Bittensor

Bittensor is a network where machine learning models compete for rewards based on how useful other participants judge them to be. It has a maximum supply of 21 million TAO and a halving schedule, applying Bitcoin's monetary structure to a completely different kind of work.

Evaluating intelligence is the central problem

Proof of work verifies a computation trivially. There is no equivalent objective test for whether a model's output is good. Bittensor's answer is peer evaluation: validators score miners' responses and rewards follow those scores. This is subjective by nature, and designing it to resist collusion is the network's hardest and most interesting engineering challenge.

Subnets specialise

Rather than one undifferentiated market, the network divides into subnets each targeting a specific task — text generation, image work, data scraping, prediction. Each defines its own evaluation criteria, so what counts as good is set by people who understand that domain rather than by a global rule.

Bitcoin-style issuance for a different asset

A 21 million cap with periodic halvings gives TAO a fixed monetary schedule. Emissions are distributed to miners and validators according to evaluated contribution, so issuance funds useful work rather than raw computation.

The recurring criticism

Any system where participants score each other invites gaming: collusion between validators and miners, models optimised for the scoring function rather than genuine usefulness. The network has iterated on its incentive design repeatedly in response, and this remains the open question a prospective participant should examine most closely.

What it is attempting

An open market for machine intelligence with permissionless participation, in a field otherwise dominated by a few very large firms. That is a genuinely ambitious goal, and the difficulty of measuring quality objectively is the reason it is hard.