Most of what happened today is housekeeping. Releases were renumbered, the wallet tool was told to stop printing nine decimal places, and a shortcut that lets a brand new computer join the network without grinding through years of history was accepted into the code. None of that changes anyone's money.
One thing does. The part of Root Reborn that decides what you are actually choosing when you stake has been switched on, and Bittensor's founder announced it on a community call yesterday, saying it had gone live about two hours earlier. It is worth explaining from the beginning, because it changes the job of an ordinary holder more than anything else this week.
What went live
Bittensor is not one project. It is a network of separate projects, called subnets, each paying computers to do a particular job: writing text, generating 3D models for games, forecasting markets, renting out spare processing power. There are more than a hundred of them. The network's own coin is TAO, and newly created TAO is handed out continuously to whoever supplies useful work and to whoever backs them. That continuous handout is what people mean by emissions. Think of it as a river of new coin flowing out every few seconds, split by rules the software enforces rather than by anybody's decision.
Staking is how a holder takes part without owning a single machine. You lock your TAO behind a subnet, or behind a validator, a validator being one of the referee machines that scores the work and decides who has earned what. In return you get a share of the flow. The awkward part has always been the choosing. Judging a hundred and thirty subnets is a research job, and most people were never realistically going to do it.
Root was the old answer. Instead of picking, you staked to the network as a whole and took a general share. It was simple, and it was also where nearly everyone ended up, which is its own problem: a system built to fund a hundred experiments kept funnelling the money into the one default option. Root Reborn is the rebuilt version, and the idea is that Root stops being a parking space and becomes a set of funds.
Here is the piece that went live. Every validator can now choose which subnets its Root yield is sold into. That chosen set is called a basket, and rather than researching the network yourself, you subscribe to a validator's basket and hand the picking to a team that has been watching these projects for years. The founder, who goes by Const, put it plainly on the call: it is a lower effort way in for people who do not want to do due diligence on more than a hundred teams. He also said it is not risk free, and that part matters. You still end up holding the subnet tokens the manager chose, with all the upside and all the downside that carries. A basket is not a savings account. It is closer to buying a fund somebody else assembles than to picking shares yourself, and a stranger's judgement is now part of your outcome.
There is a guardrail. No basket can put more than one sixteenth of its weight, a little over six percent, into any single subnet. That number is the whole point of the exercise turned into a rule. The original complaint about Root was concentration, everyone's money piled in one place, and it would have been easy to rebuild exactly that problem in a new shape by letting one manager pour a whole basket into a favourite. Six percent forces a spread at least sixteen ways.
Two smaller things arrived with it. The way yield is described on the chain was standardised, so that two dashboards stop quoting the same fund in different units and disagreeing about a number neither of them got wrong. And the basket curation tools that a few teams, Crucible and Mentat Minds among them, had already built for themselves now exist inside the protocol in one common format. That is usually a healthy sign: the network absorbing something its own users proved was necessary by building it twice.
A hotfix landed alongside the launch, and it is a good illustration of how new all of this is. 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 could fail outright and the perfectly healthy holdings sitting next to it could not be paid either. Picture a market stall that has run out of one item refusing to serve anybody in the queue. The fix teaches the code to tell a genuinely empty pot apart from an unexpected error, so one dead holding no longer jams everyone else's withdrawal. A second merged change adds tests proving that subnet owners' own accounts do not accidentally burn their rewards under the new system.
Now the honest caveat, and it applies to nearly every line above. Merged is not live. A change becomes part of Bittensor's codebase when the developers accept it, but it only reaches the running network inside a numbered release, and a release needs two signatures: one from the automated build system, which it produces by itself, and a second from a small group of trusted human signers known as the triumvirate. Until that second signature exists, the code is agreed rather than switched on.
That is where this week gets untidy. Release 449, which carried the subnet-choosing rule and the six percent cap, was written and never shipped, and its contents are now being folded into a much later bundle numbered 500. Release 450, which carries the standard way of recording fund prices, is built and waiting on its second signature. A branch for 451 was opened containing nothing at all, two release runs were cancelled, and the code has since been renumbered to 452, which is also waiting on signatures. Meanwhile the founder says baskets are live. Both of those can be true at once, because a feature can reach the network by a different route than the numbering implies, but we cannot resolve it from the public record, so we are telling you what each source says and who said it. If you take one thing from this, take this: when you read that something has been merged, that is a decision, not a delivery.
Two other merged items are worth a line each. Bittensor lets Ethereum-style smart contracts reach network features through built-in shortcuts, called precompiles, and a fix now insists that a shortcut only runs when the contract being executed is genuinely the shortcut itself, rather than one contract reaching for it while standing somewhere it should not. And new computers joining the network can now use a trusted starting point shipped in the network's own configuration file, so instead of replaying the whole history they can jump most of the way by following a trail of signed handover records. Neither changes what anyone holds. Both make the plumbing less likely to leak.
What is coming next
Nothing in this section is running. Each item is a proposal, and proposals get rewritten, delayed or dropped, so read this as a map of what is being argued about rather than a list of things that will happen.
The one to understand is a proposed fix that closes a way of quietly draining value out of a basket, and it is worth going slowly because the mechanism is not obvious.
A basket is a pooled holding. When you want to know what one share of it is worth, the chain values it as though the entire pot were sold at once. Now, the pots these tokens sit in do not have a fixed price. They price on a curve, meaning the more you take out in one go, the worse the rate you get, in the same way that dumping a large block of shares on a thin market moves the price against you while selling a handful barely registers. So valuing a share at the whole-pot rate is the harsh, pessimistic price.
The loophole falls out of that gap. Someone could deposit into a basket, immediately claim a small slice, and get out at the gentle small-sale rate while the accounting assumed the harsh one. Each round trip skims a little, and the little comes out of everyone who stayed. Nothing about it requires breaking any rule. It is just arithmetic being taken at its word by someone who read it more carefully than the people who wrote it.
If the fix ships, it helps exactly one group: people who put money in a basket and leave it there, which is the group Root Reborn exists to serve. It costs almost nobody anything, though anyone making rapid deposits and withdrawals will find the terms less generous than they briefly were. What could go wrong is the usual risk with tightening a payout path: the same code that stops a clever claim can also refuse an ordinary one if it is drawn too tightly, and this is the third correction to basket claims in a fortnight. Fixes written quickly around live money are exactly where the next problem tends to be introduced. That is not a prediction, it is a reason to read the next hotfix carefully rather than skim it.
The second proposal worth your attention is smaller but sharper. The chain supports standing orders, which are instructions parked on the network that wait for a price rather than trading immediately, the kind of thing behind a stop loss or a take profit. Those orders are signed to last a hundred and eighty days. A developer noticed that the supporting record other systems must read to act on such an order was being thrown away after only seven. So an order could still be alive and valid while the paperwork proving where it came from had already expired, leaving services unable to act on something the owner reasonably believed was still in place. The proposal raises the retention to match the order at a hundred and eighty days. It helps anyone relying on an automated instruction they set once and forgot, and the cost is simply that the chain keeps more records for much longer.
A handful of others are queued behind those. One would stop the chain wiping the forwarding note that points from an old account key to a new one when the account is also removed from a subnet shortly after, which today can leave an automated contract unable to find where funds went. One would make the developer toolkit's practice run, which shows what a transaction would do before you commit, refuse to report success when the wallet cannot actually afford the transaction, since the true cost is only settled by the network at the time. One would move WTAO, a contract that holds your TAO and hands back a standard token representing it so ordinary on-chain applications can handle it, into the main codebase so it is reviewed alongside everything else. And one would trim wallet balances on screen to three decimal places while keeping the exact figure behind the scenes. That last one changes nothing and improves the day of everyone who has squinted at nine digits.
Two subnets serve a brand new AI model within hours of its release
GLM-5.3, a new AI model, was published this week with its weights released openly, meaning the actual files that make the model work are free for anyone to download and run rather than locked behind one company's service. Two Bittensor subnets were serving it almost immediately.
Engy, subnet 53, says it is running the model at roughly a hundred tokens a second, which is about seventy-five words of output per second, faster than most people read. Its prices are the interesting part: ninety-eight cents per million words of input and three dollars and eight cents per million words of output, against a dollar forty and four dollars forty from the model's official service and from OpenRouter, the reseller most developers use as a default. That is roughly a third cheaper on both sides.
SayGm, subnet 28, is also running it, hosted through a partner called KubeTEE which runs the model inside hardware built so that the operator cannot read what users send it. If you have ever hesitated before pasting something private into a chatbot, that is the concern it is aimed at. This is the latest in a run of day-one launches from that subnet, and the claim being made is simply speed: an ordinary person can try a brand new model through Bittensor on the day it appears.
Why it matters to someone holding TAO is that both of these are the network doing the thing it was built to do, competing on price and speed with the companies that make the models, and being genuinely usable while doing it. The caution is equally simple, and we would rather say it than let the numbers stand alone: a subnet is paid in newly created TAO for participating. That means a price undercutting the official one is not automatically proof of a better business, because part of the bill may effectively be paid by emissions rather than by customers. The question worth watching is not whether the price is low today. It is whether these services still have paying users if the flow of new coin ever slows.
A playable 3D world from a single sentence
The most substantial thing said this week came from Ben of the 404 team, the 3D generation subnet, on a community call. The subnet builds AI that turns a written description into a three-dimensional object you can actually put in a game.
The interesting part is a change of direction. The team used to produce what are called Gaussian splats, which you can picture as a dense cloud of coloured dots that together look remarkably like a real object. They looked stunning, and studios could not use them, because game engines are built around meshes, a mesh being the wireframe skin most computer graphics have used for decades. The industry feedback, in Ben's telling, was that they wanted something that fitted the pipelines they already had rather than something impressive that did not. So the team rebuilt around meshes.
Their model, Titan, is built on top of an open source model called Trellis 2, and the subnet's competition rules required that whatever miners produced could actually be used commercially. That forced miners to strip out parts of the original that blocked commercial use and rebuild them, and Ben says quality improved in the process, with holes and geometry errors getting fixed along the way. They added the ability to feed in several reference views of the same thing, and support for normal maps, which are worth understanding because they are a lovely trick: a normal map paints the appearance of bumps and detail onto a surface without adding any actual geometry, so an object looks far richer while staying light enough for a phone to draw. Titan produces a model in around thirty seconds, against several minutes for the closed commercial tools he named.
On top of that sits Atlas, which does the unglamorous work of making an asset genuinely usable: matching a style, simplifying a heavy model while projecting the detailed surface back onto the simple version, and setting the pivot point so a sword ends up in the hand rather than through the wrist. The commercial proof he offered is a mobile game called Colony, made by Parallel and in early testing, where players generate their own weapons, helmets and armour while playing. The generation is hidden inside the game itself: you describe what you want to a character, and it is finished by the time you have walked across the map. Parallel report that this is the single most liked feature in testing, and Ben says he expects the game to launch within weeks. Two people also used the technology to win the first Chroma Awards, an AI gaming competition, generating over twelve thousand assets for their entry.
The subnet's current competition has moved on again, this time to generating Three.js code, meaning the object arrives as editable instructions rather than as a finished shape. Those look plainer, but they are structured, far lighter, and they understand that a door belongs to a house, which is what you need for a game that runs in a browser tab.
All of the above is Ben's account, not something we have independently checked. What makes it notable is that it is one of the few places in this ecosystem where you can point at a product a member of the public can play, rather than at a benchmark.
Turning power that would be thrown away into computers
Josh, from subnet 110, spent an interview this week explaining a genuinely unusual pitch: putting small data centres on British farms and running them on electricity that would otherwise be wasted.
The premise is that the grid is much worse at carrying power in the direction of the farm to the town than most people assume. A large field of solar panels may 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 electricity is simply not made, or not used. Josh's argument is that a computer sitting next to the turbine does not care about the substation, so the wasted generation can be pointed at AI work instead.
His numbers, and these are his figures rather than ones we can verify, are that a generator selling back to the grid earns roughly eight to twelve pence per kilowatt, which puts the payback period on a wind turbine or an anaerobic digester at twelve to fifteen years. Running compute instead, he says, has been generating sixty to eighty pence and in some cases a pound. His worked example is a farm that spent five million pounds on turbines and a further million on servers, where he claims the payback could fall to four or five years including the extra spend. The requirements are strict enough to be believable: real servers on verifiably green power, proper enterprise infrastructure including a ten gigabit connection that is equally fast in both directions, and human beings on site to fix things. Verification that the power really is green is currently a person driving to the site and looking, which is honest but obviously does not scale beyond a small region.
The distinction he draws is between rented and permanent compute. Most crypto compute is sublet from somewhere else and may not be there next month, which is fine for short bursts and useless to a company that needs a hundred machines reliably for six months. He argues nobody in Bittensor is offering the second thing, and that is the gap he is aiming at. If it works, it turns a subnet into something with a physical footprint and real contracts. If it does not, the failure mode is expensive hardware in a field, and that is worth remembering when reading the payback maths.
One date to note
Masx.ai, subnet 104, which is trying to turn the flow of information across Bittensor into forecasts and signals, has finally put a time on the beta of its version three: Monday, the thirty-first, at nine in the evening central European time. Until now there was a description and no date. A beta means an early public version expected to be rough. The date is the news here, not the product, and the thing worth judging is what it actually does when people can use it.