Programmable mining
One argument about what Bittensor is, and three instruments that record what the puzzles are, whether a reader can find them, and who says so.
Sealed: fixed at birth. Open: anyone can write it. TAO keeps Bitcoin's money on purpose: no pre-mine, 21 million, halvings, paid out by the block.
Counts, not readings: each is computed from the datasets and links to the rows it was counted from. All findings, the rules, and the allocator table
Pay strangers, every block, for work you define. Bitcoin defined one job, and the job was guessing a number. The hash bought security and nothing else. The mining was never the waste. The fixed puzzle was.
The design space moved up a layer, to contracts. A decade of chains lives there. Nobody reprogrammed the mine. They reprogrammed what its output could do.
The puzzle is a program. A subnet owner writes what the work is, how it is scored, and how the block reward splits. Miners compete on the work. Validators compete on the judgment. The chain does not run the puzzle. It pays the answer.
Ethereum's work is verified by re-running it. Bittensor's work can only be verified by judging it, and the network's own docs call the judgments subjective OTF 2024. A hash is right or wrong AS 2022. A judgment can be gamed, and on Bittensor the judges are a paid market of their own.
"Show me the subnet, and I'll tell you what the miners are optimizing for."
Ala Shaabana, June 2026. Below: pick a job. The same puzzle in three states: as written, gamed, and patched. Nothing touches the chain.
The numbers are illustrative: a toy scoring function, a toy miner, nothing on chain. The shape is not. Every subnet's incentive mechanism is a function like the one on the left, and every one of them has a right-hand panel somewhere.
A hash cannot be gamed. A judgment can. The subnets already write against it: Affine sells a mechanism that is "sybil-proof, decoy-proof, copy-proof, and overfitting-proof" SN120, and Ridges names "applying a stored answer" as hardcoding SN62. Reward hacking is to Bittensor what reentrancy was to early Ethereum: the native exploit of the new design space.
Research versus revenue, the pace of rule changes, enterprise readiness. Each asks the same thing of a different puzzle: is this work worth paying for before anyone has bought it? The foundation says it plainly: "a subnet nobody uses still collects a slice of every block" OTF 2026. Targon's target is a dollar of revenue for every dollar a miner earns SN4. Bitcoin subsidised security for years before fees mattered. The frame names the bet. It does not settle it.
The argument above is one person's. The lineage below is not.
Two of eighteen narrators use the phrase, against seven for market, the network's dominant frame. The foundation's own about page now makes the Ethereum comparison and lands it on markets. The argument lands it on mining, on purpose: a Bitcoin reader already holds that slot, and mining names the work where market names the venue. The lineage is theirs. The choice of word, and the line about work that can only be verified by judging it, are the argument's. Trace the phrase across every narrator in the metaphor index.
If a subnet is a mining program, it can be read: what a miner is paid to produce, how a validator judges it, how the pay splits, who judges, how often it turns, who buys, and what has been gamed. Seven fields per subnet, each with its source and a mark for what the claim rests on, from the chain and the code down to the site's own reading. The index beside it rates whether a first-time reader could have found the puzzle in five minutes. Both have a method page and a correction form, and neither depends on the argument being right.
A configuration file: the parameter settings for one of the supported variant callers, the software that decides which differences in a sequenced genome are real mutations. The miner never runs the genome itself. What is paid for is the recipe that, when the validator runs it on this round's synthetic genome, finds the planted mutations best.
GPU time by the second. A miner, called a provider, runs a small coordinator plus one or more NVIDIA GPU machines and sets an hourly price per machine. What is paid for is a machine that passes the validator's probe: rented machines are paid from one pool, eligible idle machines of the listed models from another.
A shopping agent: one Python file exporting agent_main that, given a shopper's request and budget, searches a catalogue, compares listings and places an order inside a sealed environment. What is paid for is that agent's performance on the day's hidden tasks. Every run's trajectory is kept by the subnet as training data for its own model.
A trained checkpoint: the saved weights of a coding model, uploaded to a private bucket and signalled on chain. The subnet pays whoever holds or recently held the crown, so the thing produced is a model that beats the reigning king in a duel.
Compute time. A miner runs a Kubernetes cluster of GPU servers inside confidential virtual machines and hosts chutes, the platform's word for a deployable model or workload. What is paid for is the seconds an instance is live and serving, weighted by the GPU class it runs on. The index lists the SDK repository; the validator lives in chutes-api.
A fine-tuned checkpoint of Qwen3.6-35B-A3B, an open 35 billion parameter model, trained further to act as a coding agent: read a real repository, run shell commands, make edits over several turns. The miner uploads it to HuggingFace or Hippius and commits a v7 reveal on chain. What is paid for is a checkpoint that beats the reigning king in a duel, twice.
All 32 programsThe legibility indexSays versus showsWhat a field means
"If you ask five people what Bittensor is, you will get five different answers" BS 2024. The map records the answers, verbatim and dated, reads every voice through one framework, and traces each frame to its first use. It has its own method page and correction form.