Regret Minimization
Overview
Most frameworks optimize expected value — pick the highest probability-weighted payoff. That math breaks down for large, asymmetric, life-defining choices where the true unit is not money: career pivots, founding decisions, relocations, relationship exits. There, the decisive question is not "what's the expected payoff?" but "which regret will I be unable to live with at 80?"
Associated with Jeff Bezos (D.E. Shaw → Amazon, 1994); rooted in Stoic Premeditatio Malorum (Seneca, Epistulae Morales 91) and formalized in Regret Theory (Loomes & Sugden, The Economic Journal, 1982).
Compose with neighbors: use first-principles to clarify what is at stake; inversion to surface failure modes; second-order-thinking to verify downstream consequences; then use regret minimization to choose when EV analyses come out close and the true cost is psychological.
When to Use
Apply when: decision is major, hard-to-reverse, asymmetric (career pivot, founding, relocation, relationship, children); EV math feels insufficient because units are joy/meaning/identity; hesitation is emotional, not analytical; user says "what would my 80-year-old self think?", "if I never try this will I regret it?", "I keep hesitating but the math is clear," "should I quit big tech to go all-in on AI / join the AI wave / start a company now (bubble or export-control fears notwithstanding)?"
When NOT to use: routine reversible decisions; EV is genuinely the right unit (portfolio allocation, pricing); framework being re-run weekly (procrastination); both regrets are unlivable (redesign the choice instead).
Coaching Novices (Adaptive Front Door)
- Engine mode: user has a concrete major decision → run The Process directly.
- Coach mode: user is unfamiliar or has no concrete case → guide step by step.
In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.
- One-line what-it-is: for big life decisions where math is close or doesn't fit, project to age 80 and ask which path will haunt you more. Commit to the option whose regret you can live with.
- Check fit against When to Use / When NOT to use — if it doesn't fit, say so and point elsewhere.
- Elicit their real decision. Ask for one if missing; never run the audit on a hypothetical.
[WAIT — do not advance until user responds]
- Walk The Process one step per turn: describe 80-year-old self → name each path's regret → test asymmetry. Never project for them.
[WAIT — do not advance until user responds]
- Close by naming the regret they identified as harder to live with, plus the commit date.
[WAIT — do not advance until user responds]
The Process
Run the Regret Audit. Project, name the regret on each path, test asymmetry, commit.
- Define the fork specifically. Not "should I change my life" — "leave job to start X by date vs stay 2 years and revisit." Vague forks produce vague regrets.
- Project to 80. Describe yourself looking back: what life, what values, who's around. The age-80 self is your evaluator function.
- For each option, imagine the path lived realistically. What did you do, learn, not do? Avoid heroic versions.
- Name the regret on each path — split failed-but-tried from never-tried. Failed regret is usually lighter; test this for your case.
- Test asymmetry — which regret is harder to live with? Load-bearing step. Most people know the answer once they ask the right question.
- Name the unit being optimized: joy, meaning, security, identity, contribution, autonomy. Step 5's answer should match the unit you actually weight most.
- Commit and pre-bind. State the date the decision is irreversible — after which you stop relitigating.
Output: the Regret Audit
Decision: Option A vs Option B (concrete, time-bounded)
Age-80 self: <one paragraph — what life, what values looking back>
Path A lived (realistic): <what I did / learned / didn't do>
Path B lived (realistic): <what I did / learned / didn't do>
Regrets — A failed: "..." | A never tried: "..."
Regrets — B failed: "..." | B never tried: "..."
Asymmetry: which regret is harder to live with at 80, and why?
Unit optimized: <joy / meaning / security / identity / autonomy>
Commitment: I commit to <option> by <date>. Stop relitigating after that.
Re-open trigger: only if <specific observable evidence> before <date>.
→ Method in Action: Jeff Bezos, D.E. Shaw → Amazon (1994) · Gilovich & Medvec, the Evidence for the Age-80 Test (1995) → 2026 lens: The "go all-in on AI" career / founder decision (2023–2026) — quit big tech for the AI wave, or stay?
Regret Packs
The Audit runs identically across domains; what differs is the typical load-bearing regret and the unit optimized.
- Career pivots / founding: load-bearing = never having tried; discount "I'll have less money than peers" (vivid at 35, faint at 80).
- Family formation: structurally asymmetric; discount "slower career progression for a decade."
- Geographic relocation: load-bearing = staying somewhere that quietly drained you; discount "I'll miss old friends."
- Sunk investments (PhD, long relationship, failing startup): load-bearing = staying past the point you knew; discount sunk-cost framing.
Applying It Well
- Project, don't fantasize. Imagine the ordinary version of each path — successful-A vs worst-B corrupts the audit.
- Failed-but-tried regret is usually lighter than never-tried — but verify for your case. Often true, not universally true.
- The unit matters more than the math. Most failed audits asked the right question against the wrong unit.
- Pre-commit or it didn't happen. Re-running weekly is postponement, not decision-making.
→ Primary sources: references/sources.md
Common Rationalizations
[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.
| Fake move | Reality |
|---|---|
| [D] Using regret minimization for trivial decisions | Theater, not analysis. The framework is for large, asymmetric, hard-to-reverse decisions. |
| [D] Asymmetric application | Running regret minimization for the option you want; running EV for the option you don't. Both options must use the same audit in the same units. |
| [D] Treating "regret" as emotional discomfort | Real regret haunts you at 80 — not passing disappointment. Mild annoyance = wrong tool. |
| [D] Conflating never-tried with tried-and-failed regret | Different magnitudes. The whole framework rests on telling them apart. |
| [D] Mistaking it for EV optimization | Different objective functions that often disagree. Picking regret because EV gives an unwelcome answer is rationalization. |
| [D] Importing someone else's 80-year-old self | Your future self is not Jeff Bezos's. Build your own age-80 evaluator first. |
| [D] Failing to pre-commit | Getting the answer then re-litigating weekly. The point is to commit and stop the inner argument. |
| [D] Fantasy projection | Imagining the successful version of A and the worst version of B. Both must be imagined at realistic outcomes. |
| → Add [O] entries here after each real use — paste the actual failure pattern | What went wrong and why |
Red Flags
- Decision is reversible or low-stakes; options not specified with concrete actions and dates
- One option imagined heroically, the other realistically
- "Regret" weighed is present-day annoyance, not haunting-at-80 regret
- Audit re-run multiple times with no commitment date; framework used to justify a decision already made
- No unit named — "asymmetry" comparison is across categories
- Age-80 self borrowed from someone famous, not the user's own values
Verification
- Specific, time-bounded fork defined; age-80 self described with values explicit
- Each path imagined realistically (including the boring middle)
- Both failed-but-tried and never-tried regrets named for each option
- Asymmetry test answered; unit optimized named; commitment date set
Part of deciqAI Knowledge Skills — 227 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/c/regret-minimization · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.
Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/regret-minimization.json