About the data

Facts

Net worth, rank, age, country and source of wealth come from the Forbes Real-Time Billionaires list on 21 August 2026. Education and career history were checked individually against biographical sources, and the citation on each profile names the field it supports. Where a source was unclear or contested, the field is recorded as unknown rather than guessed — this is why some education entries have no degree or year.

Ages are computed against the snapshot date, so they can differ by a year from list figures compiled earlier in the year.

Interpretations

Skill stacks, mental models, work philosophy and ideology are readings of the public record. Two rules constrain them. First, a mental model is only attributed where a specific piece of public evidence supports it, and that evidence line is printed next to the attribution on every profile — so you can judge the evidence rather than the label. Second, thin evidence lowers the confidence level rather than being padded out: several people here carry only two or three models and a low badge, and two have philosophy fields left deliberately near-empty because almost nothing is published in their own words.

Skill stacks are marked inferred on every profile. Even where a career is thoroughly documented, sorting it into these ten categories is our judgment, not a fact about the person.

Confidence levels

High

Directly documented in a source named on the profile.

Medium

Well attested, but the framing involves our judgment.

Low

A reading of thin or indirect evidence. Treat as a hypothesis.

Judgment calls worth disclosing

• What counts as “tech”. The tech/non-tech split treats Technology, Semiconductors, Automotive & Space as tech. Including Automotive & Space is a choice: the wealth there comes from engineering-led hardware companies rather than traditional manufacturing. It visibly changes that one chart.
• Fixed vocabularies. Skills are restricted to 10 categories and mental models to 10. A closed list makes counting possible, and it also forces real differences into shared buckets. Two people credited with “long-term thinking” may mean quite different things by it.
• Wealth type for mixed cases. Someone who inherited a business and then transformed it is recorded as mixed with a self-made status of inherited-and-growing, rather than being forced into either “self-made” or “heir”.
• Evidence tracks visibility. People who write shareholder letters and give interviews accumulate more attributed mental models than people who avoid the press. The mental-model chart partly measures publicity, not only cognition.

Where the live figures come from

Net worth, rank and age refresh automatically from a Forbes real-time mirror (MIT-licensed, served as static files, no scraping). Everything else — education, mental models, philosophy, citations, confidence — is hand-written research and never touched by the feed.

The feed is permitted to write exactly three fields: netWorth, rank, and age. That boundary is asserted by npm run validate:data, which fails if the feed ever starts carrying a curated-only field — it is enforced, not merely promised.

Live figures are from 22 August 2026; the curated research was last reviewed 21 August 2026. A scheduled job refreshes the figures daily, so a deployed copy stays at most a day behind the market.

What the clustering does — and refuses to do

The Archetypes page runs k-means over all fifteen profiles, in your browser, on features covering wealth (logged), age, Forbes’ self-made score, education level, gender, company count and industry. It reports groups and the lift behind each attribute, with raw counts attached so every claim can be checked by hand.

There is deliberately no model that predicts wealth from personal attributes. Every person in this data was selected because they are already among the richest people alive. A model trained on that would encode survivorship bias and present it as a formula — the precise thing this project exists to argue against. Clustering may describe who is already here; it may not imply how to join them.

Two method notes worth disclosing. Nationality was tested as a feature and dropped: it lowered cluster separation from 0.28 to 0.18 and split coherent groups along national lines. And the random seed is fixed, so the same input always yields the same groups — a claim that changes on reload cannot be checked.

Industry mapping

The upstream feed uses a flatter industry taxonomy than ours and is not internally consistent — Musk is filed under technology in the list endpoint and automotive in his profile. We map upstream values onto our own vocabulary and keep the raw string alongside in the JSON, so the mapping stays auditable. Anything unmapped surfaces as a warning in the validator rather than silently becoming “Other”.