on-chain generative traces of ephemeral AI reasoning
Every time an AI responds, it makes thousands of decisions you never see. At each token, it weighs alternatives, hesitates, almost says something else. Then the moment passes. The thought is gone. Ephemeral. Unreproducible.
This project captures that invisible process and freezes it on-chain forever.
You submit a prompt. The AI thinks. We capture the raw logprobs — the probability distribution at every single token. How confident was it? How many alternatives did it consider? Where did it almost break? Where did it change direction entirely?
That data becomes the art. Not a picture of a thought. The thought itself, rendered as what it actually is — a probability field collapsing, one token at a time.
The Fog
Particles drifting in a noise field. This is the probability space before any token is chosen. Everything is possible. Nothing has been decided.
The Collapse
Each token collapses a region of the field. Confident tokens burn clean, small holes. Uncertain tokens detonate — scattering possibilities, tearing through the fog. The path the thought takes is shaped by its own confidence.
The Corruption
When the model doesn't know what to say, the image breaks. Pixel sorting erupts — rows of pixels reorganize by luminosity, Asendorf-style. Data stutters. Colors crush to fewer bits. The corruption IS the uncertainty. The cleaner the image, the more the model knew what it was doing.
Phase Shifts
Sometimes the model hits a wall — confidence craters near zero, alternatives explode, and when it recovers, it's thinking about something completely different. The color palette shifts. The path breaks and reforms. These are rare. They can't be manufactured. They emerge naturally from the data.
Rarity is determined entirely by the data. No artificial scarcity. No hand-picked traits. The thought itself decides what it is.
Every piece carries on-chain traits derived from its logprob signature — fork density, minimum confidence, phase shift count, semantic drift range, crisis depth, model identity, token count. These combine into a rarity score. Some traits are common. Some combinations have never been seen.
Trait Examples
Fully on-chain SVG. No PNGs. No GIFs. No IPFS. No external servers. The renderer and your thought data live entirely in the smart contract. If every server in the world goes down, your piece still renders from the blockchain.
Multi-model. Choose which AI brain generates your thought — GPT-4o, Qwen, or others. Each model produces different uncertainty patterns. Different brain, different art.
Data-driven. Nothing is manufactured. The shape, the color, the corruption, the rarity — all derived from the actual logprob data of that specific inference pass. No two thoughts can ever be the same.
01
Write a prompt
Any thought, question, or phrase — up to 500 characters
02
Capture
The AI processes your prompt and we extract the raw logprob data
03
Preview & Mint
See your unique piece — mint it on-chain or try a different prompt