Today was a quiet day on Bittensor, and it is worth saying so rather than dressing it up. One version of the network's own software went out, and it was a clear-out rather than a new rule. Everything else was repairs, proposals and subnet announcements. The part that deserves your attention is the queue: work that is written and agreed but not yet switched on, including a change to who is allowed to take a subnet away from its owner, and a proposal that would change how easily anyone can move from one subnet to another.
What went live
Bittensor runs on a shared rulebook, and that rulebook is a piece of software called the runtime. Changing it is not like updating an app on your phone. A new version gets built and tagged, and then it needs two separate approvals before the live network, which is named finney, will run it. The automated build system gives one. A small group of trusted signers, called the triumvirate, gives the other. Until both are in place the code exists in public but the network is still following the old rules.
Version 445 is the one that is actually finished and out, and it is maintenance rather than new rules. Transactions that are guaranteed to fail are now turned away before they are packed into a block, so they stop taking up room and burning fees for nothing. Empty and out of date entries are deleted from the chain's storage. Picture a post room that has stopped stamping and loading letters addressed to houses that do not exist, and has finally cleared out the drawers of paperwork nobody is ever going to ask for. Nobody's rewards change. The network simply carries less dead weight, and keeping storage small is one of the dull things that decides whether a chain is still usable in five years.
Two further versions, 446 and 447, are tagged and sitting with one signature each, waiting for the second. That is why several of the changes below are described as merged rather than live. Merged means the code has been accepted into the project. Live means the chain is actually running it. On Bittensor those are two separate moments, and the gap between them can be days.
Several of the repairs waiting in that queue concern Root Reborn, the change that stops rewards from staking to Root, the network's default pool, being paid out and reinvests them instead. Staking means locking your TAO behind a participant who does work for the network and taking a share of what they earn. Switching Root Reborn on means moving every existing record into a new format, and that move was originally written to happen inside a single block, one twelve second step of the chain. An entire office move crammed into one lunch break tends to go badly, so the move is now spread across many blocks. Separately, the route that let someone deposit straight into a Root Reborn holding, which the code calls a basket, has been temporarily closed while its storage cost is checked, because each direct deposit can spread across a validator's full list of scores and leave a new permanent record behind.
A few smaller fixes are worth a line each. On subnets where miners, the workers who actually produce the output, send in their work sealed with a time lock so nobody can copy it before an agreed moment, malformed packages used to be stored anyway and only found to be broken when the network tried to open them. They are now rejected on arrival. A related fix on Affine, subnet 120, stops the chain throwing a sealed record away entirely when the opening step fails, which previously left a miner who had paid a registration fee and submitted correctly with nothing recorded at all. And the counter that tracks which group of signers is currently allowed to declare a block final, meaning settled for good and impossible to undo, has had a bookkeeping correction, because when that counter and the real group drift apart in the stored history, a computer joining the network cannot verify what it is being handed.
One merged change alters behaviour rather than tidying it. Validators, the participants whose job is to check miners' work, build up a running score for each miner over time, which the code calls a bond. Think of it as a reference that gets stronger the longer someone keeps doing good work. Liquid Alpha 2.1 lets each subnet decide for itself how quickly that reference builds and fades, instead of every subnet living with one shared setting. A subnet whose work changes hour to hour wants a short memory. One judging slow research wants a long one. It is not live yet, and the real effect will depend entirely on what individual subnet owners choose to set.
What is coming next
The change most worth understanding today is about who can take a subnet away from its owner. A subnet is a self-contained marketplace running on Bittensor, with its own job to do and its own token, called alpha. Somebody owns each one, and ownership is not merely a title: the owner sets the rules of that marketplace and takes a cut of what it pays out. Bittensor deliberately allows ownership to change hands, and the way you make a claim is through conviction, meaning you lock up a large amount of that subnet's alpha for a period as proof that you are serious. The locking is the whole point. Anyone can say they would run a subnet better. Locking real value for a long stretch costs something.
The merged change, which forms part of runtime 447 and is therefore not yet live, tightens that test considerably. A challenger now has to be a single account, called a hotkey, whose own locked conviction is more than 18 percent of the subnet's eligible alpha, a figure worked out by taking the alpha in circulation and subtracting the portion the protocol holds and the portion that has been burned. The old test could be satisfied by adding people together. The new one cannot. One identifiable party has to put up close to a fifth of the subnet on their own.
A second proposal cleans up the accounting behind that test, and it is less dry than it sounds. Runtime 443 could roll one person's lock forward in time, apply only that person's change to the shared total, and then mark the shared total as up to date. The result was a running figure that quietly stopped matching the individual records underneath it. The fix repairs the corrupted state, makes ownership changes update the individual records properly, and requires each challenger to clear a 10 percent conviction threshold independently. When the number that decides who controls a marketplace is known to have drifted, correcting it is not housekeeping.
So who does the higher bar help? Anyone who wants a subnet they use or hold to be stable. A takeover is disruptive: the rules can change, the payout formula can change, and the people mining there built for the old regime. Making a takeover expensive and attributable to one named party makes it far harder to assemble a surprise raid out of a loose crowd. And who could it hurt? The same rule cuts both ways. If an owner goes missing, stops maintaining their subnet or behaves badly, a community that wants them out can no longer pool its holdings to do it. It now needs one member wealthy enough to hold nearly a fifth of the subnet alone. That is a small group of people, and on most subnets it is a group that skews towards those who were already early and already large.
The other thing worth watching is the denominator. The threshold is a percentage of eligible alpha, and eligible alpha is a calculated number with burned and protocol-held tokens subtracted from it. Any rule written as a percentage of a moving figure invites the question of who can move that figure, and how cheaply. We are not saying it can be gamed. We are saying it is the sort of thing to keep an eye on rather than take on trust. None of it is live: it sits behind that second signature and could still change.
The other proposal an ordinary holder would actually feel is about moving between subnets. Every subnet has its own token, so shifting from one to another means selling out of the first and buying into the second, with TAO in the middle. The awkward part is not the two steps, it is that you have to state the size of the second step when you sign it, before you know what the first one produced. It is like being asked to write the cheque for your next house before your current one has sold, and being told the number cannot be changed afterwards.
The proposal introduces linked orders, where the second order is sized as a fraction of whatever the first one actually produced. The example the developers give is a rotation: sell my subnet 7 alpha, then put the TAO that produced into subnet 12. The mirror case is buying into a subnet now and arranging a sale on exactly the alpha that purchase created, rather than on an amount guessed in advance.
This is straightforwardly good for someone who does not run software to watch the chain on their behalf. Right now the gap between the two steps is where you end up with a worse price than you expected, and closing it makes the network fairer to people acting by hand. The honest other side is that making rotation easy makes rotation more likely. If moving between subnets becomes one instruction instead of a chore, money may move faster and more often, and small subnets with thin trading are exactly where money rushing in and out swings a price hardest. This one is not live either and could change before it lands. Nothing here is a suggestion about what anyone should do with their own holdings.
One more proposal deserves a mention purely because of how blunt today's version is. Bittensor caps how many subnets can exist at once. When someone registers a new one while the cap is full, the cheapest existing subnet is dissolved in the very same transaction, with no warning to whoever owns it. The proposal would park that registration as a priced offer for roughly a day, giving the owner in the firing line a chance to respond. That is plainly fairer to a small owner who happens to be asleep. It also, unavoidably, gives notice to owners who are only sitting on a slot rather than using it, and opens a visible window in which everyone can react, so it may turn a silent replacement into a public scramble.
The rest of the queue is plumbing, and plumbing is where reliability comes from. Root claims currently make roughly thirty eight separate internal transfers each, and a proposal collapses that into one, which means lower cost and far less noise recorded on the chain. Developer tools would gain a direct way to ask which blocks are final, which sounds trivial until you consider that getting it wrong means telling somebody their payment has settled when it has not. Upgrade rehearsals on a copy of the live network would fail automatically if a single block takes four seconds to build or the chain stops making progress for twelve. And the command line tool would let people approve group-controlled actions from a hardware wallet or a browser extension instead of needing wallet files sitting on the machine.
The subnet news
The clearest theme today is training large models on computers scattered around the world rather than sitting in one company's data centre. This is genuinely hard. Training a model means thousands of machines constantly comparing notes, and inside a data centre they are wired together with connections far faster than any home broadband. Doing it across the open internet has been widely assumed to be too slow to be worth attempting.
Subnet 3, Teutonic, released Teutonic-I, its first model built from scratch rather than adapted from someone else's. It has 10 billion parameters, the internal settings a model learns while it trains, and it was pretrained across many separate machines. The claim comes from a weekly roundup shared by the Opentensor Foundation and relayed on X, so it is from inside the ecosystem rather than from an outside test. Subnet 9, IOTA, says its model Orion-16B has now been trained on more than one hundred billion tokens, the word fragments a model reads during training, and that quality is still improving rather than levelling off. Its training ran on strangers' spare graphics cards rather than matched hardware.
Subnet 96, Verathos, is described as working on something in the same family, which it calls mesh operators: several ordinary machines joining forces to run one large model, mixing professional cards, home gaming PCs and Apple laptops in the same pool. We would treat that one as unconfirmed for now. What all three would mean, if the results hold up under outside testing, is that the assumption that only companies with enormous data centres can build serious models has a hole in it. That is a large if, and the way to settle it is independent benchmarks rather than announcements.
The second theme is revenue, and how carefully you should read it. Bitcast, subnet 93, pays online creators for content that brands can verify people actually watched. It says revenue is growing and that it is spending about 75,000 dollars this month buying back its own token, having bought back roughly thirty three thousand dollars' worth so far. A buyback means a project uses money it has earned to buy its own token back off the open market, taking some supply out of circulation. It is a real sign that money is coming in. It is also a choice about how to spend that money rather than proof of profit, and the figures come from the project itself.
The same caution applies elsewhere. TrendBTC, an app where people guess whether Bitcoin, Ethereum or Solana will be higher or lower five minutes from now, says it handled more than 200,000 of those guesses in a single day, its busiest ever. Nobody outside the project has checked that, and the app pays people in its own token for taking part, so some of that activity is being bought rather than earned. Busy is not the same as valuable.
The most interesting revenue story this week never appeared as a post at all: it came from an interview. The founder of a Google Ads optimisation business has launched it as a subnet, named AdTAO in the video title and described as Ad Tower in the conversation. His product had a gap at its centre: given an advertising account, which of several possible changes will perform best. His reason for turning to Bittensor is worth repeating because it is the plainest version of the argument for this whole model. He says he cannot find, hire or afford the right data scientists, does not know which country or company they are in, and that a single hire is one bet he cannot verify. So miners compete instead: they predict what a campaign's performance will be one month after a single change, and they have to get both the direction of the effect and its size within a range. A correct direction with a narrower range scores higher, because that shows genuine understanding rather than a lucky guess. He says the subnet has been live for five weeks, that miners are achieving over 80 percent accuracy, and that around 104 miners competed last week. Those are his figures, given in an interview, and nobody outside has checked them.
The same interview opened with two analogies worth borrowing. Bittensor as Alphabet: a holding company whose value comes from the businesses under its umbrella, except that here the umbrella holds independent AI startups, each obsessed with one narrow job rather than trying to build one giant general model. And the mechanism as a school: subnets are the classrooms, validators are the teachers, miners are the students who hand in work, and the teacher ranks them, rewards the best and replaces the worst. The presenter put the current number of subnets at 128, which is about right.
Finally, two smaller things. Score, subnet 44, posted a demonstration of its studio taking a raw volleyball clip and labelling it by itself: marking what is in each frame and separating the objects into groups, the slow manual step that normally stands between raw footage and a dataset a model can learn from, and which can take a team weeks. The demo was honest about its limits. And the official command line tool would gain a short name for Blockmachine, subnet 19, so people can type a word instead of pasting a long address. That is convenience rather than capability, but friction of exactly that kind quietly decides what people bother to use.