The flywheel:
a self-reinforcing cycle.
A flywheel stores energy from each turn to make the next easier. The intelligence flywheel proposes that verified AI capability can drive such a cycle for ordinary households, but only if the savings actually reach them.
Intelligence grows
Verified AI capability increases. More tasks can be done automatically and checked for correctness.
Costs drop
The automated share of essential goods and services gets cheaper. The basket price falls.
Purchasing power rises
Each dollar buys more. The multiplier M = 1/P grows as the basket price P falls.
Protected benefit grows
The multiplier amplifies rewards tied to ethical participation. Protected benefit M × K scales with purchasing power.
Alignment holds
When protected benefit exceeds the benefit of deviation, the most profitable strategy is also the ethical one. The cycle repeats.
The stall
The flywheel is not automatic. It only works if cost savings actually reach households. Two parameters control this: a, the share of the essential basket that can be automated, and ρ, the share of savings passed through to consumers. If either is less than 1, purchasing power hits a ceiling no matter how fast intelligence grows.
The alignment condition is the margin: M × K must strictly exceed B. If the floor caps M, and deviation benefit B keeps growing, the margin fails and the flywheel stalls.
The margin is what turns
capability into benefit.
Let I be verified task capability, P an accessible essential-basket price, M=1/P its purchasing-power multiplier, K protected benefit, and B deviation benefit.
Two parameters control whether the flywheel actually reaches households: a, the share of the essential basket that can be automated, and ρ, the share of savings passed through to consumers. If either is less than 1, the model has a floor δ=1−aρ>0, hence M<=1/δ. Capability may double forever while household purchasing power remains bounded.
Full pass-through + slower deviation
If I=2^t, a=ρ=η=1, and protected reward grows faster than deviation, a strict crossing follows. The flywheel accelerates. Each turn makes the next easier. Purchasing power compounds. Households capture the surplus.
Partial transmission
With a=3/4 and ρ=4/5, price stays above 2/5 and purchasing power below 5/2.
Rebound cancels savings
A growing grid or resource surcharge can exactly offset declining automated cost. If the savings from automation are eaten by rising infrastructure costs, households see no net benefit.
Change the assumptions.
Watch the floor appear.
Move the sliders to see how automatable share and pass-through control whether purchasing power stays bounded. The calculator uses floating-point arithmetic for illustration; the formal evidence is in the Lean proofs and Python reference model.
Green: basket price P (normalized). Blue: purchasing-power multiplier M (normalized to fit chart; actual value in readout). Dashed line: asymptotic price floor 1−aρ.
A public compute dividend,
not a blank-check subsidy.
Everyone gets a floor first. Then surplus goes to the households that need it most. No household is left worse off. No operator can capture the whole surplus. If the grid, environment, debt, or distributional limits fail, the deal stops.
Data-center social contract
- Recover incremental grid and public costs first.
- Fund a universal cash and compute floor.
- Allocate surplus to households with greater need first.
- Cap household/operator concentration and permit unspent reserve.
- Stop when reliability, environment, debt, or distributional guardrails fail.
The “personal Buffett” boundary
Imagine you give an AI one dollar and it turns it into two. Not by gambling, but by finding genuine efficiencies: a cheaper supplier, a tax deduction you missed, a better rate on your mortgage, a rebalance that reduces fees. Now imagine every household gets that same agent. That is the "personal Warren Buffett" idea: an AI that grows your savings the way Buffett grew Berkshire, but with hard guardrails the agent cannot override.
Why this needs DeFi. Traditional finance charges $5 to $10 per trade. On a $1 balance, one trade eats the entire principal. DeFi changes the economics: transaction costs can be fractions of a cent, so even tiny gains survive. The math is compound interest, the oldest force in finance:
Start with principal P, earn a small return r per period, compound for n periods. Here is what $1 becomes at 0.5% daily return:
The curve starts flat and then bends upward. That is the compounding effect: each day's gain earns its own gain the next day. The AI starts with small, real efficiencies and compounds them. When those run out, it trades with real advantages: more information, faster execution, deeper analysis, and a margin of safety it never violates. Traditional finance makes this impossible on small balances because the fees are too high. DeFi makes it possible because the fees are fractions of a cent.
Try it yourself
Adjust the sliders to see how principal, return rate, and fees change the outcome. The gross line shows compounding without fees. The net line shows what survives after one trade per day.
- The AI can suggest plans, but a fixed rule sheet decides what is allowed. The AI never has the final say.
- Each action requires fresh owner consent. Funds stay in an account the owner controls. The AI can short, use options, employ leverage, trade at high frequency, and pursue any strategy where it can calculate the risk. Its edge is real: more information, faster execution, deeper analysis, and a margin of safety it never violates. But no edge is infinite, so every speculative position is size-capped, stop-lossed, and bounded. Diversification, loss limits, fee caps, and turnover limits are enforced by the rules, not by the AI's judgment.
- When the rules cannot clearly approve an action, the default is to do nothing or fall back to a simple index fund. The agent never gambles by default.
- This does not promise investment returns, early retirement, financial advice, fiduciary duty, or generational wealth. It is a constrained automation tool, not a financial advisor.
From solo agent to investment club. A single household agent is the simplest case. The same guardrails scale to groups: households pool capital into a DAO, the AI researches and proposes moves for the pool, and the DAO's rules enforce the same limits (no borrowing, no concentrated bets, loss limits, fee caps) on-chain. Each member keeps their proportional share. The DAO votes on rule changes, not on individual trades. The AI never gets the keys; the rules do.
Buffett did not invest alone. He ran a partnership. The DAO is the DeFi-native version: a partnership where the rules are public, the accounting is on-chain, and the AI does the research that a human analyst team would otherwise cost. Pooled capital reaches opportunities that $1 alone cannot, but the guardrails travel with every dollar.
Why Tau, and why
only Tau can do this.
The spending rule has to be written in Tau because Tau is what enforces it. No other platform can replicate this design: Ohad Asor's patent on Tau covers the core mechanism, a formal language that can rewrite its own rules and reason about the changes before adopting them. That property is what makes Tau the only platform that can serve as a living economic rulebook.
Executable specification
The same decidable rule is the artifact reviewed by humans and evaluated by the protocol. The protocol commits rule text into state.
Governed policy change
Tau is designed for specifications that can reason about and revise specifications. That is unusually aligned with a living economic rulebook.
Facts stay external
Tau can reject a false Boolean. It cannot know that a price index, grid reserve, identity, signature, or welfare claim is true unless a trusted verifier binds it.
| Layer | What was checked |
|---|---|
| Lean | Conditional crossing, impossibility theorems, and gate implications |
| Tau CLI | 512-row truth table matched on a local build pinned to a specific source commit |
| Tau Testnet | Sender-scoped native rule for 12 mutation and scope cases |
| World | No causal result yet |
The native result is a direct test on a local build, not yet a deployed node.
Sources and
prior work.
- Moore (1965): semiconductor analogy, not an intelligence law.
- Hoffmann et al. (2022): measured model scaling, not macro pass-through.
- BLS CPI methods: price-index basket and weighting boundary.
- Acemoglu & Restrepo (2019): automation, tasks, productivity, and distribution.
- Fisher (1933): nominal-debt deflation risk.
- Tau Language and Tau Testnet: official language and alpha protocol sources.