Most days on Bittensor there is a change to report. Today there is a queue.
Three separate upgrades to the network's core software have been built, tested and formally put forward for release, and all three are sitting still, waiting on the same missing approval. At the same time the largest change of the month has actually gone live, and it quietly rewrites the job description of every validator on the network. Both stories are about the same thing: who gets to decide where money goes.
Root Reborn is live, and you can now follow someone else's picks
Start with what Bittensor is, because the rest only makes sense on top of it. Bittensor is a network of independent teams, called subnets, each trying to do one useful job with artificial intelligence. One generates 3D models. One serves chatbots. One rents out computing power. There are more than a hundred of them, and they compete for a share of the network's currency, TAO, based on how good a job they do. The machines that judge the work and hand out the scores are called validators.
If you hold TAO, you can stake it, meaning you lock it behind a team and earn a share of what that team is paid. The catch has always been the same one you meet with any investment: to stake well, you have to know which of a hundred and thirty teams is worth backing. Most people do not, so most people used Root, the default pool for anyone who did not want to choose.
Root Reborn is the rebuild of that pool, and on the 28th of August the network's founder, known as Const, announced on a community call that it had gone live within the previous two hours. What changed is that validators can now choose which subnets they sell Root staking yield into. That chosen set is called a basket. Rather than researching a hundred teams yourself, you subscribe to a validator's basket and let a group that has watched the network for years do the picking.
The closest everyday comparison is the difference between buying individual company shares and buying a fund. You are still exposed to the market, you have simply hired someone to decide the mix. Const framed it in almost those terms: a lower effort way in for people who do not want to do their homework on more than a hundred teams, but not a risk free one, because you still end up holding subnet tokens that can go down as well as up. He also said something worth noting for anyone who likes clean labels: a basket is not itself a token, though he said it can be what he called pseudo tokenised.
Two details came with it. First, the way yield is described on the network was standardised, so the same holding no longer gets calculated differently by different tools. Second, and Const said this plainly, the basket curation tooling that teams like Crucible and Mentat Minds had built for themselves now exists at the protocol level, in a standard format everyone can use. Work that was previously a private advantage is now a shared feature.
There is a guardrail attached. No single basket can put more than one sixteenth of its weight, a little over six percent, into any one subnet. The entire point of the rebuild was to stop everybody's stake piling into whichever subnet happened to be fashionable, so a ceiling on concentration is the mechanism that makes the promise real.
One honest caveat. The change that switches on a manager's ability to choose subnets was written as release 449, and the code record shows 449 never went out as a release of its own. It is being folded into a larger bundle numbered 500. So the founder describes baskets as live on the network, while the release notes show the specific selection rule still travelling in a bundle that has not shipped under that number. We cannot reconcile those two from here, and rather than pick one we are telling you both and who said which. If you are watching this closely, the thing to check is what your own validator's basket actually holds, not what any release note says.
The plumbing that went in behind it
Root Reborn also produced its first emergency repair, which is a good sign rather than a bad one, because it means real money is moving through it.
Developers found that if one subnet inside a basket had no usable liquidity, meaning there was not enough of its token available to trade in and out, a claim on that basket could fail outright, and healthy holdings sitting alongside the broken one could not be withdrawn either. One jammed till was closing the whole supermarket. A hotfix, runtime 451, was merged to make sure a pot that has run dry, or one that has grown unusually large, cannot block everyone else's payouts. The same work added automated tests proving that the accounts run by subnet owners do not accidentally burn their own rewards.
Two smaller pieces of engineering also went into the code this week. One tightens up what are called precompiles, which are built-in shortcuts letting smart contracts reach network features like staking without writing that logic themselves. From now on a shortcut only runs when it is genuinely being called at its own address, rather than being triggered by something pretending to be somewhere it is not. Think of a staff-only door that only opens when a member of staff is actually standing in front of it. The other gives a computer joining the network a trusted starting point written into its configuration, so it can catch up quickly instead of grinding through every block of history since the beginning. It is the difference between a new employee reading the handover notes and reading every email the company has ever sent.
And here is the word that matters for all of it: merged. Merged means the fix exists in the code. It does not mean it is running on the live network. Which brings us to the queue.
What is coming next
Every machine running Bittensor follows the same rulebook, and that rulebook is updated in numbered packages called runtime upgrades. Changing it is not something one person can do on a Tuesday afternoon. An upgrade needs two separate approvals before it takes effect: an automated build system submits the first half, and a small group of trusted human signers, known as the triumvirate, has to provide the second. Two keys, two turns, like a safety deposit box.
Right now, three upgrades are stacked behind that second turn. Release 448 was built and submitted on the 21st of August and is still unsigned. Since then 450 and 452 have been built and submitted too. Nothing is broken, nobody has lost anything, and holders do not need to do a thing. But a growing queue of finished, unshipped work is worth understanding, because it is the mechanism by which everything above eventually reaches you or does not.
What is waiting in there is not trivial. The Root Reborn hardening that stops one stuck subnet blocking everyone's payouts is in that queue. So is the change that gives Root baskets one official price, recorded on the network itself, so that every wallet, explorer and dashboard converts to the same figure instead of each doing its own arithmetic and disagreeing about the same holding. That one changes nothing about how much anyone owns. It is plumbing. But disagreeing dashboards are how people talk themselves into believing something has gone wrong when it has not.
Who does the delay help? Arguably everyone, in one specific way. A deliberate pause before changing the rules that govern real money is not a bug, it is the reason two signatures exist at all. One compromised build machine cannot rewrite how the network pays people. A queue is what caution looks like from the outside.
Who could it hurt? Anyone waiting on a fix that is already written. And there is a second, less obvious cost. The longer the queue, the larger the single jump when it finally clears, and large jumps are harder to reason about and harder to undo if something behaves unexpectedly. You can see the version numbers already getting tangled by the wait: 449 was never shipped and is being folded into a bundle numbered 500, and 451 never made it out either, so its work has been renumbered to ride inside 452. None of that is alarming on its own, but it does mean a version number on Bittensor is a plan rather than a promise, and you should not treat one as a delivery date.
The honest summary is that none of this is live yet and any of it could still change. What is worth watching over the next week or two is simply whether the queue clears in one go, or whether the numbers keep getting rebundled while the pile grows.
A loophole in the payout maths, and the small repairs behind it
One proposal in the queue deserves explaining properly, because it is the kind of thing that quietly costs ordinary people money when it is left alone.
When you hold a share of a basket, that share is priced as though the entire holding were being sold at once. But the pools underneath sit on a curved pricing formula, where selling a small slice fetches a better price per unit than selling the whole thing would. Picture a shop where the first few items are cheap and the price gets worse the more you buy at once. That gap between the small-slice price and the everything-at-once price is real, and someone could exploit it: deposit into a basket, immediately claim a small slice at the favourable rate, and walk away with slightly more than they put in. The difference comes out of everybody who stayed.
The proposed fix closes that gap. It helps everyone who holds a basket and intends to keep holding it, and it costs nothing to anyone behaving normally. The thing to keep an eye on, as with any change to payout arithmetic, is that ordinary claims may settle at figures that are slightly different from what people got used to, and a small unexplained change in a familiar number is exactly what makes people worry. It is not live yet.
A handful of smaller repairs sit alongside it. The chain supports standing orders that stay open for six months until filled or cancelled, but the internal record other systems need in order to read such an order was being thrown away after only seven days, so an order could still be live while its paperwork had already been shredded. A proposal raises that retention to the full one hundred and eighty days. Another keeps a forwarding note alive when someone swaps to a new account key, so automated contracts can still find where the money went instead of following a trail that ends nowhere. Another stops a developer tool cheerfully approving a rehearsal of a payment that the wallet cannot actually afford once the fee and the minimum balance are counted. One trims wallet balances on screen from nine decimal places to three, which changes no accounting whatsoever and removes a wall of digits from the screen most holders look at most often. And the wrapped version of TAO, a contract that holds your TAO and hands you a standard token representing it, rather like a cloakroom ticket that every venue knows how to accept, is being moved into the main codebase so it is reviewed alongside everything else.
Subnets racing to serve a brand new model
An AI model called GLM-5.3 was released this week with its weights published openly, meaning the model's files were put out for anyone to download and run rather than kept locked behind one company's service. Two Bittensor subnets picked it up almost immediately.
SayGm, subnet 28, says it was serving the model on launch day, hosted through a partner called KubeTEE that runs it inside hardware designed so the operator cannot read what users send. That is the latest in a run of day-one launches from that team. Engy, subnet 53, is also serving it, and made the sharper claim: roughly 100 tokens a second, which is about 75 words of output per second, at 98 cents per million words of input and $3.08 per million of output. The official service and the reseller OpenRouter charge $1.40 and $4.40 for the same thing. That is roughly a third cheaper.
This is the clearest version of Bittensor's pitch you will see. A brand new model, available on a decentralised network the same day the big providers have it, at a noticeably lower price. If you use AI tools at all, that is the part of this ecosystem that touches you directly. The open question is what the cheapness is made of. It could be genuine efficiency from many competing operators undercutting each other, in which case it holds. It could also be underwritten by the network's own token rewards, in which case the price is a marketing budget and the number to watch is what it looks like in six months rather than what it looks like this week.
A smaller piece of news fits here too. SOMA, subnet 114, builds matching algorithms, the sort of software that decides which video you see next or which person a dating app shows you. It says SayGm has started using SOMA's algorithm inside its own app. Subnet teams almost always build alone, so one team's output being good enough for another to depend on is a modest but genuine signal. SOMA also says an early access version of its own product is coming, with no date and no description of what it does for a user, so treat that as an intention rather than a launch.
The subnet turning a sentence into a playable 3D world
The most interesting thing said publicly this week came from Ben of the 404 team, subnet 17, which generates 3D models.
The team used to produce what are called Gaussian splats, a way of representing a 3D scene as a cloud of coloured blobs. Ben said the results looked visually stunning and studios could not use them, because game engines are built around meshes, models made of flat triangles, and splats do not slot into the pipelines and file types the industry already runs on. So they rebuilt around meshes. Their model is called Titan, built by miners on top of an open source model called Trellis 2, and because their internal competition requires commercially usable output, the miners stripped out the parts that blocked commercial deployment and rebuilt them, which by Ben's account also improved quality and fixed holes and geometry errors along the way. He says Titan produces a 3D model in about 30 seconds, against several minutes for closed competitors like Meshy.
On top of that sits Atlas, which does the unglamorous work that makes an asset actually usable in a game: shrinking a heavy model down to something light enough for a phone while painting the fine detail back on with what are called normal maps, which add apparent surface texture without adding any real geometry, and setting the pivot point so a weapon sits in the hand and a helmet sits on the head rather than floating beside it.
The reason to care is that someone is paying for this outside crypto. A mobile game called Colony, by a studio named Parallel, is in early play testing using Atlas and Titan to let players generate their own weapons, helmets and armour in real time. The clever part is that the wait is hidden inside the game: the player describes what they want to a character, and the item is finished by the time they have walked across the map. Ben said Parallel report that crafting in the player's own style is the single most liked feature in testing, that he believes the game launches within weeks, and that Google cited it early as one of the most exciting AI games around. He also said two people used the team's technology to build a game that won the inaugural Chroma Awards, generating over twelve thousand assets. All of that is the team's own account, not something we have independently checked.
Compute made out of energy nobody could use
The other substantial interview this week was with Josh of subnet 110, which is trying to build what he calls new permanent compute on entirely renewable energy, starting on farms in the United Kingdom.
His diagnosis of the problem is worth repeating because it is not the one people expect. Everyone knows AI uses enormous amounts of power. What Josh says gets missed is how poorly the grid carries power in the other direction. A large field of solar panels can sit behind a local substation too small to accept what it produces, so only a fraction gets through. Wind turbines get switched off when it is too windy. The power is simply dumped. His pitch to a farm owner with wind, solar, hydro or an anaerobic digester is to point that wasted generation at computers instead.
The numbers he gave: selling back to the grid earns roughly 8 to 12 pence per kilowatt, which puts the payback period on a turbine or a digester at 12 to 15 years. Running compute, over the past year with some clients, has generated 60, 70, 80 pence and in some cases a pound per kilowatt. His worked example is a client who spent five million pounds on wind turbines and a further million on servers, where he says the payback could fall to four or five years including the extra spend. These are his figures, from his own business, and nobody outside has checked them.
His argument for why this is different from other compute subnets is about reliability rather than price. Existing decentralised compute, he says, is good for cheap, fragmented or short jobs, but a company that needs a hundred machines dependably for six months cannot get that today, because the person renting to you may themselves be renting from someone else temporarily. To qualify as a miner on his subnet you need physical servers, verifiably green power, a ten gigabit connection that is as fast uploading as downloading, and actual people on site to fix things. That last requirement is the honest limit of the whole idea: green energy is currently verified by Josh's team driving out and looking at the site. That works for a few multi-million pound deployments in Europe. It is not obvious what it becomes at scale.
Everything else worth knowing
Teutonic, on subnet 3, has begun training a model called Teutonic-II with around 110 billion parameters, the internal settings a model adjusts as it learns and a rough proxy for how big it is. The run is permissionless, meaning anyone with suitable hardware can join and contribute computing power rather than the work happening inside one company's data centre. The team describes it as rough and early. Training a model that size on strangers' machines is one of the hardest claims in this whole field, so it is worth following, and worth judging on results rather than on the announcement.
Tatsu, subnet 84, spent the week building the money layer under its AI products: a single prepaid credit covering training a model, running one, and using its assistant, bought by card or with tokens. The team's own framing is the interesting bit, that taking the payment is the easy part and the hard part is metering, making sure nobody quietly uses more computing power than they have paid for.
Masx.ai, subnet 104, has named a time for the beta of its version three: Monday at nine in the evening, central European time. The project describes its goal as turning the flow of information across Bittensor into signals and forecasts. A first firm date is progress after weeks of description, but nothing is usable yet and a beta means an early version expected to have rough edges.
And Ditto says it has shipped a search tool that returns decisions rather than documents. Most search finds where a word appears, so asking what your team decided about pricing in March returns a heap of files that happen to mention pricing. Ditto's version runs three approaches at once: plain keyword matching, embeddings, which match on meaning rather than exact wording, and a map of how records connect to each other.