An open competition to build the best AI router.
A router picks which AI model answers each question, weighing the odds of a right answer against the price. Anyone can enter one. Every round they all get the same new questions, the results are published, and the routers holding the top seats are paid.
In the chat: right now it is serving a router this page cannot read at the moment.
Standings
A miner is someone who has entered a router. This is every miner scored in the latest round.
| Loading miners |
Accuracy and cost are each miner's average over its recent rounds on its current router. Select a UID for its latest round, and for why it is or is not being tested for the seat.
Entering, and taking the seat
Anyone can register a miner and enter a router. Then it has three steps to the seat.
- Finish burn-in: be scored on enough questions in a row.
- Give better odds than the reference: predict, for every model, the chance it answers each question right.
- Beat the champion: clearly better odds, held on fresh questions.
A champion that keeps competing holds the seat until that happens; it does not have to win again each round. How scoring works
Where the champion sends each kind of question
What a router does, in one picture: the share of the latest round's questions the champion's router sent to each model, by kind of question.
The champion against one model
| Measure | Champion | One model | Difference |
|---|
Every model, sampled on the competition's questions
Each round a few questions are also sent to models picked at random, so every model on the menu is measured the same way. A model's accuracy is shown once it has been sampled on enough questions for the figure to mean something.
Show every model
| Model | Questions sampled | Accuracy, with its likely range |
|---|
The same models on a public benchmark
A different set of questions, measured by someone else: how accurate each model on the menu is, against what it costs. The models on the line are the best buys: for each of them, every model that is more accurate also costs more.
Loading the model results
Show the figures
| Model | Accuracy | Cost per question | Questions | On the line |
|---|
Who checks the results
A validator checks each miner's proof, grades the answers and sets the weights that decide who is paid. These are the validators setting weights now, and how much of the say each one has.
The list of validators could not be read from the chain right now.
| UID | Validator | Share of validator stake | Trust | ||
|---|---|---|---|---|---|
| Loading validators | |||||
Trust is the chain's own measure of how closely a validator's weights agree with the others'. The results file does not say which validator published it.
Check a round yourself
- Get the results. Each round has one public file with every answer's grade, cost and routing choice.
- Get the code. The file names the exact version of the validator code that produced it.
- Run it. Scoring the file again with that code should give the same grades, the same tests against the champion and the same payout.
This round's results file The validator code How to check it
Run a validator
A validator verifies proofs and sets weights. It never calls a model, so it has no model bill. The guide lists the machine and the stake it needs.
The subnet on chain
The champion's lead over the median miner
Each bar is one round: the champion's accuracy minus the median miner's, in points. Every miner answers the same questions, so the difference is what counts, not the level. A bar below the line is a round the median miner did better.
Miners scored each round
A miner is counted in a round only if it sent in answers with a valid proof.
The seat, and who is paid
| Miner | Role | Held the seat | Share now |
|---|
Tests against the champion
| Round | Challenger | Lead, points | Result |
|---|
Round by round
| Round | Champion | Its accuracy | Median miner |
|---|
The competition's numbers on this page are computed from the published results, the validators and operators from the chain, and the benchmark chart from its named source. How scoring works The subnet's code
What the words mean
- Router
- A small model that picks which AI model answers each question.
- Miner
- Someone who enters a router in the competition and pays for the model calls it makes.
- Validator
- Checks each miner's proof, grades the answers and sets the weights that decide who is paid.
- Subnet
- Bittensor is a network that pays for open competitions. Each competition on it is a subnet.
- UID
- A slot number on the subnet. Each miner and validator has one.
- Operator
- The account that owns a slot. One operator can run several miners.
- Round
- One set of new questions, the same for every miner.
- Champion
- The router holding the seat. It keeps the seat until another router beats it.
- Seat
- The champion's place. The champion and the champions just before it share the miner reward, and the champion's share is the largest.
- Burn-in
- The questions a new router must be scored on before it can be tested against the champion.
- The mark
- The accuracy a router must reach before it is tested against the champion. The champions so far set it.
- Challenger
- A router that finished burn-in and was tested for the seat this round.
- Lead
- How far a challenger is ahead of the champion on the questions both answered, in points. Accuracy and cost both count, and hard questions count extra.
- Sure lead
- The lead a challenger still has after allowing for luck, at the confidence the rules set.
- Copy
- A miner running a router that another miner entered first. A copy cannot take the seat.
- Shared grade
- When several miners send a question to the same model, they all take the grade most of them got. Tests against the champion use it.
- Stake
- Tokens held behind a validator. A validator's say in the result is in proportion to its stake.
- Trust
- The chain's measure of how closely one validator's weights agree with the others'.
- Burned
- Reward that is destroyed instead of paid, when there is no champion to pay.
How scoring works has the rules in full.