No change to Bittensor itself went live in the last two days. What happened instead is that a fifth finished upgrade joined a queue of upgrades nobody has switched on, and a set of proposals arrived, one of which would let people do something with a subnet's token that the network has never allowed.
Nothing was switched on, and five upgrades are waiting
Bittensor is one program running on thousands of computers at the same time, and those computers have to agree on the rules or they stop agreeing on anything at all. The shared rulebook is called the runtime, and it is updated in numbered releases: 447, 448, 450, 452 and now 453. Swapping the rulebook out is not something anyone does casually, so it takes two separate approvals. The automated build system gives the first, confirming that the code about to be installed is exactly the code that was reviewed. A small group of trusted signers, known in Bittensor as the triumvirate, gives the second.
The first signature is attached to all five of those releases. The second is attached to none of them.
Release 453 is the newest arrival. It was merged this week, packaged, tagged, and given its first signature, and it is now sitting in the same waiting room as the four before it. Its contents are mostly tightening rather than new features, which is worth understanding because it tells you what the people maintaining the network are currently worried about.
One part of it deals with random numbers. The network needs unpredictable numbers for anything that has to be a fair draw, and it gets them from an outside service that publishes a fresh random value at regular intervals, a bit like a public lottery machine that anyone can watch. Release 453 binds each of those values to a checkable fingerprint, so that a value which has been tampered with in transit no longer looks like the real thing. Another part closes a gap in the permission system, where authority handed to one account could be used to reach actions it was never given. A third strengthens the protection on coldkey transfers. Your coldkey is the key that actually holds your funds, the equivalent of the deed to a house, as opposed to the day to day key you use for routine work, so anything that makes it harder to move funds out from under someone is worth having.
The more interesting question is what a queue five deep actually means. There are two honest readings and this week does not settle between them. The generous reading is that each of these releases touches money belonging to real people, that the second signature is the last point at which anyone can say no, and that the people holding that key are right to be slow. The less comfortable reading is that finished, reviewed, tested work is piling up and none of it is reaching anyone, and that a process which depends on a small number of individuals being available is a process with a single point of failure in it.
There is also a practical cost to batching. If four rulebook changes go live on the same day and something behaves oddly the following week, working out which of the four caused it is much harder than if they had gone out one at a time. It is the difference between a landlord fixing one thing a month and a landlord doing eight repairs in a single afternoon: cheaper in one sense, much harder to trace when the taps start making a noise.
In fairness to the slow route, this week produced a good argument for it. Release 453's permission fix, the one that closes the loophole where authority given to one account can be stretched into authority over another, turns out to break a working escrow contract already running on the network. Escrow is the arrangement where money sits with a neutral party until both sides have done what they promised, the way a deposit on a house is held by a solicitor rather than handed straight to the seller. That contract moves staked TAO on a depositor's behalf, using exactly the pattern the new rule now blocks. A patch has been proposed to allow that one specific action back through the filter. Nobody lost anything, because 453 is not live. Had it been signed the day it was tagged, a live contract holding other people's deposits would have stopped working.
The last change that genuinely did land was a week earlier, when Root Reborn's beta baskets went live. Root staking is the plain option on Bittensor: instead of choosing which of the hundred plus projects on the network to back, you park your TAO in a shared pool. Under the new arrangement the yield from that pool gets directed into subnets chosen by whichever validator you are staked with, and that curated selection is the basket. A validator is one of the machines that scores everyone's work and votes on who gets paid, usually run by a team that has watched the network for years. Const, who founded Bittensor, described the point of it on a community call as a lower effort way in for people who do not want to do their own research on more than a hundred teams. He was careful to add that it is not risk free, because you are still holding subnet tokens with all of their upside and downside. There is already a proposal in the queue to harden how those baskets handle claims when a subnet's liquidity runs thin, which is the ordinary sound of a new system meeting real conditions.
What is coming next
The proposal worth your attention is pull request 3135, which would add a way to take a dated position for or against a subnet's token.
Start with what a subnet token is. Every project on Bittensor, and there are well over a hundred, has its own token, called alpha. Holding a subnet's alpha is the way you back that particular project rather than the network as a whole. Today the only thing you can do with a view about a subnet is buy its token or sell it. If you think a team is doing excellent work, you buy. If you think a subnet is coasting on an announcement it will never deliver, and you do not already hold it, there is nothing at all you can do. Your opinion never reaches the price.
The proposal would add two dated positions. A long gains if the subnet's token rises. A short gains if it falls. Both expire after thirty days, and you cannot hold either one open indefinitely.
The mechanism matters, because it is unusual. Each subnet's token trades against TAO in a pool, which you can picture as a two sided till with TAO in one drawer and the subnet's token in the other. The price is simply the ratio between the two drawers, and ordinary buying and selling moves that ratio, which is why a large trade moves the price. Under this proposal, opening a position lifts a matched slice out of both drawers at once. Because it takes the same proportion from each side, the ratio does not change and the price does not move at the moment the position opens. One half of that slice is then traded through the pool in the normal way, and at close the whole thing is reversed. Nothing new is created and nothing is destroyed. The pool only ever gets its own liquidity back.
Who does this help? Anyone with a negative view, for the first time. There is a reasonable argument that a market where only optimism can be expressed is not really a market, it is a queue of buyers, and that prices which can only be voted up are prices that stay wrong for longer. Subnets that make big claims and quietly miss them currently face no financial contradiction from anyone. This would give them one. It also helps people who want to express a view on a subnet without committing to hold its token for months.
Who could it hurt? Small subnets with thin pools, and this is the part worth thinking hardest about. A position lifts liquidity out of the pool for as long as it is open, and a pool with less in it moves further on every ordinary trade. On a large subnet that is a rounding error. On a small one it could make the price jumpier for everyone, including people who have no interest in derivatives and just hold the token.
There is a second concern that is specific to how Bittensor works. On most networks a token price is a scoreboard and nothing more. Here it is also an input: the price of a subnet's token feeds into how much newly created TAO that subnet is allotted. That means sustained downward pressure on a small subnet's token is not only unpleasant for holders, it can reduce the income the team relies on to keep building. Whether organised selling could be used deliberately against a subnet, and whether the thirty day expiry is short enough to stop that becoming a campaign, is the question this proposal has to answer. The expiry helps, because every position is forced closed on a known date rather than sitting on a token indefinitely. Whether it helps enough is not something anyone can claim to know yet.
And there is the ordinary risk of new financial code holding real money. This would be a brand new component of the network, given its own slot in the rulebook, doing arithmetic on live pools from day one. It is proposed against release 453, which, as above, has been waiting on a signature along with four others.
None of this is live. It is a proposal, it can be changed, argued down or abandoned, and on the current evidence it would not go live quickly even if everyone agreed with it tomorrow.
Three smaller proposals are worth knowing about. One would extend how long a standing buy or sell instruction survives, from seven days to a hundred and eighty. A limit order is an instruction to trade only if the price reaches a level you name, and a linked order is a pair of them where one firing cancels the other. Expiring after a week means anyone using them has to keep re-entering the same instruction, which is both annoying and a way to end up unprotected because you forgot. Another would fix a quiet trap in the permission system: if you hand a second account the power to manage your staking, that account currently cannot perform the one specific action that makes root rewards actually arrive, so the rewards silently do not materialise. The proposal treats a staking delegation as covering the thing a staking delegation was obviously meant to cover.
Separately, and not yet a formal proposal at all, an idea called fractal emissions is circulating in the community. It would let a subnet voluntarily accept less than the full share of new TAO it is entitled to. Because that new TAO is shared out to holders of the subnet's own token, taking less would mean fewer new tokens created, and therefore less dilution of the people already holding. It is an unusual idea, in that it asks teams to turn down money in exchange for their existing holders being watered down more slowly, and it is at the discussion stage only.
The subnet news: money actually coming in
Bitcast, subnet 93, says it spent seventy five thousand dollars of customer revenue in August buying its own token back off the open market. A buyback means the team takes money earned from real customers and uses it to purchase the token its network runs on, which takes some supply out of circulation. It is the clearest signal a subnet can send that it has income rather than only ambition. This was its biggest month and more than double July.
Bitcast connects brands with creators on YouTube and X. Its founder Tom explained the point of it on the Hash Rate podcast this week: normally a brand and a creator find each other through direct messages, haggle over a rate, sign something, and then the creator chases payment and tries to prove the work happened. Bitcast publishes the brief instead, any creator in the network can pick it up, an AI checks the finished content against every point in the brief, and payment lands automatically. Miss a point and you get nothing at all. On YouTube it pays on watch time rather than raw views, because watch time is harder to fake, and on X it works out who genuinely holds influence in a niche before anyone posts, by treating reposts and mentions as one account lending credibility to another, roughly the way Google originally ranked websites by who linked to them. That map is rebuilt every two weeks. Tom said the network has just under twelve hundred creators and a combined reach of around eleven million, with three separate maps built so far, three hundred creators in the Bittensor world, six hundred and forty five in prediction markets and four hundred and fifty two in the venues where people trade with leverage.
He also announced a structural change. Creators will no longer be the miners, the competing participants who earn the subnet's rewards. Instead the miners become niche marketing platforms, each bringing their own app, their own creators and their own paying customers, with Bitcast supplying the mapping and scoring underneath. The first outside team, Test Drive, is launching for app testing. That is a subnet trying to become an engine other businesses build on rather than a single product, which is a much larger ambition and a much harder one.
A well followed community account has separately claimed Bitcast's revenue has grown 4.7 times since February. That figure came from a commentator rather than the team, and it should be held more loosely than the buyback, which the team stated itself.
The other genuine revenue story is smaller and more telling. A company called Ditto pointed out that it is not announcing a partnership with anyone, it simply pays two subnets for work it already depends on: it stores its data on Hippius, where independent operators host files instead of one company's servers, and it sends its search queries through Desearch first. Both teams made a point of the distinction. A partnership announcement is a press release. A recurring invoice is a business.
Making AI cheap enough to be worth running
Several subnets are attacking the same problem from different directions, which is that running AI models costs money and most of that cost is wasted.
Verathos has switched on a system called Sleipnir on the main network, meaning it is in production rather than testing. It lets several computers pool their graphics cards, the chips that do the heavy lifting for AI, so that together they can run models too large for any single machine to hold. The team says it has already served several large open models this way. Think of it as four people carrying a wardrobe that none of them could lift alone.
Subnet 65, run by TPN Labs, has turned on a public scoreboard, a dashboard and a gateway two weeks after going live on the main network. The subnet pays people to shrink large AI models so they run cheaper and faster, and seventeen participants have squeezed one model down to a two gigabyte limit. The scoreboard matters more than it sounds: until now the judging happened out of sight, and a competition nobody can audit is a competition people eventually stop trusting. TPN has also opened a third contest, this time to fit a twelve billion parameter model into under eight gigabytes of memory while still scoring well across five standard tests. Eight gigabytes is roughly what an ordinary laptop has, which is the entire point.
Subnet 114, SOMA, has opened early access and says its software now sits behind GitHub Copilot, the coding assistant many programmers already use. When you ask Copilot a question it normally hands the model a large slab of surrounding code so it understands the context, and most of that slab is irrelevant. SOMA trims it before it is sent. Subnet 29's Cacheon reports a new speed record in its own competition to make models run faster without breaking them, and subnet 28's SayGm has added Google's newest fast model at just over eight percent below Google's own list price, which is the ordinary daily work of a reseller.
New arrivals, and one that lost its place
Subnet 111, the Claims subnet, has completed its main network launch and had emissions switched on. Emissions are the newly created TAO the network issues every few seconds and splits between the people doing the work on a subnet and the people backing them, so that switch is the moment a project stops being an experiment and starts being paid.
Subnet 59, Babelbit, went the other way. It has been deregistered, roughly a week after announcing a dubbing service for translating audio between languages. Deregistration means losing the slot on the network, and with it every bit of the newly issued TAO. Slots are limited and a subnet holds one only by staying valuable enough to justify it, which is deliberate, and which is also why a launch announcement and a lasting business are not the same thing. It is worth reading the two stories together, a week apart, as a corrective to any sense that going live is the finish line.
Elsewhere, Targon, which rents out computing power, gave free compute credits to the new subnet 91 team to help them launch, the first to come out of a training programme for new subnet teams. A group called Kusanagi has set up as a business layer around Bittensor rather than registering a subnet, bundling a validator, a token trading site and other services under one name. And subnet 18, Zeus, told holders that the run of selling in its token came from a single early backer who had paid the fee for Zeus to get its slot and received a share of the subnet's daily tokens in return, and who has now decided to wind down. Zeus says the selling connected to that is finished. That is the team's own account of it and there is no way to verify it from outside, but it is a more useful thing to say than nothing.
Showing your work
A quieter theme this week was subnets making themselves checkable. Subnet 107, Minos, now publishes a short fingerprint of the exact settings used in each scoring run, so nobody can adjust the rules afterwards and claim they were always that way, and the tasks being scored are now picked on the blockchain itself rather than behind closed doors. Subnet 6, Numinous, where people compete to forecast future events, has added an in-house reference forecast that competitors must beat before they earn anything, and says that bar will be raised over time. Subnet 63, Enigma, which works on quantum computing problems, has dropped the requirement to run mining software and configure a crypto wallet before you can take part, and now accepts submissions through a web page, which is an obvious kindness to a researcher who has an answer but no interest in plumbing. Subnet 54, Yanez, says its private sign-in has begun appearing inside two other projects, and reports that only two percent of people signing up needed help, which is the kind of unglamorous number that tells you whether something actually works. Subnet 48 has rewritten its whitepaper after nine months of changes made the old one inaccurate, which is housekeeping, but the honest kind.
Games, big files and farm electricity
The richest material this week came from two long conversations rather than announcements.
Ben from subnet 17, the 3D generation subnet, explained why the team abandoned an approach that looked spectacular. Their earlier output was visually stunning but studios could not use it, because it did not fit the file types and production pipelines that game engines already expect. So they rebuilt around meshes, the ordinary way 3D objects are represented, and their model now produces one in about thirty seconds where closed competitors take minutes. On top of that sits a layer that makes the output actually usable in a game: reducing the detail so a phone can render it, and setting the pivot point so a weapon sits in the hand rather than floating beside it. A mobile game in early play testing uses this to let players describe an item they want and receive it mid-game, with the generation hidden behind the walk across the map, and Ben said the developers report that in-game crafting in the player's own style is the single most liked feature in testing. The subnet's current competition has moved on again, to generating 3D scenes as editable code rather than as models, which look simpler but are far lighter and can be edited afterwards. A tool to turn a written description into a playable game is being built on top of that.
Separately, Josh from subnet 110 made an argument about electricity that is worth hearing even if you never touch Bittensor. The UK grid is poor at carrying power back the other way, so a large solar field can sit behind a substation too small to accept it, and wind turbines are switched off when it is too windy. That generation is simply dumped. His subnet puts small data centres on those sites instead. He says generators currently earn something like eight to twelve pence per unit selling back to the grid, and that running compute on the same power has produced sixty to eighty pence and occasionally a pound, which on his worked example turns a fifteen year payback on a wind investment into four or five years including the extra cost of the servers. Those are his numbers, from his own clients, and nobody outside has audited them. He also drew a distinction worth keeping: most compute you can rent in this world is itself rented from someone else and may not be there next month, whereas an enterprise that needs a hundred machines reliably for six months needs something that physically exists and stays put.
Subnet 105, Beam, is wiring itself into Hugging Face, the site where most of the world's freely available AI models are kept. Beam is not offering to store them: the job it is taking is movement, shifting very large files across many machines at once rather than down a single connection, which is the difference between one person carrying boxes and twenty people forming a chain.
Finally, one claim to file carefully. Siam Kidd, an investor who spends much of his time meeting subnet teams, says that in fifteen to twenty of them he has seen the same hidden problem: an early employee who was paid or overpaid, gradually stopped doing the work, and cannot easily be removed. That is one person's account of private conversations, and there is no way to check it. It is included because if it is even partly true it describes a drag on delivery that no announcement will ever mention.