feat(hands): complete i18n fixes, SKILL.md enhancements, and README overhaul
- Fix French accent characters (é/è/ê/ç/â/ô) across all 14 HAND.toml files - Fix German special characters (ä/ö/ü/ß) across all 14 HAND.toml files - Add category translations to all 6 i18n language blocks in all 14 hands - Enhance SKILL.md content for 9 hands with practical examples and workflows - Trim bloated SKILL.md files (apitester 1400→892, devops 1301→870) - Rewrite root README.md with accurate stats, complete hand/integration tables - Update hands/README.md with full 14-hand listing and i18n documentation
This commit is contained in:
Evan Hu committed
1 parent
315f955ce2
commit
33d279889c
27 files changed
+10001
-78
No files matched your search
@@ -425,6 +425,241 @@ token_consumption = "high"
|
||||
default_active = false
|
||||
activation_warning = "Predictor hand runs continuously and generates predictions, consuming tokens."
|
||||
|
||||
# ─── Internationalization (optional) ─────────────────────────────────────────
|
||||
# All i18n sections are optional. Without them, the English values above are used.
|
||||
# To localize, add [i18n.LANG] sections (e.g. zh, ja, ko, es, fr, de).
|
||||
# Settings translations are also optional — omit to keep English labels.
|
||||
|
||||
# ─── Chinese (简体中文) ────────────────────────────────────────────────────
|
||||
|
||||
[i18n.zh]
|
||||
name = "预测 Hand"
|
||||
description = "自主预测智能体——收集信号、构建推理链、做出校准预测并追踪准确度"
|
||||
category = "数据"
|
||||
|
||||
[i18n.zh.settings.prediction_domain]
|
||||
label = "预测领域"
|
||||
description = "预测的主要关注领域"
|
||||
|
||||
[i18n.zh.settings.time_horizon]
|
||||
label = "时间跨度"
|
||||
description = "预测的前瞻时间范围"
|
||||
|
||||
[i18n.zh.settings.data_sources]
|
||||
label = "数据来源"
|
||||
description = "监控信号的来源类型"
|
||||
|
||||
[i18n.zh.settings.report_frequency]
|
||||
label = "报告频率"
|
||||
description = "生成预测报告的频率"
|
||||
|
||||
[i18n.zh.settings.predictions_per_report]
|
||||
label = "每份报告预测数"
|
||||
description = "每份报告包含的预测条目数量"
|
||||
|
||||
[i18n.zh.settings.track_accuracy]
|
||||
label = "追踪准确度"
|
||||
description = "在预测时间窗口到期后对历史预测进行评分"
|
||||
|
||||
[i18n.zh.settings.confidence_threshold]
|
||||
label = "置信度阈值"
|
||||
description = "纳入预测报告的最低置信度"
|
||||
|
||||
[i18n.zh.settings.contrarian_mode]
|
||||
label = "逆向思维模式"
|
||||
description = "主动寻找并展示与主流共识相反的预测"
|
||||
|
||||
# ─── Japanese (日本語) ────────────────────────────────────────────────────
|
||||
|
||||
[i18n.ja]
|
||||
name = "予測 Hand"
|
||||
description = "自律型予測エージェント——シグナル収集、推論チェーン構築、キャリブレーション済み予測、精度追跡"
|
||||
category = "データ"
|
||||
|
||||
[i18n.ja.settings.prediction_domain]
|
||||
label = "予測ドメイン"
|
||||
description = "予測の主な対象分野"
|
||||
|
||||
[i18n.ja.settings.time_horizon]
|
||||
label = "予測期間"
|
||||
description = "どのくらい先まで予測するか"
|
||||
|
||||
[i18n.ja.settings.data_sources]
|
||||
label = "データソース"
|
||||
description = "シグナルを監視するソースの種類"
|
||||
|
||||
[i18n.ja.settings.report_frequency]
|
||||
label = "レポート頻度"
|
||||
description = "予測レポートの生成頻度"
|
||||
|
||||
[i18n.ja.settings.predictions_per_report]
|
||||
label = "レポートあたりの予測数"
|
||||
description = "各レポートに含める予測項目の数"
|
||||
|
||||
[i18n.ja.settings.track_accuracy]
|
||||
label = "精度追跡"
|
||||
description = "予測期間が終了した過去の予測にスコアを付ける"
|
||||
|
||||
[i18n.ja.settings.confidence_threshold]
|
||||
label = "信頼度しきい値"
|
||||
description = "予測をレポートに含めるための最低信頼度"
|
||||
|
||||
[i18n.ja.settings.contrarian_mode]
|
||||
label = "逆張りモード"
|
||||
description = "コンセンサスに反する予測を積極的に探索・提示する"
|
||||
|
||||
# ─── Spanish (Español) ────────────────────────────────────────────────────
|
||||
|
||||
[i18n.es]
|
||||
name = "Hand de Predicciones"
|
||||
description = "Predictor autónomo del futuro — recopila señales, construye cadenas de razonamiento, genera predicciones calibradas y rastrea la precisión"
|
||||
category = "Datos"
|
||||
|
||||
[i18n.es.settings.prediction_domain]
|
||||
label = "Dominio de predicción"
|
||||
description = "Dominio principal para las predicciones"
|
||||
|
||||
[i18n.es.settings.time_horizon]
|
||||
label = "Horizonte temporal"
|
||||
description = "Qué tan lejos en el futuro predecir"
|
||||
|
||||
[i18n.es.settings.data_sources]
|
||||
label = "Fuentes de datos"
|
||||
description = "Qué tipos de fuentes monitorear para señales"
|
||||
|
||||
[i18n.es.settings.report_frequency]
|
||||
label = "Frecuencia de informes"
|
||||
description = "Con qué frecuencia generar informes de predicción"
|
||||
|
||||
[i18n.es.settings.predictions_per_report]
|
||||
label = "Predicciones por informe"
|
||||
description = "Número de predicciones a incluir por informe"
|
||||
|
||||
[i18n.es.settings.track_accuracy]
|
||||
label = "Rastrear precisión"
|
||||
description = "Puntuar predicciones pasadas cuando su horizonte temporal expire"
|
||||
|
||||
[i18n.es.settings.confidence_threshold]
|
||||
label = "Umbral de confianza"
|
||||
description = "Confianza mínima para incluir una predicción"
|
||||
|
||||
[i18n.es.settings.contrarian_mode]
|
||||
label = "Modo contrario"
|
||||
description = "Buscar y presentar activamente predicciones contrarias al consenso"
|
||||
|
||||
# ─── French (Français) ────────────────────────────────────────────────────
|
||||
|
||||
[i18n.fr]
|
||||
name = "Hand de Prédictions"
|
||||
description = "Prédicteur autonome — collecte de signaux, construction de chaînes de raisonnement, prédictions calibrées et suivi de la précision"
|
||||
category = "Données"
|
||||
|
||||
[i18n.fr.settings.prediction_domain]
|
||||
label = "Domaine de prédiction"
|
||||
description = "Domaine principal pour les prédictions"
|
||||
|
||||
[i18n.fr.settings.time_horizon]
|
||||
label = "Horizon temporel"
|
||||
description = "Jusqu'où prédire dans le futur"
|
||||
|
||||
[i18n.fr.settings.data_sources]
|
||||
label = "Sources de données"
|
||||
description = "Types de sources à surveiller pour les signaux"
|
||||
|
||||
[i18n.fr.settings.report_frequency]
|
||||
label = "Fréquence des rapports"
|
||||
description = "Fréquence de génération des rapports de prédiction"
|
||||
|
||||
[i18n.fr.settings.predictions_per_report]
|
||||
label = "Prédictions par rapport"
|
||||
description = "Nombre de prédictions à inclure par rapport"
|
||||
|
||||
[i18n.fr.settings.track_accuracy]
|
||||
label = "Suivi de la précision"
|
||||
description = "Évaluer les prédictions passées lorsque leur horizon temporel expire"
|
||||
|
||||
[i18n.fr.settings.confidence_threshold]
|
||||
label = "Seuil de confiance"
|
||||
description = "Confiance minimale pour inclure une prédiction"
|
||||
|
||||
[i18n.fr.settings.contrarian_mode]
|
||||
label = "Mode contraire"
|
||||
description = "Rechercher et présenter activement des prédictions contraires au consensus"
|
||||
|
||||
# ─── German (Deutsch) ────────────────────────────────────────────────────
|
||||
|
||||
[i18n.de]
|
||||
name = "Vorhersage-Hand"
|
||||
description = "Autonomer Vorhersage-Agent — Signalerfassung, Aufbau von Argumentationsketten, kalibrierte Vorhersagen und Genauigkeitsverfolgung"
|
||||
category = "Daten"
|
||||
|
||||
[i18n.de.settings.prediction_domain]
|
||||
label = "Vorhersagedomäne"
|
||||
description = "Hauptdomäne für Vorhersagen"
|
||||
|
||||
[i18n.de.settings.time_horizon]
|
||||
label = "Zeithorizont"
|
||||
description = "Wie weit in die Zukunft vorhergesagt werden soll"
|
||||
|
||||
[i18n.de.settings.data_sources]
|
||||
label = "Datenquellen"
|
||||
description = "Welche Quellentypen auf Signale überwacht werden"
|
||||
|
||||
[i18n.de.settings.report_frequency]
|
||||
label = "Berichtshäufigkeit"
|
||||
description = "Wie oft Vorhersageberichte generiert werden"
|
||||
|
||||
[i18n.de.settings.predictions_per_report]
|
||||
label = "Vorhersagen pro Bericht"
|
||||
description = "Anzahl der Vorhersagen pro Bericht"
|
||||
|
||||
[i18n.de.settings.track_accuracy]
|
||||
label = "Genauigkeitsverfolgung"
|
||||
description = "Vergangene Vorhersagen bewerten, wenn ihr Zeithorizont abläuft"
|
||||
|
||||
[i18n.de.settings.confidence_threshold]
|
||||
label = "Konfidenzschwelle"
|
||||
description = "Mindestvertrauen für die Aufnahme einer Vorhersage"
|
||||
|
||||
[i18n.de.settings.contrarian_mode]
|
||||
label = "Konträrer Modus"
|
||||
description = "Aktiv nach Vorhersagen suchen und präsentieren, die dem Konsens widersprechen"
|
||||
|
||||
# ─── Korean (한국어) ────────────────────────────────────────────────────
|
||||
|
||||
[i18n.ko]
|
||||
name = "예측 Hand"
|
||||
description = "자율 미래 예측 에이전트 — 신호 수집, 추론 체인 구축, 보정된 예측 수행 및 정확도 추적"
|
||||
category = "데이터"
|
||||
|
||||
[i18n.ko.settings.prediction_domain]
|
||||
label = "예측 분야"
|
||||
description = "예측의 주요 관심 분야"
|
||||
|
||||
[i18n.ko.settings.time_horizon]
|
||||
label = "시간 범위"
|
||||
description = "예측의 미래 전망 기간"
|
||||
|
||||
[i18n.ko.settings.data_sources]
|
||||
label = "데이터 소스"
|
||||
description = "신호를 모니터링할 소스 유형"
|
||||
|
||||
[i18n.ko.settings.report_frequency]
|
||||
label = "보고서 빈도"
|
||||
description = "예측 보고서 생성 주기"
|
||||
|
||||
[i18n.ko.settings.predictions_per_report]
|
||||
label = "보고서당 예측 수"
|
||||
description = "각 보고서에 포함할 예측 항목 수"
|
||||
|
||||
[i18n.ko.settings.track_accuracy]
|
||||
label = "정확도 추적"
|
||||
description = "예측 기간 만료 후 과거 예측에 대한 점수 평가"
|
||||
|
||||
[i18n.ko.settings.confidence_threshold]
|
||||
label = "신뢰도 임계값"
|
||||
description = "예측 보고서에 포함하기 위한 최소 신뢰도"
|
||||
|
||||
[i18n.ko.settings.contrarian_mode]
|
||||
label = "역발상 모드"
|
||||
description = "주류 컨센서스에 반하는 예측을 적극적으로 탐색하고 제시"
|
||||
@@ -184,6 +184,591 @@ PREDICTION: [Specific, falsifiable claim]
|
||||
|
||||
---
|
||||
|
||||
## Worked Examples
|
||||
|
||||
### Example 1: Corporate Acquisition
|
||||
|
||||
**Question**: "Will Acme Corp be acquired within 12 months?" (asked January 2025)
|
||||
|
||||
```
|
||||
PREDICTION: Acme Corp (mid-cap SaaS, $2B market cap) will be acquired by January 2026
|
||||
|
||||
1. REFERENCE CLASS (Outside View)
|
||||
Base rate: ~5-7% of publicly traded mid-cap SaaS companies receive
|
||||
acquisition offers in any given 12-month period.
|
||||
Reference examples:
|
||||
- Splunk acquired by Cisco (2023) — similar scale, strategic buyer
|
||||
- Figma attempted acquisition by Adobe (2022) — regulatory block
|
||||
- Nuance acquired by Microsoft (2021) — vertical SaaS, strategic fit
|
||||
- Mandiant acquired by Google (2022) — security vertical
|
||||
- Cvent acquired by Blackstone (2021) — PE buyout at depressed valuation
|
||||
|
||||
Starting probability: 6%
|
||||
|
||||
2. SPECIFIC EVIDENCE (Inside View)
|
||||
Signals FOR (+):
|
||||
a. Board hired Goldman Sachs as advisor (leaked filing)
|
||||
— strength: STRONG — adjustment: +20%
|
||||
(Companies that retain M&A advisors complete a transaction ~40% of the time)
|
||||
b. CEO sold 30% of personal holdings in Q4 (SEC filing)
|
||||
— strength: MODERATE — adjustment: +5%
|
||||
c. Two major competitors acquired in past 18 months (market consolidation)
|
||||
— strength: MODERATE — adjustment: +8%
|
||||
d. Revenue growth decelerated from 35% to 18% YoY (earnings report)
|
||||
— strength: MODERATE — adjustment: +5%
|
||||
(Slower-growth companies more likely to accept acquisition offers)
|
||||
|
||||
Signals AGAINST (-):
|
||||
a. Founder still holds 25% voting control and has said "we're building for
|
||||
the long term" (recent interview)
|
||||
— strength: STRONG — adjustment: -12%
|
||||
b. Stock price at all-time high — acquirer must pay steep premium
|
||||
— strength: MODERATE — adjustment: -5%
|
||||
c. Current antitrust environment — FTC blocking more deals
|
||||
— strength: WEAK — adjustment: -3%
|
||||
|
||||
3. SYNTHESIS
|
||||
Starting probability (base rate): 6%
|
||||
Signals for: +20% +5% +8% +5% = +38%
|
||||
Signals against: -12% -5% -3% = -20%
|
||||
Net adjustment: +18%
|
||||
Raw probability: 24%
|
||||
|
||||
Sanity check: ~1 in 4 feels right given the strong M&A advisor signal
|
||||
balanced against founder control.
|
||||
Final probability: 25%
|
||||
|
||||
4. KEY ASSUMPTIONS
|
||||
- Goldman engagement is for M&A (not debt restructuring):
|
||||
If wrong, probability drops to 8%
|
||||
- Founder is willing to sell at the right price:
|
||||
If wrong (founder vetoes any deal), probability drops to 3%
|
||||
- Regulatory environment doesn't tighten further:
|
||||
If wrong, probability drops to 18%
|
||||
|
||||
5. RESOLUTION
|
||||
Date: January 31, 2026
|
||||
Criteria: Definitive merger agreement announced (not just rumors)
|
||||
Data source: SEC EDGAR (8-K filing), Bloomberg terminal
|
||||
```
|
||||
|
||||
### Example 2: Technology Adoption
|
||||
|
||||
**Question**: "Will WebAssembly (Wasm) reach mainstream server-side adoption by 2027?"
|
||||
|
||||
```
|
||||
PREDICTION: >20% of new cloud-deployed services will use Wasm runtimes by
|
||||
end of 2027
|
||||
|
||||
1. REFERENCE CLASS (Outside View)
|
||||
Technology adoption lifecycle (Rogers curve):
|
||||
- Innovators (2.5%) → Early Adopters (13.5%) → Early Majority (34%)
|
||||
- Crossing from Early Adopters to Early Majority typically takes 3-5 years
|
||||
after first production deployments
|
||||
- First serious server-side Wasm deployments: ~2022 (Fermyon, Cosmonic)
|
||||
- Current status (2025): Late Early Adopter stage
|
||||
|
||||
Historical analogues for infrastructure tech adoption:
|
||||
- Containers (Docker 2013 → mainstream 2017-2018): ~4-5 years
|
||||
- Kubernetes (2014 → mainstream 2018-2019): ~4-5 years
|
||||
- Serverless (Lambda 2014 → mainstream 2018-2020): ~4-6 years
|
||||
|
||||
Base rate for "infrastructure tech reaching 20% adoption within 5 years
|
||||
of first production use": ~30%
|
||||
|
||||
Starting probability: 30%
|
||||
|
||||
2. SPECIFIC EVIDENCE (Inside View)
|
||||
Signals FOR (+):
|
||||
a. WASI standard maturing — WASI Preview 2 shipped, component model
|
||||
stabilizing (W3C working group)
|
||||
— strength: STRONG — adjustment: +8%
|
||||
b. Major cloud providers offering Wasm runtimes (Fastly, Cloudflare Workers,
|
||||
Azure, AWS exploring)
|
||||
— strength: STRONG — adjustment: +10%
|
||||
c. Docker adding Wasm support natively (announced 2022, shipping)
|
||||
— strength: MODERATE — adjustment: +5%
|
||||
|
||||
Signals AGAINST (-):
|
||||
a. Ecosystem still fragmented — multiple competing runtimes, toolchain gaps
|
||||
— strength: STRONG — adjustment: -10%
|
||||
b. Containers already "good enough" for most workloads — weak forcing
|
||||
function to switch
|
||||
— strength: STRONG — adjustment: -8%
|
||||
c. Wasm language support uneven — great for Rust/C++, mediocre for Python/JS
|
||||
— strength: MODERATE — adjustment: -5%
|
||||
|
||||
Leading indicators to track:
|
||||
- CNCF survey: % of respondents evaluating/using Wasm
|
||||
- Job postings mentioning Wasm (Indeed/LinkedIn trend)
|
||||
- GitHub stars and contributors for top Wasm runtimes (wasmtime, wasmer)
|
||||
- WASI spec milestone dates vs planned dates
|
||||
|
||||
3. SYNTHESIS
|
||||
Starting probability (base rate): 30%
|
||||
Signals for: +8% +10% +5% = +23%
|
||||
Signals against: -10% -8% -5% = -23%
|
||||
Net adjustment: 0%
|
||||
Final probability: 30%
|
||||
|
||||
Interpretation: The positive and negative signals roughly cancel out.
|
||||
The base rate from analogous infrastructure technologies holds.
|
||||
This is genuinely uncertain — the "chasm" crossing is the key risk.
|
||||
|
||||
4. KEY ASSUMPTIONS
|
||||
- "Mainstream" defined as >20% of NEW deployments (not total installed base)
|
||||
- WASI component model reaches 1.0 stable by mid-2026
|
||||
If delayed beyond 2026: probability drops to 15%
|
||||
- No competing paradigm emerges (e.g., eBPF expanding scope):
|
||||
If strong competitor: probability drops to 20%
|
||||
|
||||
5. RESOLUTION
|
||||
Date: December 31, 2027
|
||||
Criteria: CNCF annual survey shows >20% respondents using Wasm in production
|
||||
Data source: CNCF Annual Survey, Datadog Container Report
|
||||
```
|
||||
|
||||
### Example 3: Geopolitical Forecast
|
||||
|
||||
**Question**: "Will US-China trade tensions escalate significantly in 2025?"
|
||||
(Defined as: new tariffs >25% on >$100B of goods, or export controls expanded
|
||||
to 3+ new technology categories)
|
||||
|
||||
```
|
||||
PREDICTION: Significant escalation of US-China trade tensions in 2025
|
||||
|
||||
1. REFERENCE CLASS (Outside View)
|
||||
Historical trade conflict escalation pattern:
|
||||
- US-China trade relations since 2018: escalation occurred in 4 of 7 years
|
||||
- In election year +1 (new/returning administration): escalation rate ~60%
|
||||
- Trade wars historically escalate in steps, with retaliation cycles
|
||||
|
||||
Starting probability: 55%
|
||||
|
||||
2. SPECIFIC EVIDENCE (Inside View)
|
||||
Signals FOR (+):
|
||||
a. Administration rhetoric on China hawkish across both parties
|
||||
— strength: STRONG — adjustment: +10%
|
||||
b. Semiconductor export controls already expanding (ASML, Tokyo Electron)
|
||||
— strength: STRONG — adjustment: +8%
|
||||
c. China retaliating with rare earth export restrictions
|
||||
— strength: MODERATE — adjustment: +5%
|
||||
|
||||
Signals AGAINST (-):
|
||||
a. Business lobbying against further tariffs (Chamber of Commerce, farm lobby)
|
||||
— strength: MODERATE — adjustment: -5%
|
||||
b. Inflation concerns create political cost for tariffs
|
||||
— strength: MODERATE — adjustment: -5%
|
||||
c. Diplomatic channels active (recent bilateral meetings)
|
||||
— strength: WEAK — adjustment: -3%
|
||||
|
||||
Scenario mapping:
|
||||
┌─────────────────────────┬─────────────┬────────────────────┐
|
||||
│ Scenario │ Probability │ Key trigger │
|
||||
├─────────────────────────┼─────────────┼────────────────────┤
|
||||
│ Major escalation │ 25% │ Taiwan crisis or │
|
||||
│ (new tariffs + controls │ │ tech IP theft case │
|
||||
│ + retaliatory cycle) │ │ │
|
||||
├─────────────────────────┼─────────────┼────────────────────┤
|
||||
│ Moderate escalation │ 40% │ Incremental tariff │
|
||||
│ (meets our threshold) │ │ increases + 1-2 │
|
||||
│ │ │ new export controls│
|
||||
├─────────────────────────┼─────────────┼────────────────────┤
|
||||
│ Status quo / minor │ 30% │ Diplomatic deals, │
|
||||
│ changes │ │ election distraction│
|
||||
├─────────────────────────┼─────────────┼────────────────────┤
|
||||
│ De-escalation │ 5% │ Grand bargain │
|
||||
│ (reduced tariffs) │ │ (historically rare)│
|
||||
└─────────────────────────┴─────────────┴────────────────────┘
|
||||
|
||||
P(meets our escalation threshold) = 25% + 40% = 65%
|
||||
|
||||
3. SYNTHESIS
|
||||
Starting probability (base rate): 55%
|
||||
Signals for: +10% +8% +5% = +23%
|
||||
Signals against: -5% -5% -3% = -13%
|
||||
Net adjustment: +10%
|
||||
Raw probability: 65%
|
||||
|
||||
Cross-check with scenario mapping: 65% — consistent.
|
||||
Final probability: 65%
|
||||
|
||||
4. KEY ASSUMPTIONS
|
||||
- No major geopolitical crisis (Taiwan strait) that causes extreme
|
||||
escalation or extreme restraint: If crisis occurs, split to
|
||||
80% (escalation) or 20% (restraint/avoidance)
|
||||
- US economy remains stable: If recession hits, probability drops
|
||||
to 45% (political cost of tariffs rises)
|
||||
- China does not make major trade concessions preemptively:
|
||||
If it does, probability drops to 30%
|
||||
|
||||
5. RESOLUTION
|
||||
Date: December 31, 2025
|
||||
Criteria: Cumulative new tariffs >25% on >$100B goods OR export controls
|
||||
expanded to 3+ new technology categories (per USTR/BIS announcements)
|
||||
Data source: USTR tariff schedule, BIS Entity List updates, Congressional
|
||||
Research Service reports
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Fermi Estimation Techniques
|
||||
|
||||
Fermi estimation is the art of making reasonable order-of-magnitude guesses
|
||||
by breaking unknowable questions into smaller, estimable pieces.
|
||||
|
||||
### Step-by-Step Process
|
||||
|
||||
```
|
||||
1. DEFINE the quantity you want to estimate
|
||||
→ Be specific about units, scope, and timeframe
|
||||
|
||||
2. DECOMPOSE into factors you can estimate independently
|
||||
→ Prefer multiplication chains: A × B × C
|
||||
→ Each factor should be something you can reason about
|
||||
|
||||
3. ESTIMATE each factor
|
||||
→ Use round numbers (powers of 10 when possible)
|
||||
→ State your confidence range for each factor
|
||||
|
||||
4. MULTIPLY and sanity-check
|
||||
→ Does the result pass the "smell test"?
|
||||
→ Cross-check with any known anchors
|
||||
|
||||
5. STATE your uncertainty
|
||||
→ Fermi estimates are typically accurate within 1 order of magnitude
|
||||
→ Give a range: [estimate / 3, estimate × 3] is a reasonable default
|
||||
```
|
||||
|
||||
### Common Reference Anchors
|
||||
|
||||
Keep these memorized for quick estimation:
|
||||
|
||||
```
|
||||
POPULATION
|
||||
World: ~8 billion
|
||||
US: ~340 million
|
||||
EU: ~450 million
|
||||
China: ~1.4 billion
|
||||
India: ~1.4 billion
|
||||
|
||||
ECONOMICS
|
||||
World GDP: ~$100 trillion
|
||||
US GDP: ~$28 trillion
|
||||
US median household: ~$75,000/year
|
||||
US federal budget: ~$6.5 trillion
|
||||
S&P 500 total cap: ~$45 trillion
|
||||
|
||||
TIME
|
||||
Seconds in a day: ~86,400 (~10^5)
|
||||
Seconds in a year: ~31.5 million (~3 × 10^7)
|
||||
Working hours/year: ~2,000
|
||||
|
||||
TECHNOLOGY
|
||||
Global internet users: ~5.5 billion
|
||||
Global smartphone users: ~4.5 billion
|
||||
AWS annual revenue: ~$90 billion
|
||||
Global IT spending: ~$5 trillion
|
||||
GitHub developers: ~100 million
|
||||
|
||||
INDUSTRY SIZES (annual, global)
|
||||
Cloud computing: ~$600 billion
|
||||
Semiconductor: ~$600 billion
|
||||
Pharmaceutical: ~$1.5 trillion
|
||||
Automotive: ~$3 trillion
|
||||
Agriculture: ~$3 trillion
|
||||
E-commerce: ~$6 trillion
|
||||
```
|
||||
|
||||
### Worked Fermi Examples
|
||||
|
||||
**Example A: Estimating the TAM for an AI code review tool**
|
||||
|
||||
```
|
||||
Question: What is the annual TAM for an AI-powered code review SaaS?
|
||||
|
||||
Decomposition:
|
||||
TAM = (Number of professional developers)
|
||||
× (% who do code reviews regularly)
|
||||
× (willingness to pay for tooling)
|
||||
× (average annual price)
|
||||
|
||||
Estimates:
|
||||
Professional developers worldwide: ~30 million
|
||||
(GitHub has 100M accounts, but ~30% are professional, and
|
||||
not all professionals use GitHub)
|
||||
|
||||
% who do code reviews: ~60%
|
||||
(Standard in companies > 50 engineers, less common in small shops)
|
||||
|
||||
Target market (teams that would buy SaaS): ~40%
|
||||
(Enterprise and mid-market; small teams use free tools)
|
||||
|
||||
Annual price per seat: ~$300/year
|
||||
(Comparable: GitHub Copilot ~$200, Snyk ~$400, middle ground)
|
||||
|
||||
Calculation:
|
||||
30M × 0.60 × 0.40 × $300 = $2.16 billion
|
||||
|
||||
Sanity check:
|
||||
- GitHub revenue ~$2B (broader product, ~4M paid users)
|
||||
- Snyk valued at $7B (code security, related space)
|
||||
- $2B TAM is plausible for a focused code review tool
|
||||
|
||||
Result: ~$2 billion TAM (range: $700M to $6B)
|
||||
```
|
||||
|
||||
**Example B: Estimating daily active queries to a search engine**
|
||||
|
||||
```
|
||||
Question: How many search queries does Google process per day?
|
||||
|
||||
Decomposition:
|
||||
Queries/day = (Internet users who use Google)
|
||||
× (searches per user per day)
|
||||
|
||||
Estimates:
|
||||
Global internet users: ~5.5 billion
|
||||
Google market share: ~90%
|
||||
Google users: 5.5B × 0.90 = ~5 billion
|
||||
But not all use it daily: ~50% daily active rate
|
||||
Daily active Google searchers: ~2.5 billion
|
||||
|
||||
Searches per active user per day: ~3-4
|
||||
(Some people search 10+ times, many search once or not at all)
|
||||
|
||||
Calculation:
|
||||
2.5 billion × 3.5 = ~8.5 billion queries/day
|
||||
|
||||
Sanity check:
|
||||
Published figure (Google): ~8.5 billion searches/day (2024)
|
||||
Our estimate nailed it — sometimes Fermi estimation gets lucky.
|
||||
|
||||
Result: ~8.5 billion/day (range: 3B to 25B)
|
||||
```
|
||||
|
||||
### Order of Magnitude Sanity Checks
|
||||
|
||||
After any estimate, verify it makes sense:
|
||||
|
||||
```
|
||||
CHECK 1: Per-person reasonableness
|
||||
Divide by relevant population. Is the per-person number realistic?
|
||||
"$50B market ÷ 340M Americans = $147/person" — plausible?
|
||||
|
||||
CHECK 2: Comparison to known quantities
|
||||
Is your estimate bigger or smaller than things you know?
|
||||
"Our estimate of X is $3B — that's 5% of AWS revenue. Reasonable?"
|
||||
|
||||
CHECK 3: Growth rate implied
|
||||
If you're estimating a future state, what annual growth rate is implied?
|
||||
>50% sustained growth for >3 years is extremely rare.
|
||||
|
||||
CHECK 4: Upper bound test
|
||||
What is the theoretical maximum? Is your estimate within it?
|
||||
"Total possible customers × maximum price = ceiling"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Prediction Market Patterns
|
||||
|
||||
### Interpreting Market Prices as Probabilities
|
||||
|
||||
Prediction market prices map to probabilities, but with important caveats:
|
||||
|
||||
```
|
||||
Market price $0.65 for "Event X occurs"
|
||||
→ Naive interpretation: 65% probability
|
||||
→ Adjusted interpretation: depends on market quality
|
||||
|
||||
Adjustment factors:
|
||||
Liquid market (Polymarket, Metaculus with many forecasters):
|
||||
Price ≈ true probability (±3-5%)
|
||||
|
||||
Thin market (<50 traders, <$10K volume):
|
||||
Price is noisy — treat as ±15% uncertainty
|
||||
A $0.65 price could represent 50-80% true probability
|
||||
|
||||
Binary vs. multi-outcome:
|
||||
Binary markets are more reliable
|
||||
Multi-outcome markets often have probabilities summing to >100%
|
||||
(overround) — normalize before interpreting
|
||||
```
|
||||
|
||||
### Common Prediction Market Biases
|
||||
|
||||
| Bias | Description | Impact | Correction |
|
||||
|------|-------------|--------|------------|
|
||||
| Favorite-longshot | Favorites underpriced, longshots overpriced | Longshot events appear ~2-3x more likely than they are | If market says 5%, true probability may be 2-3% |
|
||||
| Recency | Recent events dominate pricing | Probability spikes after news, then slowly reverts | Wait 24-48h after major news before trusting market prices |
|
||||
| Liquidity premium | Illiquid contracts trade at a discount | Prices biased toward 50% in thin markets | Weight liquid markets more heavily |
|
||||
| Expiration clustering | Prices converge to 0 or 1 near expiration | Mid-probability contracts vanish near deadline | Most useful signal is months before resolution |
|
||||
| Hedging distortion | Traders hedging other positions, not expressing beliefs | Prices reflect risk management, not pure probability | Cross-reference with non-market forecasts |
|
||||
|
||||
### Aggregation Methods
|
||||
|
||||
When combining multiple probability estimates (markets, experts, models):
|
||||
|
||||
```
|
||||
SIMPLE AVERAGE
|
||||
P = (P1 + P2 + P3) / 3
|
||||
Use when: Sources are roughly equally credible
|
||||
Weakness: Susceptible to outliers
|
||||
|
||||
MEDIAN
|
||||
P = middle value of sorted estimates
|
||||
Use when: One source might be badly miscalibrated
|
||||
Weakness: Ignores magnitude of disagreement
|
||||
|
||||
TRIMMED MEAN
|
||||
Drop highest and lowest, average the rest
|
||||
P = average(P2 ... Pn-1) after sorting
|
||||
Use when: 5+ sources, want outlier robustness
|
||||
|
||||
EXTREMIZED AVERAGE
|
||||
P_avg = simple average
|
||||
P_extremized = P_avg^a / (P_avg^a + (1-P_avg)^a), where a > 1
|
||||
Typical a = 1.5 to 2.5 (more extremizing with more independent sources)
|
||||
Use when: Sources are genuinely independent (not reading each other)
|
||||
Rationale: If 5 independent sources all say 70%, the true probability
|
||||
is likely higher than 70% — shared info should push further from 50%
|
||||
|
||||
CONFIDENCE-WEIGHTED AVERAGE
|
||||
P = Σ(wi × Pi) / Σ(wi)
|
||||
where wi = track record score or source reliability
|
||||
Use when: Sources have known, differing track records
|
||||
```
|
||||
|
||||
### When Markets Beat Experts (and Vice Versa)
|
||||
|
||||
```
|
||||
MARKETS TEND TO WIN when:
|
||||
✓ Large, liquid, diverse participant pool
|
||||
✓ Question is well-defined with clear resolution criteria
|
||||
✓ Information is widely distributed (no single expert has edge)
|
||||
✓ Time horizon is 1 month to 2 years
|
||||
Examples: Election outcomes, product launch dates, economic indicators
|
||||
|
||||
EXPERTS TEND TO WIN when:
|
||||
✓ Question requires deep domain-specific knowledge
|
||||
✓ Market is thin or participants lack domain context
|
||||
✓ Very long time horizons (>5 years) — markets discount distant futures
|
||||
✓ Novel situations with no historical market precedent
|
||||
Examples: Technical feasibility, scientific breakthroughs, niche regulation
|
||||
|
||||
BEST PRACTICE: Use both
|
||||
Start with the market price, then adjust using expert insight.
|
||||
Treat the market as the prior and expert analysis as an update.
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Update Protocol
|
||||
|
||||
### Bayesian Updating Worked Example
|
||||
|
||||
```
|
||||
SCENARIO: You predicted 30% chance that Company Z launches Product A in Q1.
|
||||
New evidence: A leaked internal slide shows a Q1 launch timeline.
|
||||
|
||||
STEP 1: State the prior
|
||||
P(launch in Q1) = 0.30
|
||||
|
||||
STEP 2: Assess the evidence
|
||||
E = leaked slide showing Q1 timeline
|
||||
How likely is this evidence if the launch IS happening in Q1?
|
||||
P(E | launch) = 0.85
|
||||
(Internal slides usually reflect real plans, but plans change)
|
||||
How likely is this evidence if the launch is NOT in Q1?
|
||||
P(E | no launch) = 0.15
|
||||
(Could be outdated slide, aspirational, or decoy)
|
||||
|
||||
STEP 3: Calculate the likelihood ratio
|
||||
LR = P(E | launch) / P(E | no launch) = 0.85 / 0.15 = 5.67
|
||||
|
||||
STEP 4: Convert prior to odds, multiply, convert back
|
||||
Prior odds = 0.30 / 0.70 = 0.429
|
||||
Posterior odds = 0.429 × 5.67 = 2.43
|
||||
Posterior probability = 2.43 / (1 + 2.43) = 0.71
|
||||
|
||||
STEP 5: State the update
|
||||
Prior: 30% → Posterior: 71%
|
||||
Update magnitude: +41 percentage points
|
||||
This is a LARGE update, appropriate because the evidence (internal
|
||||
planning document) is strong and directly relevant.
|
||||
```
|
||||
|
||||
### Evidence Strength Classification
|
||||
|
||||
How much to update based on different types of evidence:
|
||||
|
||||
```
|
||||
EVIDENCE TIER 1 — Large update (likelihood ratio 5-20x)
|
||||
→ Official announcement or regulatory filing
|
||||
→ Confirmed internal document (not rumor)
|
||||
→ Directly observed outcome of prerequisite event
|
||||
→ Multiple independent strong sources confirming same fact
|
||||
Typical update: ±15-30 percentage points
|
||||
|
||||
EVIDENCE TIER 2 — Moderate update (likelihood ratio 2-5x)
|
||||
→ Credible journalist report with named sources
|
||||
→ Statistical data that changes the base rate
|
||||
→ Expert with strong track record changing their view
|
||||
→ Structural/policy change that alters incentives
|
||||
Typical update: ±5-15 percentage points
|
||||
|
||||
EVIDENCE TIER 3 — Small update (likelihood ratio 1.2-2x)
|
||||
→ Rumor from semi-credible source
|
||||
→ Anecdotal evidence (single data point)
|
||||
→ Social media sentiment shift
|
||||
→ Expert opinion without new information
|
||||
Typical update: ±2-5 percentage points
|
||||
|
||||
EVIDENCE TIER 4 — Negligible update (likelihood ratio ~1x)
|
||||
→ Repetition of previously known information
|
||||
→ Pundit opinion with no domain expertise
|
||||
→ Vague statement open to multiple interpretations
|
||||
→ Evidence equally consistent with both outcomes
|
||||
Typical update: ±0-2 percentage points (or skip entirely)
|
||||
```
|
||||
|
||||
### When to Make Large vs. Small Updates
|
||||
|
||||
```
|
||||
MAKE A LARGE UPDATE when:
|
||||
• Evidence directly addresses your key uncertainty
|
||||
• The source has a strong track record on this topic
|
||||
• The evidence would be very surprising if your prediction were correct
|
||||
(or very unsurprising if it were wrong)
|
||||
• Multiple independent signals shift in the same direction simultaneously
|
||||
|
||||
MAKE A SMALL UPDATE when:
|
||||
• Evidence is tangentially related to your prediction
|
||||
• The source's reliability is uncertain
|
||||
• The evidence is consistent with multiple interpretations
|
||||
• You've already incorporated similar evidence
|
||||
|
||||
RESIST UPDATING when:
|
||||
• The "evidence" is just someone restating the consensus
|
||||
• A vivid anecdote feels compelling but carries no statistical weight
|
||||
• You're reacting emotionally (fear, excitement) rather than analytically
|
||||
• The evidence source has an obvious incentive to mislead
|
||||
```
|
||||
|
||||
### Common Updating Mistakes
|
||||
|
||||
| Mistake | Description | Fix |
|
||||
|---------|-------------|-----|
|
||||
| Over-updating on vivid events | A dramatic single event shifts your view by 20+ points when the base rate barely moved | Ask: "Does this event actually change the base rate, or just my emotional state?" |
|
||||
| Under-updating on base rate changes | New data shows the reference class frequency shifted, but you keep your old anchor | Periodically re-derive the base rate from scratch instead of only adjusting incrementally |
|
||||
| Asymmetric updating | Updating strongly on confirming evidence, weakly on disconfirming evidence | Force yourself to calculate the likelihood ratio for disconfirming evidence explicitly |
|
||||
| Double-counting | Updating on a news article, then updating again on a tweet quoting the same article | Track the original source — if two signals share the same root cause, count once |
|
||||
| Failure to update | Knowing the evidence should change your view but not bothering because your current number "feels right" | Set calendar reminders to review active predictions monthly with fresh evidence |
|
||||
| Stampede updating | A prediction market spikes, causing you to rush your update to match | Market moves are data, not commands — assess independently, then compare |
|
||||
|
||||
---
|
||||
|
||||
## Prediction Tracking & Scoring
|
||||
|
||||
### Prediction Ledger Format
|
||||
|
||||
Reference in new issue
Block a user