Kompas · Hypothesis

Is Bayesian reasoning practical for everyday decisions?

Not the math — the habit of treating beliefs as probabilities and updating them as evidence arrives.

vs
Next question: Is this an expected outcome? →
How this page picks what you see

The homepage hero question is selected by a Thompson Sampling model. Each question is a bandit arm. The model maintains a Beta posterior over its click-through rate and samples from it to decide which question to show next.

Posterior per arm: Beta(α, β) where α = votes + 1 and β = max(1, impressions − votes + 1). Thompson Sampling draws one scalar from each posterior and shows the arm with the highest draw.

Current arms
Question α β mean 90% CI Win %
ai-math 12 238 0.048 0.028–0.072 27.6%
rationality 5 127 0.038 0.015–0.068 12.4%
deno-deploy 2 65 0.030 0.005–0.070 9.6%
abtest-react 3 90 0.032 0.009–0.067 9.2%
bits 3 91 0.032 0.009–0.066 8.9%
cookie-consent 2 69 0.028 0.005–0.067 7.8%
transcripts 2 70 0.028 0.005–0.065 7.5%
bluesky 2 71 0.027 0.005–0.065 7.2%
entropy 1 46 0.021 0.001–0.064 5.9%
bayesian 1 53 0.019 0.001–0.055 3.8%

α = votes + 1 · β = max(1, impressions − votes + 1) · mean = α / (α + β) · sorted by win %

Concentration: top arm wins 27.6% of 10,000 simulated selections → exploring

Field notes

The same reasoning, applied to the decision to run this campaign. Each row: what was believed before, what the data said, what's believed now.

Pomodoro searchers engage with a philosophical question about breaks

Prior No prior data — uniform over [0, 1]
Prior (no data yet) mean 50.0%, 90% CI [2.5%, 97.5%] — campaign pending

Bayesian-reasoning searchers vote on a philosophical question

Prior No prior data — uniform over [0, 1]
Prior (no data yet) mean 50.0%, 90% CI [2.5%, 97.5%] — campaign pending

Thompson-Sampling searchers follow the DAG link to bayesian-meta

Prior Slight optimism: technical audience is curious. Beta(0.6, 4.4) → ~12%
Prior (no data yet) mean 12.0%, 90% CI [0.0%, 33.8%] — campaign pending

Cold philosophical search ads get ~1–2% CTR on bayesian-reasoning keywords

Prior No expectation — uninformed prior
Evidence 13 / 829 — Experiment 1, src=ads_bayesian, 14 days
Posterior mean 1.7%, 90% CI [1.0%, 2.4%]

Of the people who click the ad, more than half vote on the question

Prior Mild optimism — self-selected clickers are curious. Beta(0.6, 0.4) → ~60%
Prior (no data yet) mean 60.0%, 90% CI [3.0%, 100.0%] — campaign pending

The cost per newsletter subscriber via this channel is under €5

Prior No prior — depends on click-to-vote and vote-to-subscribe rates
Prior (no data yet) mean 50.0%, 90% CI [2.5%, 97.5%] — campaign pending

Run your own numbers → beliefs playground