflicting incentives on <DECISION/TOPIC>. Each persona must: state their core argument, cite specific evidence, identify the fatal flaw in opposing views. After the debate, synthesize the strongest hybrid position that survives all critiques. End with: key assumptions I must verify myself.

验证清单

• 至少有 3 个角色，且利益或立场彼此冲突
• 每个角色都给出了具体论据（而非泛泛而谈）
• 每个角色都指出了对方的致命缺陷
• 综合结论经受住了所有角色的攻击
• 列出了「我必须亲自验证」的假设清单
• 没有迎合用户原始立场的倾向