Bigger and Fairer

Every setting tested is a dot below. The gold line is the efficient frontier: settings no other setting beats on both size and equality, all with inflation under 5%. It runs almost horizontally across the top. Moving left along it, from 0.75 to 0.14, costs less than 1% of output.

Pure capitalism is not the richest corner: the recommended mix doubles its output. The poorest 40% show the pattern even more sharply: their consumption tends to be highest where inequality is lowest.

Under Extreme AI about half of households have no job and wages are a sliver of income. In these runs output follows household spending power. The AI grant (government.ubiShare) pays a share of the AI slice of output equally to every household. Switch it off and output stalls at about 63× today in all 20 seeds, the median household consumes 0.47× today’s median, and the top 1% hold 77% of wealth. At the 25% default, output reaches 114×; at 50% or more it levels off at about 125×. With growth levers maxed and broad ownership, the same dial takes output from 147× to 294×.

No inflation cliff. Bigger grants did not push prices up. Across every Extreme-AI setting, median inflation stayed at or below 0.5% a year, and no single run exceeded 1.2%.

Profits and claims on firm capital are split in proportion to skill raised to a power: 1.5 plus ai.ownershipConcentration × (AI factor − 1). Under Extreme AI the AI factor ends near 17.5, so the default 0.8 gives an exponent of about 15. Profits flow to a handful of the most-skilled households. At 0.1 the exponent is about 3.1. The top 1%’s share falls from 66% to 19%, the wealth Gini from 0.88 to 0.71, and output rises from 114× to 127×. That fits the spending-power pattern in section 02: more of the profit reaches households who spend it.

Skills are drawn from a lognormal spread (household.skillSigma), capped, then scaled so the average is 1. With the default steep profit tilt, the rule is “most skilled takes nearly all”, so what matters is how many households share the top. At the default 0.7, about 5 households tie at the cap and split the profits. At 1.1, 36 tie, so the top 1% hold only 15%, but the top 10% still hold 86%. At 0.3 or 0.5 nobody ties: one household stands alone at the top and takes 41–66% of all wealth. Output falls to 101× and 83×, the same spending-power pattern again. Only at 0.1, where even the top household is barely above average, does the spread truly compress wealth (Gini 0.56, output 126×).

With broad ownership (0.1) the tie effect disappears. Narrowing the skill spread then lowers the wealth Gini steadily (0.72 at 1.1 to 0.17 at 0.1), and output stays at 126× throughout. In the real world this lever stands for broad access to education and tools, not a policy switch.

One working hypothesis was that a bond-financed grant leaks back to the rich through bond coupons (paid to banks and passed on to large deposits), so the income tax should rise with the grant. The runs say otherwise. Setting the bond rate to zero removes the coupons and barely changes anything. Raising the tax to fund the grant leaves everyone’s consumption the same but pushes wealth inequality back up.

Why. A bond-financed grant leaves new money in every household’s account, so savings spread widely. A tax-financed grant takes the same money back out of household incomes, the treasury surplus is returned in proportion to income, and more of what is left of household wealth is claims on firm capital, which are concentrated (claims are 11% of household wealth at a 20% tax, 38% at 50%). The income Gini is 0.50 in every row: only who holds the savings changes. Raising the flat tax with the grant at its 25% default is worse still: at 35% the top 1% hold 83% (default 66%).

The price is public debt. Bonds held by banks reach 3.0× a year’s output at a 100% grant, against 0.8× at the defaults. The recommended mix stays at 0.9×. These runs show no inflation from this debt; a real economy might react differently.

Base productivity growth is the strongest growth lever: 4% a year takes output from 114× to 213×. Falling compute prices are essential: if the price of compute stops falling, Extreme AI never takes hold and output reaches only 2.9×. Compute productivity, the final automatable share and adoption timing also matter. Adoption steepness, the physical-task limit, robotics timing and a faster agent-cost half-life barely move output, because Extreme AI already saturates them.

All growth levers maxed together give 234× with a wealth Gini of 0.86 (defaults: 0.88). They don’t steepen the profit tilt in these runs: the AI factor ends near 17.7 either way. Growth levers enlarge the pie but don’t touch the slices.

One representative run (seed 11). The owner is the wealthiest household in year 20 under Extreme AI defaults and starts among the most skilled. The median family sits mid-pack on skill, starting wealth and final consumption. The struggling household started in the bottom 40% and ends near the bottom in consumption. The same households are traced through each world. In the recommended mix their skills are closer together, so the owner starts less far ahead.

Under pure capitalism the median family consumes less than today, and the owner shares the top with about 36 tied households in an economy half the size. In the recommended mix all three end up hundreds of times richer than today’s median household. Only the Extreme-defaults owner does better, by taking most of a smaller pie.

How the working hypotheses fared