#39 - Spencer Greenberg on the scientific approach to solving difficult everyday questions

Sparkwave Portfolio & Personal Updates

• Spencer Greenberg manages four active portfolio companies at Sparkwave, having recruited CEOs for all and spun them out as separate entities.
  • Uplift: An automated depression care program run by Eddie Lou; currently in closed beta with plans for imminent public release.
  • Mind Ease: A software product designed to alleviate significant anxiety, managed by Peter Breitbart.
  • Clear Thinking: A tool for improving decision-making and reducing bias, led by Aurora Quinn Elmore.
  • Positively: A recruitment platform for social science research, run by Luke Freeman.
    • Currently utilizes Amazon Mechanical Turk as its first backend to layer features for researchers.
    • Plans to add new backends to secure representative national samples or specific demographic cohorts beyond Turk.
• Clear Thinking Research & Products:
  • Conducted long-term studies on "happiness habits" using environmental triggers; Greenberg reported a 5% personal happiness increase by associating a daily trigger (checking social media) with gratitude.
  • Testing a methodology where pre-existing daily triggers are used to induce specific positive thoughts (e.g., gratitude) to create lasting habits.
  • The goal is to release this as a public tool if study results validate the efficacy of the trigger-based happiness technique.

Intrinsic vs. Instrumental Values Study

• Methodology: Conducted a rigorous survey of Effective Altruists (EAs), partial EAs, and non-EAs across demographics (age, gender, politics) to identify intrinsic values.
  • Excluded participants who failed a quiz on defining "intrinsic value" (valuing something for its own sake, not its effects).
  • Participants were required to define the concept in their own words to ensure comprehension.
• Key Demographic Findings:
  • Conservatives: Report higher intrinsic valuation of religion, retribution, and the preservation of existing values.
  • Liberals: Report higher intrinsic valuation of animal well-being, nature, and the happiness of strangers.
  • Females: More frequently report intrinsic values regarding kindness, caring, diversity, and human freedom.
  • Males: More frequently report intrinsic values regarding selfish interests, in-group interests, and stranger pleasure.
  • Older Adults: Value being trusted/cared for and general societal morality.
  • Younger Adults: Value animal lifespans, personal admiration, and the pleasure of those they know.
  • Effective Altruists: Show a distinct pattern of devaluing personal/in-group interests while strongly valuing the suffering and happiness of all conscious beings.
• Universal Intrinsic Values:
  • 82% of non-EAs report "I love other people" as an intrinsic value; 71% value "beautiful things continuing to exist" even if unseen.
  • Universal values (those not tied to self or specific in-groups) serve as a potential foundation for global cooperation.
• Practical Applications of Value Identification:
  • Avoiding Value Traps: Distinguishing intrinsic from instrumental values prevents pursuing careers or goals (e.g., high money) that do not actually deliver the desired intrinsic experience (e.g., autonomy).
  • Goal Factoring: Identifying the core intrinsic value allows for more efficient planning than pursuing intermediate goals (e.g., becoming a tenured professor) that were assumed to be the end-goal.
  • Social Guilt: Understanding that different groups hold different intrinsic values reduces feelings of alienation or "wrongness" when one's values diverge from family or community expectations.
  • Preventing Doublethink: Separating subjective intrinsic values from objective moral beliefs prevents self-deception where individuals claim to value only "acceptable" things (e.g., global welfare) while ignoring their genuine self-interest.
  • Future Vision: Building a desirable future requires balancing multiple intrinsic values to ensure broad appeal, avoiding "hedonic treadmill" scenarios like a world of only pleasure machines that ignore autonomy or beauty.

Overconfidence and Calibration Studies

• Overconfidence Prediction Model: Greenberg identified five traits of a skill that predict whether people will be overconfident or underconfident regarding their performance relative to others.
  • High Self-Perceived Ability: Skills people believe they are good at tend to yield overconfidence.
  • Subjectivity: Skills viewed as matters of personal opinion (e.g., writing a novel) correlate with overconfidence.
  • Experience: Higher self-reported experience correlates with overconfidence.
  • Personality Connection: Skills viewed as reflective of character (e.g., making friends) correlate with overconfidence.
  • Perceived Difficulty: Skills viewed as very difficult correlate with underconfidence (e.g., running a marathon, knitting).
• Empirical Findings:
  • People are generally overconfident across a broad range of skills but consistently underconfident in skills perceived as difficult, objective, or unrelated to personality.
  • A new tool is being developed on ClearThinking.org to predict overconfidence based on these five traits.
• Calibration Training:
  • Greenberg created a quiz distinguishing common misconceptions from true facts to train users in calibration (predicting confidence intervals).
  • Results from the quiz highlighted the difficulty of verifying truth, citing the "spinach iron myth" which involved a chain of errors where the original error was a decimal shift, and the subsequent debunking was also a myth.
  • Calibration training demonstrates that people often give ranges that are too narrow (overprecision), though they can learn to improve this.

Bayesian Updating and Evidence Evaluation

• The "Question of Evidence": A framework for estimating the strength of evidence using the Bayes Factor.
  • Formula: Posterior Odds = Prior Odds × Bayes Factor.
  • Bayes Factor Calculation: "How much more likely is this evidence if the hypothesis is true compared to if it is false?"
  • Ratios indicate strength: 3:1 (moderate), 30:1 (strong), 1:30 (strong against).
• Common Cognitive Errors in Updating:
  • Ignoring the Comparative: Focusing only on the likelihood of evidence given a true hypothesis, ignoring the likelihood given a false one.
  • Dismissing Weak Evidence: Accumulating weak, contradictory evidence over time can shift beliefs significantly; ignoring "trickles" of evidence leads to stagnation or error.
  • Neglecting Priors: Failing to account for the base rate of a hypothesis (e.g., mistaking a friend in a random city for a known friend due to low prior probability).
• Reliability of Priors and Intuition:
  • Trust in Intuition: Intuition is reliable only when there is high-frequency, low-noise feedback (e.g., a therapist reading patient cues). It fails in domains with delayed feedback (e.g., long-term therapy outcomes) or no feedback (e.g., philosophy, macroeconomics).
  • Reference Class Forecasting:
    • Start with an "outside view" (base rate of similar events) rather than an "inside view" (specific project details).
    • Bias-Variance Trade-off: Narrower reference classes reduce bias but increase variance due to smaller sample sizes; optimal forecasting balances these.
    • Example: Estimating startup success or project duration requires weighting historical failure rates (e.g., 90% of startups fail) against specific mitigating factors.
• Handling Specific Evidence Types:
  • Models/Theories: Treat all models as wrong but useful; averaging predictions from multiple independent models reduces individual bias and noise.
  • Empirical Studies: Evaluate studies by estimating the Bayes Factor, considering sample size, methodology, and publication bias (e.g., "p-hacking").
  • Heuristics: Qualitative frameworks (e.g., "do I feel excited hiring this person?") can add value to predictive machines even if they resist quantification.
  • Sanity Checks: Compare estimates to broad constraints (e.g., a company's value cannot exceed global wealth) to detect citation errors or "citation worms."

Case Studies in Updating

• Trump-Russia Collusion:
  • Prior: Set low (e.g., <10%) based on the rarity of presidents colluding with foreign powers.
  • Update: The meeting between Trump Jr. and a Kremlin contact increased the probability (Base Factor ~3-4:1), moving the odds from ~10% to ~30-40%.
  • Context: Updated by distinguishing between the plausibility of the act and the evidence for it.
• US-China War:
  • Prior: Reference class suggests ~70% war likelihood during major power transitions (historical data).
  • Update: Adjusted downward due to the "regime change" of nuclear weapons (mutually assured destruction reduces war probability).
  • Warning: Avoid double-counting evidence (e.g., multiple tariff disputes may not be independent signals).
• North Korea Nuclear Disarmament:
  • Prior: Low (~10%) based on few historical examples of nations voluntarily giving up nukes.
  • Update: Statements of intent to disarm were given a moderate update (3:1) but counterbalanced by the strategic incentive to keep weapons for security.
  • Outcome: Net probability likely remains low due to strong strategic incentives to retain weapons.
• Theranos:
  • Analysis: The company's failure to release a product for years ("vaporware") shifted the hypothesis from "incompetent" to "fraudulent."
  • Evidence: Poorly written scientific papers suggested incompetence (sincere failure) rather than calculated fraud, as a fraudster might produce better-looking fake data.
  • Investment Signals: Lack of interest from knowledgeable Silicon Valley biotech investors served as a negative signal.
• Power Posing:
  • Context: Original study claimed large effects on hormones and behavior; subsequent pre-registered replications largely failed to find these effects.
  • Greenberg's Analysis: Suggests a small, variable effect exists (n=1000 study found mood/power boost), but the original effect size was likely overstated.
  • Placebo Factor: Acknowledged that the "power" effect may be a placebo or self-fulfilling prophecy, which is still valuable if it improves mood.
• Dietary Advice:
  • Conclusion: Strong evidence exists only for specific deficiencies (e.g., Vitamin C for scurvy, Vitamin D for older women).
  • General Advice: Most specific food claims lack high base factors; randomized trials are difficult to conduct, making causal links hard to prove.
  • Anecdotes: Rapid recovery from a long-term condition after a diet change can provide a high Bayes Factor for an individual, even if it doesn't generalize.