What actually went live
Only one upgrade reached the main network this week, and it is the least interesting of the three sitting in the queue. Runtime 445 is now running. The runtime is the rulebook every machine on Bittensor follows, and upgrading it is like handing every referee in a league the same new edition of the rules at the same moment, so nobody is still playing to an old page.
Nothing in 445 changes what anyone earns. It throws out transactions that were always going to fail before they take up space, deletes storage entries the chain no longer uses, clears out old miner commitments, and makes proxy accounts behave the same way across more actions. A proxy account is a wallet allowed to act on behalf of another, like a signed permission slip left with a neighbour. Useful, invisible, and not the reason you are reading this.
The two upgrades people are actually waiting on, 446 and 447, are written, packaged and tagged for the live network, and neither is switched on. Deploying to the main network takes two signatures. The automated build system has given one. The second belongs to a small signing group at the Opentensor Foundation, the organisation that maintains Bittensor. Until that signature exists, the code sits exactly where it is. It is the difference between a law that has passed the vote and a law that has been signed: agreed, printed, not yet in force.
One of our own notes this week described a piece of 446 as already live. The release records say it is still a signature short, and we are going with the records. We would rather correct ourselves in public than leave you working from the wrong version of events.
Hold onto that gap between merged and live, because everything in the next two sections is real code that real developers have argued over and accepted, and none of it is running yet. It can still change.
What is coming next: root staking stops paying out
Bittensor prints new TAO every day and shares it among the projects running on it. There are two ways to take part in that. You can stake your TAO behind one particular subnet, meaning one of the independent projects that make up the network, or you can stake to the network as a whole, an option called root. Staking means locking your TAO behind a participant so their voice counts for more, and taking a share of the rewards that follow. Root has always been the option for people who did not want to pick winners: back the whole thing, take the yield, get on with your life.
A change called Root Reborn, now merged into the network's code, changes what happens to those rewards. Instead of arriving as a payout you can spend, root dividends are automatically reinvested on your behalf into a basket of subnet holdings. Every subnet issues its own token, called alpha, and its price moves with how that subnet is doing. A basket is a spread of those tokens, drawn from the list of picks held by the validator you staked with. At current sizes a basket holds roughly 38 of them, a number that comes from a separate proposal describing how many internal transfers a single reward claim triggers today.
The closest everyday comparison is a dividend reinvestment plan on a share. Your dividend never lands in your bank account, it quietly buys more of the thing that paid it. The difference here is that it does not buy more of what you already had. It buys a spread of other things, and somebody else drew up the list.
Why this matters if you hold TAO and have never touched a subnet: root was the low-decision choice, and after this, choosing not to decide is itself a decision. Your position slowly changes shape without you doing anything. It stops being a pile of TAO earning TAO and becomes a pile of TAO plus a growing spread of subnet tokens, each of which can move much harder in both directions than TAO does. If you had come to think of root as the plain savings option, the plain savings option now has a portfolio attached.
The argument for it is not hard to see. Under the old arrangement, rewards paid out in TAO could simply leave, and root looked like a way to earn from the network without ever funding anything inside it. Reinvestment turns every payout into a purchase, and those purchases land on the projects doing the work. It also means an ordinary holder ends up with exposure to the subnets that grow, without having to research well over a hundred of them.
What could this lead to? One possibility is that root quietly becomes the network's index fund, the sensible default for people who do not want to pick, with the validator acting as the fund manager. That would put real weight on questions nobody has had to ask before. Who chooses what goes in the basket, how visible are those choices, how often do they change, and can you get out. A validator picked because it never goes offline is not necessarily a validator whose investment judgment you would want. Another possibility is more mundane: it works, most people barely notice, and the main effect is a steadier flow of money into subnet tokens.
There is one early warning sign worth reporting honestly. One path inside Root Reborn, the ability to deposit straight into a basket, has been temporarily switched off by the developers while they measure how much data each deposit leaves permanently on the chain. A single direct deposit can spread across a validator's entire list of picks and open a new record for every one of them, and a blockchain has to carry every record it creates forever. A separate proposal attacks the same problem from the other end: today a single root claim moves money around 38 separate times inside the chain, which is like writing yourself 38 cheques to settle one payment, and the proposal replaces that with one entry. Both tell you the same thing. The mechanism is heavy, and the people building it know it.
None of this is final. It is merged rather than deployed, one part of it is switched off on purpose, and the efficiency fix behind it is still only proposed.
What is coming next: who is allowed to take a subnet away from its owner
Every subnet has an owner, and the owner matters more than the word suggests. They steer what the project does and they take a share of what it earns. So there has to be a rule for how ownership can change hands against the current owner's wishes, and that rule has now been rewritten twice in a matter of weeks.
The test is built on something the code calls conviction. Conviction is not a show of hands, it is locked money: alpha you commit and cannot immediately take back. The idea is sound. If you want to take control of a project, you should have to put your own capital at risk and leave it there, rather than borrowing support for an afternoon.
Until now, the test added up the conviction of everyone backing a challenge and fired once the total passed 10 percent of the subnet's own token. The version merged into upgrade 447 replaces that with a much narrower gate: a single account key must hold conviction above 18 percent, measured against the subnet's token that is genuinely out in the world, with the protocol's own holdings and everything that has been destroyed taken out of the count first. A key here is the thing you sign with, closer to a named door key than to the vault behind it.
So the change does two things at once. It nearly doubles the threshold, and more importantly it stops a group from pooling to reach it. A thousand small holders who between them own a third of a subnet no longer clear the bar. One holder with 18 percent does.
There is a reason for that, and it is a real one. Adding up locks invites gaming: conviction can be split, layered, borrowed or rented, and a total assembled out of many keys is much harder to verify than a single stake sitting in one place. The developers found that out the hard way. A separate proposal, still only proposed, exists specifically to repair state that an earlier upgrade corrupted by mishandling those aggregate totals, and it would additionally require any challenger to clear a 10 percent threshold on their own before being counted. The two pieces of work overlap, they do not yet agree on a single number, and the final shape of this rule is not settled.
Why it matters to someone holding a subnet's token rather than running one: ownership is the only lever the outside world has over a project that has stopped delivering. Make takeover too easy and anyone with a chequebook can seize something other people built. Make it too hard and a subnet whose owner has lost interest, or is taking a cut while shipping nothing, becomes untouchable, and the people holding its token have no recourse at all. It is the same question a shareholder asks about how many shares it takes to force out a board, with the same uncomfortable answer that both extremes are bad.
What could this lead to? The honest reading is that the network has chosen to protect incumbent owners from coalitions, and has accepted concentration as the price. After this change, the only realistic challenger to a subnet owner is somebody already very large. That could mean fewer opportunistic raids on projects with genuine momentum. It could equally mean that when a subnet does need new management, the only candidates are whales, and ordinary holders watch from the sidelines regardless of how many of them agree.
Again: not live. It is merged into 447, 447 is waiting on a signature, and a competing proposal in the same area is still open.
Also coming: miner pay, standing orders, and a day of grace
One more merged change touches money directly. The network works out each subnet's share of the daily TAO issue largely from that subnet's price and demand relative to the others. Subnet owners are allowed to hold back part of the pay meant for their miners, the people running the machines that do the actual work, either recycling that value or destroying it. A change merged on 12 August restores an older adjustment: the portion an owner withholds now scales down that subnet's share of the network's payouts too. In plain terms, a subnet can no longer quietly keep money set aside for its workers without the network noticing. If it matters to you which subnets treat their miners well, this is the change that puts a price on the answer.
Then there is a proposal that would change how people move between subnets. Today, swapping one subnet's token for another means doing it by hand at whatever price is on offer in that moment. The proposal introduces limit orders, which are standing instructions to trade only at a price you set in advance, the way you might tell a shop to buy something only if it drops below a certain figure. On top of that it adds linked orders, where a second order is sized as a share of whatever the first one actually produced rather than an amount you had to guess when you signed. The driving example is a rotation: sell my subnet 7 token, then put exactly the TAO that produced into subnet 12. Anyone who has tried that by hand knows the awkward part is that you do not know the middle number until the first trade has happened.
It would make moving around easier and less dependent on watching a screen. It is worth saying that easier movement cuts both ways, because standing instructions also execute during a panic, and automation tends to make crowds move faster than they otherwise would. We are describing the mechanism, not suggesting anyone use it.
A smaller proposal has an outsized effect on the people it touches. There is a hard cap on how many subnets can exist at once, and today registering a new one when the network is full instantly dissolves the cheapest existing subnet inside the very same transaction. No warning, no notice, no chance to respond. The proposal would park the new registration as a priced offer for roughly a day, giving the sitting owner the option to match it and keep their slot. It is the difference between arriving to find your market stall gone and being told what someone has offered for your pitch and having until tomorrow to match it.
The rest of what is in flight is plumbing, and plumbing is what decides whether an upgrade lands quietly or takes the network down with it. Multisig approvals could be given from a hardware wallet or browser extension instead of requiring wallet files on disk. Developers get a single instruction that returns the most recent finalised block, meaning the latest batch of records everyone agrees is permanent. A test job would rehearse each upgrade against every live subnet and fail it if a block takes four seconds to build or the chain stalls for twelve. And a stale approval for an old runtime would stop blocking a newer one, which is precisely the sort of queue that leaves a fix stuck behind something nobody is waiting for any more.
Money arriving from outside
The most useful question to ask about any subnet is whether anyone outside the crypto world is paying it. Three answered that this week, in different ways, and the differences matter.
Bitcast, subnet 93, pays online creators for posts that earn genuine, verified attention on YouTube and X rather than for follower counts, and sells that reach to brands. It says it spent about 33,000 dollars last month buying its own token back off the open market, has spent about 10,000 dollars so far this month, and intends to reach roughly 75,000 dollars before August is out. A buyback means a project takes money it has earned and uses it to purchase its own token, taking some supply out of circulation. What makes it interesting is not the price effect, it is what it implies: money is coming in from customers, and some of it is being pointed back at the token. These are the team's own figures and we have not audited them, and 75,000 dollars is a modest sum in absolute terms. Bitcast also released version 2.0.0 of its connecting software, aimed at outside developers building their own products on its plumbing, and two of its miners say the campaign machinery is being opened up so teams other than the owner can bring in advertisers. That last part is a claim from miners rather than an announcement, so treat it as unconfirmed, but if it is true it removes the ceiling on how much work miners can ever be paid for.
NIOME, subnet 55, puts medical research problems in front of the network to solve, and says it has signed three commercial partners from three different corners of healthcare in under four months. The first named one is Floré, which builds personalised products around a person's gut bacteria.
Investing88, subnet 88, has opened a public fund with a deposit ceiling of ten thousand TAO and routes a fifth of what comes in into buying its own token. Keep this one mentally separate from Bitcast. Buying your token with profit you earned is a different act from buying it with money depositors just handed you. The first is revenue finding its way home. The second is inflow, and inflow can stop.
New tools, and one partnership both sides confirmed
404-GEN, subnet 17, has launched Atlas, which turns written instructions into three dimensional assets for games and film. On a subnet, independent operators called miners compete to do one job well and are paid in that subnet's token for doing it better than the rest, so what Atlas really is, is a front door to that competition for people who build props, characters and scenery. The Opentensor Foundation says 404-GEN already works with several large studios including Square Enix, the Japanese publisher.
OpenRoboto, subnet 80, has opened a shared robotics data pool with Axis Robotics as its first named partner, and both companies announced it from their own accounts, which is worth considerably more than one side saying so. The team's argument is that robotics is no longer held back by secret software, because capable models are published freely, but by data: recordings of real robots doing real tasks are scarce, expensive, and locked inside individual firms. A shared pool pays contributors for adding to it.
Chutes, subnet 64, which rents out AI models over the internet, has launched a tool that reads documents into clean data. Hand it a PDF or a photograph of an invoice, a contract or a bank statement and it hands back tidy text or structured data in a shape you specify. It says the work runs inside a hardware attested secure enclave, meaning a sealed area of the chip that can prove what it ran, which is the kind of claim that matters when the document is somebody's bank statement.
Actual Computer, subnet 95, lets people run AI models on their own machine rather than sending their work to somebody's servers, and had the newly released Qwen 3.8 available on the day it came out. Qwen is a family of freely downloadable models from the Chinese company Alibaba. Day one availability is a small thing that tells you something larger: adding a new model is quick rather than a project. The team has also trailed a one step install of another tool, Hermes, in a sealed copy that cannot disturb existing software, but that came from an outside developer and is described as coming soon, so it is a trail rather than a shipment.
Nodexo, subnet 106, which assembles rentable computing power out of independently owned machines, says it has passed one hundred verified graphics cards across seven card types and sixteen countries, with uptime around 99.8 percent over the previous day. Verified means the subnet has checked the hardware really is what its owner claims, which is the recurring problem in renting computers from strangers. One day of uptime is one day, not a record.
The big experiments, treated as experiments
IOTA, subnet 9, is training a large language model called Orion-16B on ordinary graphics cards volunteered by people all over the world instead of inside one company's data centre, and says it has now fed the model more than one hundred billion tokens, the word fragments a model learns from, with quality still improving rather than levelling off. The team believes it is the largest model ever pretrained this way. If that holds up it is genuinely significant, because the assumption that serious AI training requires a warehouse of matched machines is the assumption behind most of the industry's cost structure. The quality claim is theirs and we have not seen a benchmark.
dFusion, subnet 70, has launched Personas, AI stand-ins built from a contributor's own online activity, where the more you contribute the more the stand-in resembles you. Because each Persona carries reputation and earns rewards, the project needs confidence that one person cannot quietly run hundreds, so identity checking is handled by another subnet, Yanez, through what it calls proof of humanhood. A figure of 750,000 people being checked is circulating in the community, and we are reporting it as a claim from a post rather than as a confirmed number.
Numinous, subnet 6, has opened a public scoreboard running the leading commercial AI models against models built by its own miners under identical conditions, the same news archive and the same memory of what each has already seen. Whatever the result, a public head to head that anyone can check is rarer than it should be in this industry.
MentatLend has opened what it describes as the first lending market on Bittensor, starting with the subnet Score. The idea is simple once unpicked: someone holding TAO can lend it and earn more than plain staking pays, and someone holding a subnet's token can borrow TAO against that holding instead of selling it. The mechanics carry an obvious edge, which is that borrowing against something volatile means a fall in its price can force the position to be closed at the worst moment. That is how lending markets work everywhere, and it is worth understanding before rather than after.
Masx.ai, subnet 104, previewed a third version of its forecasting work built around separate specialist agents for compute, trading, media and security. It is a preview of a plan, not a launch.
The bug that made valid work disappear
Some subnets ask miners to submit work sealed, encrypted so nobody can copy it until a set time passes. Subnet 120, Affine, depends on it. A bug meant that if the unsealing step failed for any reason, the chain threw the sealed submission away entirely rather than keeping it. Miners who had paid to register and had published perfectly good work appeared to validators as having submitted nothing at all. The fix keeps the sealed commitment when decryption fails, and a companion change now rejects malformed sealed messages at the door rather than storing them.
It is a small entry in a release note and it was somebody's wages. Worth remembering when reading any list of merged fixes: on a network that pays people automatically, a silent failure is not an inconvenience, it is unpaid work.
Why one founder said he came here instead of hiring
The clearest explanation of what Bittensor is for this week came from a founder on a podcast, not from a release note. He runs a Google Ads optimisation business whose whole purpose is to make an existing advertising budget produce more sales, and he says it had a hole at the centre: at any moment there is a basket of possible changes to an ad account, and choosing between them needs a prediction engine that his product did not have. Advertisers, he says, currently pick on face value from tiny data sets with no real view across the wider ecosystem.
His reason for building it on Bittensor rather than hiring is worth quoting in substance. He cannot find, afford, or verify the right data scientists, does not know which country or company they are in, and one hire is a single bet whose quality he cannot check. On a subnet, miners compete to predict what a campaign's performance will be one month after a single change is made, getting both the direction of the effect and its size within a range, with narrower correct ranges scoring higher. He says the subnet has been live five weeks, that around 104 miners competed last week, and that models are hitting over 80 percent accuracy. Those are his figures, stated on a podcast, and we have not seen the underlying data.
The programme's own framing is the most useful beginner explanation we have heard: think of Bittensor as an umbrella company holding well over a hundred independent AI startups, each obsessed with one narrow job, rather than one company trying to build a single model that does everything. Inside each one, the validators are teachers, the miners are students, and every round the best is rewarded and the worst is replaced.