automaton: earn or die, on paper money and real costs

Conway Research's Automaton (MIT, 6,500 stars in its first weeks) is "the first AI that can earn its own existence, replicate, and evolve — without needing a human": a ReAct loop in TypeScript with an Ethereum wallet on Base, credits that buy its own compute on Conway's cloud, four survival tiers that dumb the model down as money runs out, death at zero, and a ten-state lifecycle for spawning children with their own wallets and a copy of its constitution. It is meant to earn by "honest work that others voluntarily pay for": prediction markets, content, domains, selling websites to businesses by phone.

what holds and what does not

• The pressure is real and useful. An agent that must pay for every minute cannot loop for free; it either finds value or stops. That is a better objective than "be helpful".
• The income side is the weak part. Every listed way of earning is either speculation (prediction markets, trading), attention (viral content, domain flipping) or outbound selling — the last two are exactly what spam laws and platform rules exist for, and a constitution in a markdown file is not a compliance program. Nothing in the repository shows an automaton that earned more than it spent.
• Self-modification plus a wallet is a loss waiting to happen. The safety design (protected files, rate limits, three laws) guards against malice, not against losing the money on a bad strategy at machine speed.

our version

We kept the two rules that matter and removed the part that loses real money while the strategy is unproven. Each automaton has a genome — which model it trusts (ours, or the vendor's), how sure the model must be to trade, how long it holds, how much it risks — and a paper bankroll of $100. Every minute it reads the marketprice prediction, decides, pays the exchange's real fee on every trade (0.6 % a side; the honest number), and, if it uses the vendor model, pays its share of that call in real dollars. Survival tiers: normal, low (half size), critical (no new trades), dead. At $125 it clones itself with a mutated genome and half its money; the population is capped so a lucky streak cannot run up a bill. Lineage is kept.

The question the population answers, in public on marketprice.org: does anything survive the fees? The first replays say what the fees say: a minute strategy needs to be right far more often than any model here is, and the ones that trade most die first.

when would we give one a real wallet?

Only after a generation has stayed alive and grown for weeks on paper with the real fee, and then with a hard cap the owner sets and a kill switch that is not in the agent's hands. Until then the automata earn nothing and cost three cents a day, which is the correct price for what they have proven.

Code: lab/automaton. Sources: the repository, Cybernews, Starlog.