Hi,

You walk out of the fair, hand your judging badge to the guy at the gate, and take a receipt to another fair where you judge nothing. That is how the hosts put it on this week's episode, and I have not been able to shake it.

On an open AI network the badge is a staked token. It looks like savings. It pays like savings. Underneath, it is a vote about which AI work deserves the money, cast every few minutes, forever.

This week money arrived at that network by two roads. One ran through buttons inside exchange apps. The other ran across a bridge to other chains. Both delivered the money. Both left the vote with somebody else.

Nobody did anything wrong this week, and that is the part worth sitting with. So who is actually doing the judging?

Listen:

Apple Podcasts: https://podcasts.apple.com/us/podcast/plain-strata/id6783455764?i=1000791245032

YouTube: https://youtu.be/yDjm35APTHg

The full piece, no need to click through:

Picture the judging table at a county fair. There is a fixed pot of prize money, there are dozens of entries, and there are judges who taste each one and write down a score. The scores decide the money. Nothing else does. Now suppose the badges that let you sit at that table are also the thing people buy and sell as an investment, and suppose a large, friendly service offers to hold your badge for you and sit in your chair, so you never have to taste anything. Most people would take the offer. It is a good offer. And after enough people take it, the prize money is being decided by a handful of parties, none of whom will ever tell you what they tasted.

That is the week this essay is about, and it happened on Bittensor, the largest network trying to run artificial intelligence as an open market of strangers rather than as a product from one company.

Two things happened, and they are the same story. An institutional operator announced that more than two hundred million dollars of the network's tokens are now staked to its single validator, most of it arriving from ordinary people through buttons inside cryptocurrency exchanges. And the network's token became available on a second and then a third chain, wrapped by a bridge, where it trades as a price and sits in liquidity pools earning somebody else's rewards. The spine of the week is this: on this network a staked token is not a savings account, it is a vote about which AI work deserves to be paid, and every route money took into the network this week delivered the money while quietly leaving the vote behind in somebody else's hands.

Here is the minimum you need, and it takes one breath.

Bittensor pays people to do AI work. The work is split across sub-networks, each one a market for a different job: serving model answers, training a model, forecasting the weather, screening molecules. Every sub-network has miners, which are just machines running AI workloads, and validators, which score how well those machines did. The chain itself understands none of it. It is a shared ledger, a spreadsheet nobody is allowed to lie to, and all it knows is the scores it was handed. Every few minutes it reads those scores and mints new tokens to whoever scored well. That minting is called emission, from the Latin emittere, ex plus mittere, to send out. The money is sent out, and the scores decide where.

A validator's score is not one person one vote. It is weighted by stake, which is the tokens locked behind that validator. Stake began life as a security deposit: lock money, behave, or lose it. But on this network it does a second job that a deposit never did. It is the weight on a judgment about the quality of somebody's AI. More stake means your opinion about which work was good counts for more, and therefore means more of the network's money flows where you point it.

So the token in your hand is two things at once. It is an asset with a price. And it is a share of a judgment about what artificial intelligence is worth paying for. The price is easy to see. The judgment is the part that quietly changed hands this week.

On 14 September, Yuma, a subsidiary of Digital Currency Group, reported more than two hundred million dollars in tokens staked to its validator, which it says makes it the largest validator on the network running its own hardware, less than two years after the company was founded.

Read that against the mechanism rather than the number. A validator here is not a passive machine stamping transactions. It evaluates other people's model outputs and submits its own scores, and those scores route the emissions. Yuma says as much in its own materials: it grades performance firsthand, and its proprietary research is what lets it commit stake toward the sub-networks it judges most productive. That is an honest description of the job. It is also a description of an opinion that has never been published.

The route the money took matters more than the size. In July, one exchange with a reported forty million users switched on staking through this same validator, so a customer can now stake with two taps inside an app they already have. An institutional custody provider added the same thing for its clients. Others followed. None of those flows involve choosing a scoring policy, because none of those interfaces contain one. There is nothing to choose. You press stake, you see a yield, and a vote you did not know you owned is cast on your behalf, every few minutes, forever.

The word for this is delegate, from the Latin delegare, to send someone in your place carrying your authority. The envoy is supposed to be carrying instructions. Here there are none, because nobody was ever asked for any.

The other half of the week ran in the opposite direction. The network's token is now live one to one on Robinhood Chain, bridged by an operator called ForeverMoney over Chainlink's cross-chain messaging system, following the same move onto Base in August. Over the weekend a new robotics sub-network launched itself on Base through the same plumbing.

The bridge operator is unusually clear about what this is, and the clarity is worth repeating because most such announcements are not clear at all. The bridged token is wrapped access, not a migration. The network's own consensus is untouched. What sits on the other chain is a claim, and the risk of holding it is the bridge's code and the bridge's willingness to redeem.

Follow what happens to the vote. The bridge's own pitch, written plainly on its blog, is that you should move your tokens to Base and put them into a liquidity pool, where a different chain's weekly emissions will pay you far more than staking at home would. Its table compares the two directly: roughly one hundred and thirty-seven percent a year in the pool against about sixteen percent from staking on the network itself. Take that trade and your tokens are now a price sitting in a pool on another chain. They judge nothing. The scoring table at the fair has one fewer badge on it, and the ledger back home has slightly fewer opinions to weigh.

There is a sharper version of this in the bridge's own developer documentation. To bridge a staked position in that robotics sub-network, your stake must first be moved onto the bridge vault's own validator key, and stake held with any other validator cannot be bridged until you move it. So making your position portable requires, as a precondition, consolidating your judgment onto one party's key. The vote is not destroyed. It is gathered.

And the destination chain rhymes, which is the part worth sitting with. On Base, the pool rewards are decided by a weekly vote among people who have locked the local token, and the bridge openly paid about thirty-eight tokens in what that ecosystem calls bribes so that voters would direct emissions to its pool. In one round the pool won about eight tenths of one percent of the vote and collected roughly eighteen thousand dollars for the people supplying it. Two chains, two voting systems over where new money goes, and in both of them the vote turns out to be the liquid, tradable, purchasable part.

This is not a crypto problem. It is the separation of ownership from control, described for companies by Berle and Means in 1932, when they noticed that American corporations had become so widely held that the owners were a scattered crowd and the actual decisions had migrated to a small group of managers nobody had really chosen. The modern version is the index fund: hundreds of millions of people own shares, almost none of them vote, and a handful of asset managers cast those votes at every company in the economy.

The shape is always the same. A financial instrument bundles a cash flow with a governance right. The cash flow is legible, comparable, and easy to shop for. The governance right is illegible, unpriced, and costly to exercise. So the market optimizes for the first and gives the second away for free, and it accumulates wherever the plumbing happens to deposit it.

What makes this instance sharper than the index fund is what the vote decides. A shareholder vote settles who sits on a board. This vote settles which artificial intelligence work is good enough to be paid for, on a network whose entire pitch is that no single company gets to make that call. The network's defenders point at open participation: anyone can run a validator, anyone can register a sub-network, no permission required. That is true, and it is not the same as anyone actually doing it. Roughly seventy-two percent of staked tokens still sit in the network's oldest, safest staking pool rather than pointed at any particular sub-network at all, and the top ten sub-networks collect a little over half of all emissions, with three of them run by a single team.

The honest version cuts both ways, so here it is.

Delegation is not theft and it is not new. Professional validators exist because scoring AI outputs well is real work, and a large operator with staff and research probably does it better than a person with a phone would. There is no evidence anyone behaved badly this week. Every party involved described what it was doing accurately, which is more than this field usually manages.

The uncomfortable part is narrower than a scandal and harder to fix than one. A network built on the claim that strangers can check each other has, at the layer that decides who gets paid, almost nothing to check. The scoring policy is the product, and it is proprietary. Nobody publishes a weight-setting policy in a form a delegator could read, disagree with, and leave over. And every new on-ramp built this month, the exchange button, the custody product, the bridge, the pending exchange-traded wrapper, makes the asset easier to hold and the judgment easier to skip.

It is worth noticing what else happened this week, in a different corner of the same field: an analytics firm published the first serious attempt to measure how much of the traffic on agent payment rails is actually agents, and the honest answer came back as a range rather than a number. That is the same instinct pointed at demand. Nobody has yet pointed it at governance. The open question is small and answerable: does any large validator on any of these networks publish what it scores and why, in enough detail that a person could disagree with it?

Until one does, the most decentralized thing about the network is the money, and the most concentrated thing about it is the opinion.

Two things to watch. Whether any large validator publishes a scoring policy specific enough to disagree with. And whether the Bittensor community's own gathering in Montreal on 28 and 29 September produces anything technical on this, since the last long quiet stretch on the research side makes that event the likeliest place for the question to be asked out loud.

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

Decentralized AI, layer by layer.

Dastan,