Hi,

Picture a county fair that posts a prize pot every seventy-two minutes. For years every fair on the circuit cut that pot the same way: eighteen parts to the organizer, forty-one to the judges in the tent, forty-one to the exhibitors who brought the entries.

This month the circuit let each organizer write their own cut. An organizer who sweetens the exhibitors' share wins the best machines from the fair down the road, and the judges' tent gets smaller.

One slice does not get smaller. A toll paid to holders who sit outside every fair stays exactly the size it was, and the chain refuses any cut that would shrink it, by as little as two tenths of a percentage point.

So when the tent shrinks and the toll does not, who pays the difference?

Listen:

Apple Podcasts: https://podcasts.apple.com/kg/podcast/plain-strata/id6783455764?i=1000798386602

YouTube: https://youtu.be/BgdgLqbjPC0

The full piece, no need to click through:

On 20 October a change went live in Bittensor, the largest of the networks that pay strangers to do AI work. Inside its test file sit two checks, written one under the other. The first says that on a particular sub-network the judges can be paid eleven point nine percent of each round's new tokens, and the chain should accept that. The second says that eleven point seven percent should be refused.

Two tenths of a percentage point is the whole difference between a rule that passes and a rule that is rejected. The line was not drawn there to keep the judges fairly paid. It was drawn there to keep somebody else whole, somebody who does no judging and no work on that sub-network at all.

Until this month every sub-network split its new tokens the same way: eighteen percent to the owner, forty-one percent to the judges and the people who back them, forty-one percent to the machines doing the work. Now each owner can write that split down for themselves. Each sub-network can now set its own pay split, and the one share the rules anchor in place belongs to holders outside every sub-network.

Take a county fair where a prize pot is posted every seventy-two minutes, and three groups divide it: the organizer who runs the fair, the judges in the tent who decide which exhibits are good, and the exhibitors who brought them.

On this network the fair is a sub-network. The exhibitors are miners, programs running on graphics cards and doing the AI work. The judges are validators, who score that work, and behind each judge stand holders who have put their tokens behind it. The pot is not money collected at the gate. It is new tokens the network prints, which is what emission means: the word comes from the Latin emittere, to send out. Each sub-network has its own token, called alpha, and the ledger, a shared spreadsheet nobody is allowed to lie to, decides how big each pot is.

Until October the bylaws were identical at every fair. Eighteen to the organizer, forty-one to the tent, forty-one to the exhibitors. Nobody voted on it each time. It was simply how every pot was cut.

One more party stands outside all the fairs. The network has a base layer called the root, and holders who put their tokens there are not backing any one fair. Think of them as owning a piece of the association that charters all of them. In return, the association takes a toll out of the judges' tent at every fair, and that toll is what root holders are paid. Their money works at no particular fair, and they are paid by all of them. So who decides how each pot is cut?

Under the new rule the owner of a fair writes its bylaws, inside limits. The organizer can take anything from nothing up to a third. The tent can get up to half. The exhibitors must get at least a third. A split can change about once a week. Fairs that already exist keep the old split until their organizer chooses otherwise, and the size of the pot never changes. Only the cut does.

Why would anyone move? The plainest reason is competition for machines. A fair that pays its exhibitors sixty-two percent will look better to a graphics card owner than a rival paying forty-one, and the machines are what the pot exists to buy. An organizer can also pay themselves less to make the offer richer. The example in the code does both: the organizer keeps eighteen percent, the tent drops from forty-one to twenty, and the exhibitors rise to sixty-two.

So a pay rate used to be a rule of the network. It is now a setting, like a price on a menu. But the network did not leave every price to the menu, and before it let the tent shrink it had to answer one question: if the tent gets smaller, what happens to the toll?

Left alone, the toll would shrink with the tent. It is a share of what the tent holds, so halve the tent and the association is paid half as much. The designers did not allow that. Each time an organizer changes the split, the chain recalculates a multiplier that decides how heavily the association's stake counts inside that tent, and raises it until the toll comes out exactly as large as it would have been under the old bylaws.

The worked example in the code is a real sub-network, the fourth one on the network. Under the old split its toll was about five point four percent of the pot. Move the tent from forty-one to twenty, and the multiplier climbs from about eighteen percent to about forty-four, and the toll stays at five point four. The tent holds less than half as much. The toll has not moved.

The same rule explains the two checks in the test file. The multiplier cannot go past one hundred percent, because the association cannot count for more than all of its stake. Work backward from that ceiling and you get the smallest tent the chain will accept, which for that sub-network is eleven point eight one percent. Eleven point nine passes. Eleven point seven fails. The minimum exists to keep the toll whole. It was never meant to keep the judges whole.

So the tent can shrink, but only so far, and the toll does not shrink with it. If one share cannot move, who pays when the tent does?

The numbers in the example are published; the subtraction is mine. Under the old split the tent held forty-one points of the pot. The toll took five point four of them. That left thirty-five point six for the judges and everyone who had backed them.

Under the new split the tent holds twenty points. The toll still takes five point four. What is left for the judges and their backers is fourteen point six. That is a drop of nearly six in ten. The toll was thirteen percent of the tent, and it is now twenty-seven percent.

The organizer did not lose anything. The exhibitors gained twenty-one points. The cost fell almost entirely on the people behind the judges.

Finance has an old name for this arrangement. A bond is sold in tranches, from the French trancher, to cut: a slice cut off from the whole. The senior tranche is paid first and in full, and senior is the Latin word for older, the claim that came first in line. The junior tranche takes whatever is left, and it takes the first loss when there is one. A bankruptcy court works the same way, with secured creditors paid before anyone else. When one claim is fixed exactly, every change anywhere else in the pot is absorbed by the claims that are not fixed.

Here, the first-loss slice is the judges' tent minus the toll. Who sits in it? The judges, who are paid for the work of scoring. And behind them the holders who put tokens behind a judge, usually by pressing a staking button on an exchange or in a wallet. Most of them never chose a judge for its scoring policy, and none of them is in the room when an organizer sets the split. They can leave, but leaving means selling their share of the sub-network's own token into its own trading pool, and that pushes the price down for everyone, including themselves.

Nothing in this is hidden. The chain refuses a split that would cut the toll, names the error, and publishes the arithmetic. It simply decided in advance which claim is senior. What does that do to the job of judging?

Two weeks earlier the same network handed owners another switch. An owner can turn off the comparison between judges, so that one judge, the one with the largest stake, writes the scores for everybody. That change made the measurement a local choice. This one makes the price of the measurement a local choice. How the work is scored and how the scorers are paid are both settings now, held by the person who runs the fair. Inside the ranges, nearly everything is a setting. The toll is not.

The best defence of all this is that it is a market. A single split for every sub-network was a decree. Some fairs have a tent with one judge and a competition that is a formality, and paying that tent forty-one percent is a gift. Other fairs have exhibits that are hard to make and judges that are easy to find. Letting each organizer find the price is how working markets find prices, and the ranges and the weekly limit mean nobody can rewrite a fair's bylaws overnight. All of that is true.

A price only means something if the people paying it can see it and refuse it, and if someone can tell whether the thing being bought is any good. The holders who bear this cost can leave, but only by selling into their own pool. And the instrument that measured whether judging was any good was the comparison between judges, which an owner can now remove. When the pay for judging and the measure of judging are both set by the organizer, the price is for something nobody outside can check.

That is not an accusation. Where one team already runs the only real judge, a smaller tent only writes down what was true. And as of Monday morning I could find no report of a sub-network using the new split. Existing ones keep the old bylaws until their organizers act, and it may be some never do.

The change merged in the middle of October with an automated reviewer still objecting, the second in two weeks to do so. This reviewer's open finding was narrow, about how the cost of making the call is metered, and had nothing to do with the split's arithmetic. Still, the rules about money changed quickly, and the checks arrived afterwards. So what would a change in practice look like?

The first thing to watch is which sub-network moves its split first, and which way. A move toward the exhibitors says organizers are competing for hardware. A move toward the organizer says something else. The chain will show any split to whoever reads its state, which is not the same as telling the people who hold tokens on that sub-network before the change lands.

The second is whether anyone publishes a list of the sub-networks that have left the default, the way nobody has yet published which ones have switched off the comparison between judges. When the first organizer shrinks the tent, does the difference go to the machines, or does it stay with the organizer?

The two voices are AI. The research and writing are mine.

Decentralized AI, layer by layer.

Dastan,