Optimal Stopping Theory for the AI Era
최적 정지 이론 - AI 시대의 의사결정
The 37% rule says explore 37%, then commit. But in AI era, N becomes infinite. How do you know when to stop when infinite options exist? The framework that applies to hiring, investing, and life decisions.
Also available on 애당초 4개의 시선 (Ethan Cho: Four Lenses on Everything) on Substack.
Read on Substack →Optimal Stopping Theory for the AI Era
The classic secretary problem: interview N candidates, reject the first 37%, then hire the next one better than all previous. Mathematically optimal.
But AI Changes Everything
In AI era, N → ∞. You can always: - Find one more candidate (LinkedIn AI recruiter) - Analyze one more startup (AI deal flow) - Research one more market (infinite data)
The paradox: When you can always explore more, when do you stop?
The New Framework: Artificial Constraints
Since N is infinite, you must artificially cap it:
1. Set your N - "I'll look at exactly 20 candidates" 2. Apply 37% rule - Reject first 7 3. Commit to next best - Don't second-guess
Applications
Hiring: Don't interview forever. Set N=20, apply the rule.
Investing: Define your pipeline (N=50 startups/year), explore 37%, commit to next exceptional.
Life decisions: Moving cities, choosing partners, big purchases - set your N, stop when you should.
The Discipline
AI removes natural constraints. The skill is imposing them yourself. Know when to stop exploring and start executing.
[Read full article on Substack →](https://ethancho12.substack.com/p/optimal-stopping-theory)
🔑Key Takeaways
- ✓37% rule: Explore first 37% of N candidates, then commit to next one better than all previous
- ✓AI era problem: N → ∞ (infinite candidates via LinkedIn AI, infinite data, infinite options)
- ✓Solution: Artificially cap N before applying 37% rule - discipline over abundance
- ✓Applications: Hiring (N=20), investing (N=50 startups/year), life decisions (moving, partners)
- ✓Meta-skill: Imposing constraints when technology removes natural limits
📋How to Apply This Framework
Define Your N Upfront (Artificial Cap)
Before starting, commit to a specific number: Hiring? N=20 candidates max. Investing? N=50 startups/quarter. Moving cities? N=5 cities. DON'T base N on 'how many exist' (infinite in AI era). Base it on: (1) Time you can dedicate, (2) Decision urgency, (3) Diminishing returns point. Write it down BEFORE you start exploring. Example: 'I will interview exactly 20 candidates for this role, no more.'
Calculate Your 37% Threshold (Exploration Phase)
Math: 0.37 × N = exploration count. Examples: N=20 candidates → Explore first 7, commit after. N=50 startups → Explore first 18, commit after. N=5 cities → Explore first 1-2, commit after. During exploration phase: Take notes, score each candidate, but REJECT all of them. This builds your calibration baseline. Resist temptation to 'just pick one' early—exploration phase is for learning, not deciding.
Build Your Evaluation Rubric During Exploration
As you explore first 37%, track what matters: (1) Create scoring criteria (5-10 factors), (2) Note your best candidate so far (benchmark), (3) Identify deal-breakers (red flags), (4) Calibrate your expectations (realistic vs ideal). Example for hiring: After interviewing 7 candidates, you know 'best so far' scored 8/10 on technical, 6/10 on culture fit. That's your new benchmark.
Commit to Next Best (Commitment Phase)
After exploration phase (37%), commit to the VERY NEXT candidate who beats your benchmark. Critical rules: (1) Don't wait for 'perfect'—just 'better than all previous', (2) Don't go back to exploration (sunk cost fallacy), (3) Don't extend your N (defeats the purpose), (4) Trust the math—it's optimal. Example: Candidate #8 scores 8.5/10 technical, 7/10 culture fit → Beats benchmark (8/6) → HIRE IMMEDIATELY.
Review and Adjust N for Next Time
After each decision cycle, reflect: (1) Was N too small? (Committed too early, low confidence), (2) Was N too large? (Decision paralysis, diminishing returns), (3) Did exploration phase calibrate you well? Adjust for next time. Example: After 3 hiring cycles with N=20, you notice best candidates always appear in slots 8-12 → Maybe reduce N=15 next time (faster, same quality).