View methodology & research sources
AI Productivity Multiplier (PM) Scoring Methodology
═══════════════════════════════════════════════
CORE FORMULA: Geometric-mean interaction + sigmoid compression
Step 1: Normalize inputs
D = AI Exposure / 10 (digital task fraction, from Karpathy's LLM scoring)
C = weighted composite of:
- Education requirement (35%): No degree=0 → Doctoral=10
- Median pay percentile (45%): log-scale ranking across 342 occupations
- Category complexity bonus (20%): CS/Math/Legal get boost, manual labor gets reduction
Step 2: Compute base empowerment
base = geometric_mean(D, C) × 0.6 + harmonic_mean(D, C) × 0.4
- Geometric mean naturally penalizes if EITHER dimension is low
- This prevents: high exposure + low complexity → falsely high score
Step 3: Apply research-backed adjustments
- Automation penalty: if D>0.6 and C<0.35, reduce score (data entry clerk effect)
- Expert leverage bonus: if D≥0.6 and C≥0.6, boost score (Harvard "inside the frontier")
- Category adoption boost: CS +0.08, Math +0.06, Legal +0.05 (from Anthropic usage data)
Step 4: Sigmoid compression → 1–10 scale
sigmoid(adjusted, center=0.50, steepness=7.0) → maps to [1, 10]
- Top ~21% get PM 9-10 (knowledge workers in CS, finance, legal, science)
- Bottom ~28% get PM 1-2 (manual labor)
- Middle has good spread across 3-8
FOUR QUADRANTS:
Exposure ≥ 6 + PM ≥ 5.5 → "AI Power Users" (transform & thrive)
Exposure < 6 + PM ≥ 5.5 → "AI-Enhanced Specialists" (physical + AI assist)
Exposure ≥ 6 + PM < 5.5 → "Automation Risk" (routine digital → replaced)
Exposure < 6 + PM < 5.5 → "AI-Resistant" (physical, minimal AI impact)
RESEARCH FOUNDATION:
[1] Anthropic "Labor Market Impacts" (2025)
[2] Anthropic "Estimating Productivity Gains" (2025)
[3] OpenAI "GPTs are GPTs" (Eloundou et al., 2023)
[4] Harvard/BCS "Jagged Frontier" (Dell'Acqua et al., 2023)
[5] MIT/Brynjolfsson "Generative AI at Work" (2023)
[6] Goldman Sachs (2023)
[7] McKinsey "Superagency" (2025)
[8] Stanford AI Index 2025
DATA SOURCES:
- Employment, pay, education, outlook: BLS Occupational Outlook Handbook (2024-2034 projections)
- AI Exposure scores: Karpathy/jobs repo (Gemini LLM-scored via OpenRouter, 0-10 scale)
- Occupation metadata: 342 occupations from BLS OOH, scraped via Karpathy's pipeline