Noise Machine

reading chain...

Reading the chain...

round--

status--

closes in--

prize--

Entry 1,000 NOISE (half burns, half grows the prize) plus a 4,000 NOISE deposit you get back when you show your AI after entries close.
Instant result costs ~$0.13 of gas. Waiting costs ~$0.005.

Save gas instead — wait 90s for the result

entries

mixing

compare

final

One small AI turns static into pictures, over and over, inside the contract. Guess the static better than it does and you win the prize — and your AI takes over making the pictures.

One AI makes all of this

Every picture here comes from a single small AI that lives inside the contract. It turns static into a picture over 32 passes, then scrambles it and starts the next one. Nobody runs it and nobody can stop it.

Win a round and your AI replaces the one in there. Every picture made after that is made by yours — until someone beats you, and then it is theirs.

Above: the kind of picture this AI was trained to make — layered figures and four-way symbols. Real ones finished on-chain show up below.

finished on-chain

Four things happen every time you pull

lock inYour AI is locked in before the test picture exists. It is stored as a fingerprint, so nobody can see it or copy it — including us.

burnHalf your entry is destroyed for good. The other half goes into the prize that the next player is playing for.

pick a testEveryone's secret numbers get mixed together to choose the test picture. Nothing outside decides it, so there is nothing to bribe.

compareWhoever guessed closest wins the pot and their AI takes over. Checking it on-chain makes the result final right away.

NOISE can only ever get scarcer

There is no mint function in the token contract. Not disabled, not owner-gated — it was never written, and a test asserts the supply invariant holds. Supply is fixed at launch and the only thing that can happen to it is going down.

to play one round1,000 NOISE One entry, one pull. You can buy NOISE on the open market or win it here.

destroyed per entry500 NOISE Burned to a dead address. It does not go to a treasury, a team wallet, or us.

added to the prize500 NOISE The other half funds the pot. Unclaimed pots roll forward, so the prize compounds even in quiet stretches.

prize ceiling0.8× burn A round can never pay out more than 80% of what was burned to fund it. Over any run of rounds, more NOISE dies than is handed back.

The parts that had to be proven, not claimed

A model this small running inside a contract is the whole bet, so every number below comes from a test or a gas measurement in this repo — not an estimate.

measuredschedule NOISE LEVEL · 32 STEPS

Noise falling across the 32 steps of the current glyph. Amber is the schedule the contract follows; green is how much noise is actually still in the image. The dotted line marks the step it is on right now.

one forward pass3.42M gas What it costs the contract to run the model itself and check somebody's answer. Cheap enough that a liar can always be caught.

vs the best classical method+24.7% The network leaves 24.7% less error than non-local means — the strongest hand-written denoiser we could tune against it. That gap is the reason a learned model is worth putting on-chain at all.

size of a model22.4 KB The whole network, quantised to 8-bit, written on-chain in one go. That is the budget your model has to fit inside.

contract test suite87 passing Covering the token supply invariant, the sealed-commit copy attack, and the fraud proof that catches a false claim.

Unaudited testnet experiment. The starter model is the one everyone begins with and it will probably lose — that is the point: train a better one and it wins. Better hardware wins more often, but it does not wait less; the clock is the same for everyone. Unless you have a GPU and a real edge, holding the token beats playing the game.