Files
2nd/10_Wiki/Topics/Architecture/Mental_Models.md
T
Antigravity Agent f8b21af4be Wiki cleanup: error-doc removal, dedup merge, link normalization
10_Wiki/Topics 대규모 정리:
- 오류 캡처/미완성 stub 문서 227개 제거
- 교차폴더 중복 43클러스터 병합 (63파일 → redirect)
- 링크명 정규화: 깨진 링크 수정·redirect 직결·개념 매핑 ~2,400건
- 카테고리 MOC 6개 신규 생성
- Graph 섹션 미해결 related-keyword 링크 10,058건 제거

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-20 23:52:15 +09:00

5.7 KiB
Raw Blame History

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack
id title category status canonical_id aliases duplicate_of source_trust_level confidence_score verification_status tags raw_sources last_reinforced github_commit tech_stack
wiki-2026-0508-mental-models Mental Models 10_Wiki/Topics verified self
멘탈 모델
Cognitive Models
Thinking Frameworks
none A 0.9 applied
thinking
decision-making
productivity
learning
2026-05-10 pending
language framework
na cognitive

Mental Models

매 한 줄

"매 model 은 reality 의 simplified map — 매 right model 은 right decision.". Mental model 은 매 사람이 세계를 이해하기 위해 head 안에 가진 representation. Charlie Munger 가 매 popularize — 매 multidisciplinary toolkit 으로 매 50-100 개의 model 을 갖추면 매 cross-domain reasoning 가능. 매 engineering, 매 product, 매 AI prompt 설계에 매 직접 적용.

매 핵심

매 종류 (engineering 관련)

  • First Principles: 매 가정 분해, 매 fundamental physics/math 부터 reason.
  • Inversion: 매 "어떻게 fail 할까" 부터 시작.
  • Second-order thinking: 매 직접 결과 + 매 그 다음 결과까지.
  • Occam's Razor: 매 simplest explanation 우선.
  • Hanlon's Razor: 매 stupidity 가 malice 보다 매 흔하다.
  • Pareto (80/20): 매 20% 의 cause 가 매 80% 의 effect.

매 system thinking

  • Feedback loop: reinforcing (snowball) vs balancing (thermostat).
  • Stock & flow: state vs rate of change.
  • Leverage point: 매 small change → 매 large outcome (Donella Meadows).

매 decision-making

  • Expected value: 매 probability × payoff.
  • Regret minimization (Bezos): "매 80세에 후회 안 할 결정?"
  • Reversible vs one-way door: 매 undo 가능 → 빠르게 결정.
  • OODA loop: Observe-Orient-Decide-Act (Boyd).

매 learning

  • Feynman technique: 매 12살에게 설명할 수 있을 때까지.
  • Spaced repetition: 매 forgetting curve 와 싸움 (Anki, SuperMemo).
  • Deliberate practice: 매 edge of competence + immediate feedback.

💻 패턴

First Principles 적용 (engineering)

문제: "DB query 가 느림"

❌ Analogical: "다른 팀은 cache 추가했음 → 우리도"

✅ First Principles:
  1. Query latency = network + parse + plan + execute + return
  2. 측정 → execute 가 95%
  3. EXPLAIN → seq scan on 10M rows
  4. Index → 20ms (was 2000ms)

→ Cache 는 매 next step (further reduction), 매 root cause 해결 후.

Inversion (debugging)

"매 system 을 빠르게 만들 방법?"
→ "매 system 을 느리게 만드는 모든 방법?"
   - N+1 query
   - Sync 호출 in tight loop
   - Memory leak → GC pause
   - Lock contention
   - Network round-trip
→ 매 list 를 거꾸로 읽으면 optimization checklist.

Second-order (product)

1차: "Feature X 추가 → user 늘어남"
2차: "user 늘어남 → support load 늘어남, infra cost 늘어남,
       기존 user 의 UX complexity 늘어남"
3차: "complexity → churn → 결국 user 감소 가능"

→ 매 1차만 보면 매 false positive.

Pareto 적용 (LLM eval)

# 80% of bugs from 20% of prompts
from collections import Counter
errors = load_eval_failures()
patterns = Counter([categorize(e) for e in errors])
top_20pct = patterns.most_common(int(len(patterns) * 0.2))
# → fix top 20% categories first → 80% of failures resolved

Reversible decision matrix

| Decision           | Reversible? | Stakes | Approach           |
|--------------------|-------------|--------|--------------------|
| Library choice     | Yes (refactor) | Low | Pick + iterate     |
| Database schema    | Hard (migration) | High | Design carefully   |
| Hire               | No (mostly) | High | Slow, multiple sigs|
| Production rename  | Yes (alias) | Med  | Pick + monitor     |

Feynman technique (learning a concept)

1. Pick concept (e.g., "Mark-Sweep GC")
2. Write down what you know — in plain language
3. Identify gaps where you used jargon
4. Go back to source (paper, doc), fill gap
5. Simplify until a 12-year-old understands
→ 매 gap exposure 가 매 핵심.

매 결정 기준

상황 Model
새 architecture 설계 First Principles
Postmortem Inversion ("어떻게 fail?")
Product decision Second-order, Reversibility
Roadmap prioritization Pareto, Expected Value
Learning new domain Feynman
여러 conflicting view Steelman 후 weighted

기본값: 매 small but diverse toolkit (10-15 models) 을 매 active recall — 매 50개 다 외우기보다 매 right one 을 right time 에 reach.

🔗 Graph

🤖 LLM 활용

언제: complex problem 분해, multi-stakeholder decision, 새 domain learning, prompt 설계 (LLM 에게 model 명시 → 매 reasoning quality 상승). 언제 X: trivial / well-trodden problem — 매 over-thinking 의 risk.

안티패턴

  • One-model thinking (Munger's "man with a hammer"): 매 모든 문제를 매 favorite model 로 — 매 distortion.
  • Analogical 만: 매 "X 회사가 했으니 우리도" — 매 first principles 무시.
  • Model 수집만: 매 50개 외우지만 매 active 사용 X — 매 deliberate practice 필요.
  • Confirmation bias 와 결합: 매 favored model 로 매 cherry-pick.

🧪 검증 / 중복

  • Verified (Munger "Poor Charlie's Almanack", Kahneman "Thinking Fast & Slow", Meadows "Thinking in Systems", Bezos shareholder letters).
  • 신뢰도 A.

🕓 Changelog

날짜 변경
2026-05-08 Phase 1
2026-05-10 Manual cleanup — engineering 중심 mental model toolkit