Today brought one change that quietly alters what happens to a lot of people's money, a second one sitting at the final gate before it switches on, and a week of subnet announcements that, unusually, were mostly about paying customers rather than plumbing.
If you are new here, the short version of the whole thing. TAO is the token of a network called Bittensor, and Bittensor is best pictured as an umbrella company. Underneath the umbrella sit roughly 128 small independent AI startups, each one called a subnet, each one focused hard on a single job. The speakers on this week's TAO Pod suggested that if you find the word subnet unhelpful, you should mentally replace it with the words independent AI startup, and that is honestly the most useful single piece of advice for a newcomer we heard all week.
What went live
Before the changes themselves, it is worth knowing how a change actually reaches you, because the language around it is confusing and it explains why some of today's news is marked live and some is not.
A change to Bittensor starts as a written proposal, published in the open, which anyone can read and argue with. If it survives that, it is merged, meaning accepted into the official code. Merged code is then bundled into a numbered package called a runtime release, and this week's numbers are 446 and 447. Even then it is not running. The last step is that two independent sets of key holders have to sign off, rather like a bank vault that needs two managers to turn their keys at the same time. One half of that signature comes from the automated build system. The other has to come from a small group of humans. Until both turn, the code sits in a queue doing nothing at all. So when you see a release described as proposed and awaiting signatures, that does not mean somebody is still thinking about it. It means it is written, finished, packaged and waiting at the door.
The biggest of this week's changes is called Root Reborn, and it changes what the network does with newly created TAO.
You can stake TAO, meaning lock it to the network to help secure it and earn a return, in two different ways. You can back one particular subnet, which is a bet on that specific team, or you can back the network as a whole, which is called staking to root. Root has always been the cautious option, the index fund of Bittensor: no view on which startup wins, just exposure to everything under the umbrella. Until now, the rewards a root position earned arrived as a payout, money that landed in your hands. Under Root Reborn, those rewards are instead automatically reinvested on your behalf into a basket of subnet holdings.
The reasoning behind it was laid out on this week's TAO Pod by Mark Jeffrey, and it is worth hearing in his terms while remembering that this is one well informed person's reading rather than an official statement. He says the old arrangement had an unfortunate side effect: in order to produce those payouts, the chain itself was selling subnet tokens on the open market with every single block, which he put at roughly a thousand TAO worth of selling per block. His description was that it felt like trading in a market where the central bank is printing money specifically to bet against you. Under the new arrangement, the chain takes that newly printed money and buys into subnets instead of selling them.
The buying is weighted by market signals, so the subnets already performing best receive the most of it. Mark described this as the chain force multiplying what the market is already doing rather than a committee picking winners. He also noted, and this is the part worth carrying with you, that everybody outside that top group dislikes the change.
What this means if you hold TAO and stake it to root is a real change in the shape of what you own. Your return no longer arrives as something you can spend. It arrives as a growing spread of holdings across many subnets. Less like interest paid into a savings account, more like a fund that automatically reinvests its dividends whether you asked it to or not. The second effect, as the host of that episode put it, is that the daily selling pressure the protocol itself was creating simply goes away, which matters most to the sort of large, patient investor who buys TAO and stakes to root rather than trading it.
What it could lead to is the open question. A network that buys its own subnets is a friendlier place to hold something. It is also a network where a meaningful share of the buying is coming from the protocol rather than from customers, and the two are not the same thing. We come back to that at the end.
Release 446, now merged, does the least glamorous and possibly most necessary job on the list: it repairs the network's own arithmetic. Every subnet issues its own token, called alpha. Think of alpha as shares in that particular startup, separate from TAO, which is the share in the umbrella. The network keeps running totals of how much alpha exists for each subnet, how much of it the protocol itself holds, and how much has been permanently destroyed, which is called burning. Two of those historical totals had drifted away from reality, and alpha that was burned before the network started keeping proper track of burns was never counted at all. Release 446 repairs both errors and backfills the missing history.
A bookkeeping fix sounds like something to skip past, but it is load bearing. Other rules are written as a percentage of those totals. The rule about who can take control of a subnet, which is the big story in the next section, is measured as a share of the eligible alpha supply. If the supply figure is wrong, the threshold is wrong too, and a rule can then be passed or failed by an accounting error rather than by anybody's actual holdings. Getting the counters right is what makes the next rule mean anything.
Also inside release 446, and live now, is a small piece of control handed to each subnet individually. Bittensor pays validators partly on how closely their scoring of miners agrees with everyone else's. The school analogy used on the shows this week lands well here: the subnet is a classroom, the miners are the students who hand in the work, and the validators are the teachers who rank that work from best to worst, reward the top of the class, and remove the bottom one and replace them. A bond, in this context, is the record of a validator's backing for a particular miner, built up over time. The change, called Liquid Alpha 2.1, lets each subnet choose which round of scores that bond calculation compares itself against: the current round, the previous round, or an automatic setting that switches between the two. What it really does is let each subnet tune how quickly a validator is allowed to change its mind about a miner without being punished for it. That matters more in some kinds of work than others, because a subnet judging fast moving predictions has a very different problem from one judging slow, careful research.
The most human of this week's fixes concerns work that simply vanished. Some subnets ask participants to submit their work sealed, so that nobody can read it until a set time has passed. That stops one competitor copying another's answer before the deadline, the same reason exam papers stay in the envelope. The sealing is done with something called a timelock, an encryption that can only be opened once a certain moment arrives.
A bug meant that if the network failed to open one of those sealed envelopes on its first attempt, it did not simply try again later. It threw the envelope away. On subnet 120 that is exactly what happened: participants paid the fee to register, submitted valid work, and that work was destroyed before anybody could read it. They were charged for a race they were then not allowed to run. The fix keeps the sealed submission in place when an opening attempt fails, so a later successful attempt still works. If you have ever had a form time out after you filled it in, you know the feeling, except that here it cost real money.
The last live item is deep plumbing and can be short. Bittensor has a mechanism with the unlovely name GRANDPA whose job is finality: the moment at which a transaction stops being probably done and becomes permanently, unarguably done. Finality depends on knowing exactly which group of authorities is currently allowed to sign off on it, and there was a flaw in how the network moved from one group to the next. It has been corrected so that the group and its identifying number always change together, in the same instant. There is nothing here for a holder to do, but this is the kind of fix you want made quietly and early rather than loudly and late.
What is coming next
Nothing in this section is live. All of it can still be rewritten, amended or dropped. That is exactly why it is worth reading now, because by the time a change lands, the chance to form a view on it has already gone.
The most consequential proposal on the table changes who is allowed to take control of a subnet.
Every subnet has an owner, and ownership is not ceremonial. The owner takes a share of the rewards that flow to that subnet and sets its direction. Bittensor deliberately does not make that ownership permanent. A challenger can take a subnet from its current owner by demonstrating enough commitment, which the code measures with something it calls conviction. Conviction means locking up that subnet's own alpha tokens, not merely holding them. You have to tie your money to the mast, where you cannot quickly sell it, for it to count at all. The design principle is simple enough: talk is cheap, a locked position is not.
The old rule added together the conviction of every account key involved and compared that total against 10 percent of the eligible supply. The new rule, packaged into runtime 447, says instead that a single account key must on its own hold conviction above 18 percent of the eligible supply, where eligible supply means the subnet's alpha minus the portion the protocol itself holds and minus everything that has been burned. Two things change at once, then. The bar nearly doubles, and it can no longer be cleared by a crowd.
There is a good reason the old version was distrusted, and it comes out in a separate proposal that is fixing the wreckage. An earlier upgrade, version 443, could carry one individual lock forward while applying only that one member's change to the group total, and then mark the group total as up to date. The result was a running total that had quietly stopped matching the individual locks sitting underneath it. When the rule that decides who owns a business is measured against a number that has drifted away from the truth, the problem is not really about ownership at all. That same proposal would repair the corrupted state, make ownership changes update the underlying individual locks properly, and require a challenger to satisfy the threshold in its own right rather than inheriting somebody else's.
Who does the new rule help. Mostly a subnet team building something that takes years rather than weeks. Under the old rule, a group of holders who happened to disagree with the founders during a rough patch could pool their positions and take the business. Under the new one, somebody has to put nearly a fifth of the entire supply in one place, in one key, and lock it there. That is an expensive, visible and difficult thing to do. Founders can plan. Employees can be hired with a straight face. It is easier to build when the ground underneath you is not moving.
Who could it hurt. Everybody who thought the ability to replace an owner was a feature rather than a nuisance. That mechanism is the only real accountability a subnet's holders have. If a team goes quiet, stops shipping, or simply loses interest while continuing to collect the owner's share of the rewards, the takeover rule is the fire escape. Nearly doubling the bar, and requiring the whole of it to sit in one pair of hands, makes that fire escape considerably harder to reach. Small holders, who could previously only ever have acted together, lose the ability to act at all.
Three things could go wrong, and we would rather raise them as questions than pretend to know. First, does concentrating the requirement into a single key create a new fragility, given that a key can be lost or stolen, and that one key would now be enough to control an entire subnet. Second, does an 18 percent single key requirement simply hand the mechanism to whoever is already largest, so that takeovers do not stop happening but instead become the preserve of the wealthiest participants. Third, what happens the first time a genuinely abandoned subnet cannot be rescued because nobody can assemble a fifth of its supply in one place. None of those are predictions. They are the things we will be watching for.
To be completely clear about its status: the code is merged and packaged into runtime 447, and it is sitting in the signature queue described earlier. It has not been switched on, and between now and then it can still be amended or pulled.
The second proposal worth understanding properly is far less dramatic and touches many more people. It concerns limit orders.
An ordinary trade takes whatever price the market gives you at the moment it goes through. A limit order is an instruction to trade only at a price you name, or better, and if the price never gets there, nothing happens. It is the difference between telling a shopkeeper you will take the item whatever it costs, and telling them you will take it if it ever comes down to forty pounds.
The genuinely new part is linked orders. Today, if you want to move a holding from one subnet to another, you have to do it in two disconnected steps. Sell your alpha in the first subnet, wait to see how much TAO that actually produced, then place a second order to buy into the other one. The awkward part is the join in the middle. When you sign that second order, you have to name an amount, and you do not yet know what the first sale will yield. In a thin market the difference can be substantial. Guess low and you leave TAO sitting idle. Guess high and the order fails.
A linked order fixes the join. The second order is sized as a share of whatever the first one actually produced, so the whole thing can be expressed as a single intention: sell my holding in this subnet, and put the proceeds into that one. The mirror case works too. Buy into a subnet now, and set a sell on exactly the tokens that particular purchase produced, rather than on your holding as a whole.
For anybody who is not sitting in front of a screen all day, this is the difference between being able to express a plan and having to babysit it. For smaller holders it removes a quiet disadvantage, because somebody running software that watches the chain could already stitch these steps together and somebody doing it by hand could not.
What could go wrong is worth saying out loud. Making it easier for one person to rotate between subnets also makes it easier for everybody to rotate at the same time, in the same direction, into whatever performed well last week. Bittensor's subnet markets are not deep. Tools that make it frictionless to chase performance can amplify the very swings being chased. And any new order type is new code sitting directly next to people's money, which is a good reason for the careful review this one is currently getting. It is a proposal, not a feature.
A third proposal is smaller but has a sharp edge to it. There is a hard limit on how many subnets can exist at any one time. Today, when somebody registers a new subnet while the network is already at that limit, the cheapest existing subnet is dissolved instantly, inside the very same transaction. The owner gets no warning, no chance to respond and no say. One moment there is a business, the next moment there is not.
The proposal would turn that instant execution into an offer. The new registration would be parked at a stated price, and the existing owner would have roughly a day to match that price and keep their slot. It replaces an eviction with a right of first refusal, which is how most of the rest of the world handles the same situation, from tenancies to company takeovers. This matters even if you will never own a subnet, because it changes what it is like to build one. A team that could be erased without notice on any given afternoon is a team that will struggle to hire, to sign a customer, or to promise anybody anything. The counter argument is real too: the limit exists for a reason, and a grace period does slow down the replacement of weak subnets with new ones. There is a genuine trade off here between protecting what already exists and making room for what comes next.
Two more proposals deserve a sentence each, mostly for what they reveal. One would change how claims from that reinvested root basket are processed. At the moment, a single claim performs one separate transfer for every holding in the basket, which currently works out at around thirty eight transfers, and the proposal replaces all of that with one aggregate movement. The number is the interesting part, because it tells you the basket Root Reborn is quietly building on your behalf spans roughly thirty eight subnets. That is a genuinely broad spread rather than a token gesture.
The other adds proper support in the command line tool for signing with hardware wallets and vaults, and for approving multi signature transactions without keeping wallet files on the machine doing the signing. Given that the switch which turns on every single network upgrade is a multi signature approval, making that approval both safer and less fiddly matters more than it sounds. A related change would stop an unanswered approval for an older runtime from blocking a newer one in the release queue, which is precisely the sort of thing that leaves a finished upgrade sitting at the door.
Who is actually earning money
The most striking thing about this week's subnet news is how much of it was about revenue rather than releases. Treat all of it with the caution due to self reported numbers. These are teams describing their own businesses, not audited accounts.
Bitcast, subnet 93, pays social media creators to post about products and sells that audience reach to advertisers. It says it spent about thirty three thousand dollars last month buying its own token back off the open market, has spent about ten thousand dollars doing the same so far this month, and intends to reach roughly seventy five thousand dollars before August is out.
A buyback means a project takes money it has earned and uses it to purchase its own token from whoever is selling. It is the opposite of issuing more of it. Companies on ordinary stock markets do the same thing with their shares. The stated logic is that if a business is genuinely earning, spending some of those earnings on its own token converts outside revenue into demand for the thing its holders own. The caution is equally simple. A buyback tells you money is being spent. It does not, on its own, tell you where that money came from, and it is not the same thing as a profit. The number worth following over the next few months is not the size of the buyback, it is whether the revenue sitting behind it keeps growing.
Separately, and this one is a rumour rather than an announcement, two people who mine on Bitcast say the subnet has removed its biggest constraint. Until now, only the subnet's owner could bring in advertisers and set up campaigns by hand, which put a ceiling on how much work the miners could ever be paid for, no matter how good they got at it. The claim is that other teams will now be able to build their own creator, ambassador and affiliate products on top of that same machinery. If it is true, it turns Bitcast from a business into a platform, which is a much larger thing to be. We have not seen it confirmed by the team, and it is reported here as what two miners said rather than as fact.
Investing88, subnet 88, which runs investment strategies on the network, has opened a public fund it calls the 88 Public Index, with a deposit ceiling of ten thousand TAO. A fifth of everything deposited is automatically used to buy the subnet's own token on the open market. The team says it is one of the few subnets already earning real money. The deposit cap is the detail worth noticing, because a capped fund is usually the sign of a team that would rather not take in more money than its strategy can sensibly handle, and that is a more encouraging signal than an uncapped one.
NIOME, subnet 55, puts medical research problems in front of the network to be solved, and says it has taken on three commercial partners in under four months, each from a different corner of healthcare. The first named one is Floré, a company that builds personalised products around a person's gut bacteria. Three customers is not a large business. It is, however, the thing that matters most about any subnet, which is whether anybody outside this ecosystem is willing to pay for what it produces. A subnet earning from Bittensor is being funded. A subnet earning from a customer is being bought from. Those are very different things, and only one of them can grow indefinitely.
The week also brought what its builders describe as the first lending market on Bittensor. MentatLend has published its design and opened the market starting with a single subnet, Score, which is subnet 44 and specialises in computer vision, meaning software that interprets images and video. The idea is simpler than the vocabulary suggests. Somebody holding TAO can lend it out and earn more than they would by simply staking it. Somebody holding a subnet's own alpha token can borrow TAO against that holding rather than selling it. If you have ever taken a loan against a house instead of selling the house, this is the same shape of thing.
It is also, and this needs saying plainly, the moment borrowing arrives in a market that did not previously have it. Borrowing against a volatile asset means that if the asset falls far enough, the loan is closed out and the holding is sold at the worst possible moment and not by choice. That is how lending markets work everywhere, it is not a flaw peculiar to this one, and it is a genuinely useful facility for people who understand what they are taking on. It is simply not a free way to hold on to something.
New tools, out in the open
404-GEN, subnet 17, has launched a tool called Atlas for making 3D models, running on the same machines that compete on its subnet. One of the models inside it, Titan v1, turns a flat image into a 3D mesh, which is the wireframe skeleton a game character or a film asset gets built on top of. Turning a single photograph into a usable 3D object is a job that traditionally costs an artist hours. The team says Atlas is aimed at professional studios, with cleaner geometry and better textures than its earlier work managed. Whether studios actually adopt it is the only test that counts, and that will take months to know.
Chutes, subnet 64, which rents out AI models over the internet and is the largest subnet by most measures, has launched something called docuextract. You give it a PDF or a photograph of a document, an invoice, a contract, a bank statement, and it gives you back the contents as tidy text, or as structured data arranged into a shape you specify so that other software can use it directly. Anybody who has retyped numbers off a scanned invoice knows exactly what problem that solves. The notable part is where it runs: inside what is called a hardware attested secure enclave, a sealed area of the processor that can prove to you which code it is running and that keeps the contents out of sight even from whoever operates the machine. Given that the documents people most want processed automatically are precisely the ones they least want a stranger reading, that detail is the product rather than a footnote to it.
Actual Computer, subnet 95, which lets people run AI models on their own machine instead of sending their data to somebody else's, made a model called Qwen 3.8 available on the day it was published. Qwen is a family of freely available models from the Chinese company Alibaba, and the mid sized version has been well received by early users. Same day availability is a small thing that says something larger, which is that adding a new model there is quick plumbing rather than a project. The same team has also trailed a tie up in which installing Actual would set up a fresh, sealed off copy of another AI tool called Hermes, in a way that cannot see or disturb whatever software is already on the computer. That claim came from an outside developer and was shared by Actual's founder, and it is described as coming soon rather than done.
One subnet appeared nowhere in this week's posts and only in a long interview, which makes it the most interesting thing we watched. Its founder runs an existing business that optimises Google Ads campaigns, given as Ad Tower in the transcript and AdTAO in the video title, and he has brought it onto Bittensor as a subnet.
The problem he describes is specific. An advertiser at any given moment has a basket of possible changes they could make to a campaign and no reliable way to rank them. In his telling, advertisers pick on face value from very small data sets, with no real perspective across the wider ads ecosystem. What his product was missing at its centre was a prediction engine. So rather than hiring data scientists, which he says he cannot find, cannot verify and cannot afford, and where a single hire is one bet with no way to check it, he has put the problem to the network instead. Miners on his subnet predict the causal impact of a change to a campaign: if you make this one change today, what will this campaign's performance be one month from now. They have to get two things right, the direction of the effect and its size within a range, and a correct direction inside a narrower range scores more heavily, because that shows the miner genuinely understands the specific case rather than guessing broadly.
His figures are that the subnet has been live for five weeks, that miners are hitting over 80 percent accuracy on those predictions, and that roughly 104 miners competed for the prize in the previous week, a number he gave with a slight hedge himself. Those are his claims and we have not verified them. But the shape of it is the clearest illustration of the whole Bittensor pitch we have seen in a while: a real business with a specific, valuable, genuinely hard prediction problem, renting a global competition to solve it instead of a payroll.
Pooling what nobody has enough of
Four of this week's announcements are variations on one idea: a resource that is scarce and locked inside individual companies, gathered into a shared pool that pays the people who contribute to it.
Nodexo, subnet 106, assembles rentable computing power out of independently owned machines. It says it has passed one hundred verified graphics cards, across seven different card types, in sixteen countries, with uptime of about 99.8 percent over the previous day. Verified is the word doing the work in that sentence, because the subnet checks that the hardware really is what its owner claims. Renting computing power from strangers has one obvious failure mode, which is people saying they have a fast card and supplying a slow one, and checking is the only real answer to it.
OpenRoboto, subnet 80, has opened a shared robotics data pool, with Axis Robotics as its first named partner. The team's argument is worth repeating because it is not obvious. Robotics is no longer held back by secret software, since capable models are now published freely for anyone to use, but by data. Recordings of real robots doing real physical tasks are scarce, expensive to produce and locked inside individual companies. A shared pool pays contributors for that footage so that everybody can train on far more of it than any single company could ever gather alone.
dFusion, subnet 70, has launched Personas, which are AI stand ins built from a contributor's own online activity, so that the more somebody contributes the more their Persona resembles them. Because each Persona carries a reputation and earns rewards, the project has an obvious problem to solve: if one person can quietly operate hundreds of Personas, the reputation means nothing and the rewards get farmed. This is one of the oldest problems on the internet and it has a name, the sybil attack, which is simply one person wearing many masks. dFusion has handed the checking to another subnet, Yanez, which does what it calls proof of personhood, establishing that there is one real human behind an account without necessarily learning who that human is. The stated scale is 750,000 people to be checked. Identity systems are where good intentions meet hard practicalities, and the details of how this is done, and what gets kept afterwards, will matter a great deal more than the headline number.
IOTA, subnet 9, is training a large language model called Orion-16B on ordinary graphics cards volunteered by people all over the world, rather than inside one company's data centre. It has now fed the model more than 100 billion tokens, tokens being the chunks of text a model learns from, roughly a word or part of a word each, and the team says quality is still improving rather than levelling off. They believe it is the largest model ever pretrained this way. The reason this is hard was explained well on the pod this week. Training normally requires an enormous number of chips talking to each other constantly, and communication between chips is roughly a tenth the speed of staying inside one chip, which is why they are usually bolted together in a single building. Doing it over the ordinary internet should not really work. That it is still improving at 100 billion tokens is the genuinely interesting part.
What was said on the shows
Beyond Root Reborn, the rest of that TAO Pod episode circled a question worth thinking about even if you never touch a subnet. What happens to these small specialist AI projects if the big frontier models keep getting better.
Mark Jeffrey's answer had two parts. The first is that a general model is trained across a wide sweep of everything, while a specialist holds an entire vertical of data the generalist has never seen. His example was fire detection, and specifically the footage that teaches a vision model to tell an actual fire from a man holding a flashlight. The second part is less about models and more about business. A customer such as a city school system or a state fire agency does not want a model, it wants a system that is built, installed, maintained and legally underwritten by somebody who will pick up the phone when it breaks. A frontier lab will not take that call. His analogy was Linux and Red Hat: the operating system was free to anybody who wanted it, and companies still paid handsomely for a specialist to fit it to their needs. In that framing, the frontier model is Linux and the subnet is Red Hat.
The subnet he applied that to is Score, subnet 44, the computer vision specialist, which he says is being distributed through the professional services firm PricewaterhouseCoopers. Both speakers said they met the person running that at a conference in Paris and were told he has not had a single client say no yet. That is a secondhand anecdote and should be held as one. But distribution through a firm that already sits in the room with large public sector buyers is a fundamentally different route to market from posting on X, and it is the sort of thing that would show up in revenue long before it showed up in a headline.
The second theme was that frontier models are commoditising. Mark's claim is that a lab ships something impressive and within a week or so a freely available model matches it, so the advantage evaporates and the value moves to whoever can serve inference cheapest and fastest. Inference simply means running a model to answer a question, as opposed to training it in the first place. Every time you ask a chatbot something, that is one inference, and somebody somewhere is paying for the electricity.
The example given was a subnet transcribed as NG, live for about a day and a half at the time of recording, serving two well regarded open models at roughly half the price of the next cheapest provider. Their explanation of how was twofold: tuning the models to run on consumer grade 5090 graphics cards rather than expensive data centre hardware, and the fact that miners scattered around the world can sit right next to the cheapest electricity available, which a data centre in one fixed location cannot. The claim is that this makes it profitable from the first day rather than subsidised into existence.
The host put the obvious counter argument, and it is a good one. The price of inference could fall towards nothing, the way internet bandwidth did, and the large labs are already undercutting on enterprise deals and can afford to lose money for years while they do it. Mark's reply was that the labs are not profitable either, and have never before had to compete with a decentralised network that is. Reasonable people can land on either side of that, and this week produced no evidence that settles it.
One more number came out of that conversation, offered as an observation rather than a valuation. They noted that Chutes is the largest supplier of open source models to a service called OpenRouter, which routes requests to whichever model provider suits, and that OpenRouter has reportedly been valued at around ten billion dollars by Stripe, while Chutes sits at around one hundred million. Comparing a company to a subnet is not a like for like comparison, and they did not claim that it was.
The episode closed with an explicit chain of arithmetic about what a single breakout subnet would imply for the price of TAO, drawing on reports the speakers attributed to Greg Chavea at Yuma and to Sami Kassab. We will report that the conversation happened and who was cited, and we will not repeat the figures as anything more than that. Both speakers said themselves that it was speculation and not financial advice, and this newsletter does not forecast prices at all. What is worth taking from it is not the number but the claim underneath it, which is that the entire investment case currently rests on something that has not happened yet: one subnet becoming a business large enough that people outside crypto notice it. Everything in the section above about paying customers is a partial answer to that question. It is a long way from a complete one.