Keith Singery & Garrett Oetken: TAO.com Wallet, Bittensor, TAO Flow, Governance, Subnets | Ep. 77

Welcome to the Ventura Labs podcast. Today we are joined by Garrett Edkin and Keith Singery. We discussed the evolution of tensora and tao.com, the shifting roles of validators and subnet owners in product development and governance, changes in subnet designs and mining incentives, subnet collaborations for decentralized AI stacks, opinions on talflow verse pricebased admissions addressing exploits and the iterative process of improving bit tensor and the idea future governance emphasizing it, social elements and advice for identifying trustworthy actors in the ecosystem. Now, back to Garrett and Keith with the 77th Ventura Labs podcast.

Host: Okay, the Bit Sensor Guru and Garrett. I guess I should refer to you guys as the carrot and the stick now from the new podcast. How are you guys doing?

Keith Singery: We're doing well.

Garrett Oetken: Doing well. Yeah, we got the first episode of our new pod out and it's been a fun ride so far and it's just a cool experience to be able to share on that kind of a platform. It's a new venture for us, but it's fun.

Host: Yeah, you can't even pay for this kind of access to Garrett. If you've listened to my podcast, you know that I'm a little bit weaker on the technical side. I don't have any coding experience or anything. I ran the validator for two years, but relied very heavily on subnet owners and AI to basically get me there. And now that I've got this set time set aside every week to pick Garrett's brain on how everything actually works in Bit Tensor, I think the podcast will allow us to basically bring a much greater level of understanding and fill in all the gaps that I've had honestly on my side. It's easy for me to ask the right questions because I don't know. When I see these developments or announcements come across on the official Discord, I'm like, "All right, here's what I think that means. Here's what I think it means for all the alpha and the trading and how should I treat this as a change to the ecosystem." To be able to run that stuff by Garrett and get a better understanding of it, that's really invaluable for not only for me, but for everybody involved in Bit Tensor.

Host: I was saying before the pod started, I think this is the perfect combo like you guys alluded to on your podcast of Garrett and his immense technical capabilities and then Keith being one of the great orators of Bit Tensor being the first podcaster to ever do it and now you guys are chief evangelist of Tao.com and I believe CTO of tow.com. So what is yeah I'm actually head of protocol uh at telecom so it's a it's a slightly more refined uh focus in there but um but yeah it's it's a interesting position and and really it's um just around our involvement as far as to.com goes um in the ecosystem as a whole. That's kind of actually the angle that we're like going at it with with our podcast is is from that perspective of protocol involvement and ecosystem advancement. Um, and that's really what my kind of more refined focus is uh at Tensor largely.

Host: And what is this TensorOttow.com entity like? How is this makeup actually structured? Because my understanding is Giles and Jack, they made Tinsora back in the day. They're co-founders of that. And then Garrett and Yav made some other entity. And now these entities are one.

Garrett Oetken: That's right. Yeah. So yeah, Yov and I partnered up about this time last year. Technically a little bit before then, but when we were getting things going. We started having some conversations with Jack and Giles and we just thought it was like what we were doing was so compatible. We realized we were both attacking really the same problems in just different ways and we realized that it just made a lot of sense to merge our companies together. So we merged and created a new entity which is still called Tensor but it's a new entity now and so we're we're all moving forward as one and then we have tow.com as kind of our brand. So everything that we're doing productwise and whatnot is under the tow.com flag. But we still are that core tensor entity as far as the business is structured.

Host: Keith, how do you feel about this? Back in 2022, if I told you you'd be part of this mega conglomerate that is created in Bit Tensor, what would you say?

Keith Singery: You know, I got into Bit Tensor so I didn't have to be a part of conglomerates anymore or at least corporatized conglomerates. But after having spent a couple years in the project the need for professionalism and organization around the development of tools to support bit tensor was just totally obvious to me. So like and to give people a little background as to how you develop that like as as running a valid when you run a validator you're in the discords constantly you're con you're looking at the code bases whether you understand them or not you're looking at the readmes for everything that's going on within the subnetss you have to look at the minor readmmes the validator readmes you get to know everybody involved in the project from const on down to the smallest minor who wants to get started and it became obvious to me after, you know, really towards the end of 2024 that the most valuable thing for the validator that, you know, to basically leverage the the size and the capabilities of that validator was to put it to work in one of these larger organizations and and sort of use it as a not even just a store of value, but a creation of value, not just for itself, but for Bit Tensor as a whole.

So when I started looking for things to do with the validator towards the end of 2024 and I came across Tensor you know it was it was just one of these things and there was some synchronicities around it also but it was one of these things where I'm like okay somebody's got to take us out of the crypto Twitter uh you know like just just the echo chamber type thing. And now a year, you fast forward a year later and here we are with you know really what is a I would say um industry or spaceleading application in tow.com that has the ability to onboard you know thousands eventually tens of thousands who knows how far we can take this of new users into bit tensor. It's like you know makes things so much easier for the average person to get involved. And when I kind of made the bet on um you know transferring ownership of the validator this in my greatest of hopes was to be where we are today you know and we finally landed here. Uh and so for me it's the fruition of many years um of kind of dedication to the ecosystem and using the part that that kind of I've developed to help bolster this organization's position uh within the ecosystem.

So, like I love it like and it it wasn't easy and from the beginning uh I think it you know it's we had subnet 81 going of course and um you know that ended up shutting that down. So like it wasn't it wasn't just like smooth sailing up and to the right always you know looking perfect but um when you put you know the goals that we had as to.com around the group of 30 or so individuals that we have working for the organization and Tensor at this point um is really a foregone conclusion that we're eventually going to get there you know and now we got there and it feels awesome and I think it's really going to pay uh dividends and bear boot for everybody involved in Bit Tensor, you know, over the next few months and years.

Host: And this crypto bit tensor bubble that you we were stuck in a couple years ago, tao.com is for sure the current easiest interface for new people to use and get to subnets. How do you guys feel about the sentiment that used to exist around validators were the ones that should be making products on subnets and how it has kind of pivoted towards subnet owners being pressured to make the products? Do you think we'll see a pivot back towards the old sentiment of hey validators should use this bandwidth that they have to build the products.

Garrett Oetken: I think that the reason why validators were the like original product owners is because they were the ones that were operating the hardware that all of the output was being funneled to. And so it was kind of perceived that they were the ones with the opportunity to build products. But I think we pretty quickly realized that validators do not have the bandwidth to build products on every subnet and find those market fits in all those areas. Like that's a whole business. Like after you've already gotten the minor outputs from the subnet, now building a product off of that is an entire startup. And that is much more in line with the subnet owner's goal than it is with the validator's goal. Cuz a subnet owner will come in and they'll say, "Hey, I want to solve a problem." They'll start a subnet. The problem will get solved and it's like, "Well, then what do we do with that?" Then some validator is going to build a product off of it. Um like excuse me, hold on, wait a minute. I built a subnet so that I could solve this problem.

I think that's where we're just seeing that the incentive really is for owners to be building products because they're the ones with the vested interest in solving that problem in the first place. That's why they made the subnet. Whereas the validators have this role in the ecosystem where they kind of have to I hate to use the word shepherd but they kind of have to um kind of deal with and and work with um almost every sub owner and that makes it really difficult to to focus when it comes to product development. You can't build 128 products or more um as a single company. the validators would have to become such large entities. They'd have to raise so much capital and the companies would be just so you know diverse in focus and I don't think that would really work. So I think the focus that has been readjusted back to subnet owners is correct and I I don't think that's going to go back to the validators. I think the validator's role is being much more refined now in terms of operation of the subnet and governance rather than product ownership.

Keith Singery: Yeah. And actually I would add to that that the you know the old school way of designing a subnet was always to use the validator as essentially bandwidth for access to the intelligence. So if your validator has a certain amount of stake, the miners are incentivized to respond to you because if they don't, you know, they're going to take a hit on their emissions, right? So the way that always worked is there was sort of like a floor, call it like 20 30,000 where anyone there as far as like weight on the subnet where if any if a validator has less than that, the miners can just blacklist you and then they don't have to respond to you or whatever. So it created this like incentive to to build these larger validators. But if you look at the way that subnets have started to um you know sort of evolve in these different directions. For example, there's multiple subnets in the top 10 that don't actually use that as the methodology for creating access to their intelligence and maybe they don't have any actual miners in the traditional way that you would have seen it on the network a year ago. Right?

And now that's created um sort of a bifurcation in the way that people see bit tensor working optimally where you have these little fftoms in as each subnet where people can just kind of like owners can just say here's how we're running this. We're not having any actual miners. You don't have to register you know UIDs in the traditional way. We're going to control the alpha. We take it all and then distribute it at the end of the month. That requires of course quite a bit of u additional trust than the original design of bit tensor but that's their prerogative and if they can then associate some sort of value with that or gain that trust from the community then maybe they've earned the ability um to operate in that way but I still think and we talked about it actually on our on our first podcast that Garrett and I did with the the carrot and the stick that you know I think that we're going to see a transition towards um putting more not responsibility necessarily but just um like value on the validator that you're selecting.

The first thing would be that now the validators default select the the alpha you know um swapping when you're staked to root. So if you have your TOW stake to root on a certain validator, that validator will make the default decision to whether you're earning alpha on subnet one or whether you're earning TOAO through the traditional autoell mechanism that you have on route on subnet one. So that's the first thing that's a new development um you know just in in just recently with Bit Tensor that never existed before. So that creates a

Garrett Oetken: I will clarify that that's not on chain yet. So that's a that's a to be implemented thing. Um but uh that is the direction that we're going right.

Keith Singery: So then that you know leads us in the direction of creating additional decision-making parameters around which validator you're going to be staking to because there's going to be these sort of alliances where the validator will decide to essentially lower the amount of auto sell based upon taking the entire root stake on their validator and deciding to keep the alpha instead of just swapping it to tow. So for subnet owners that that creates the and and for stakers that creates the incentive to learn a little bit more about what your validator is doing, learn a little little bit more about the subnet and create some real relationships. I expect that will also evolve into a kind of return to the traditional way of operating on bit tensor where the validators do have access to the bandwidth and that the and that the products that are being built on these individual subnetss, the access to those products does potentially come through the validator. And I think that that's a healthy way for a new subnet to sort of design themselves around because there's kind of this question of like as a subnet owner, how do I create value for my alpha token? And this is one of always the first questions. It's like how do I ascribe value to my alpha? I've got all this stuff built. I've got this intelligence commodity, but does the does holding the alpha actually mean anything? Well, if you just stick to the principles that that were like in the original design, you know, that got us to this point, the value in alpha is inherent in that design. So, that's what I would suggest if you're designing a subnet on bit tensor is to take a really close look at, you know, sticking to that and and designing around it. And I think then the chain takes care of some of these questions around how how do I create value for my alpha token.

Host: And with the how to create value for the alpha token and as we talk about validators here what do you guys view as the role of validators in the future because if subnet owners bear the responsibility of building the product are they now not incentivized to try to garner all the alpha stake and say don't stake to anybody else just give us the stake and basically block off every other validator. What kind of bit tensor do you want to build? Do you want a subnet where you've got decentralization and you've got strength through that and you've got 15 20 validators with you know more or less equivalent stake or do you want to build you know one behem one single valid I mean it can run both ways. So as the as someone who's, you know, got some tow in your wallet and deciding which subnets you want to swap into, it's like what kind of subnets do I want to support? It's a question really for the community than it is and the community will decide. Um, and if we want to do the research and highlight these differences in the way these subnets are built and that that downstream leads to different decisions as to which subnets are supported uh by this by the swapping of tow. I think that will will eventually determine you know the value of the validator.

Garrett Oetken: Yeah, I I will say that um I think that validators are playing a much bigger role in governance now than they've been in the past. And a big kind of central central point that I think validators are participating in or or that they're um the role they're playing uh is centered around trust. A lot of people are like well these sub owners are doing some really cool things or at least they say they are but are they really? And the validators are the ones who are actually running the subnet owner's code to communicate with the miners and get ultimately the intelligence from that subnet. And so they can really opine on if that's legit or not in a way that a typical maybe casual investor wouldn't really be able to um to know. And so a validator can say hey you know we we back this subnet or we are um uh you know we we are uh committed to this subnet in running the hardware. Um and you verify that through having a number of other validators. Um and that's kind of where the original Yuma white paper with Yuma consensus uh kind of drove towards is that hey we are doing the best we can to verify intelligence by having everyone else run the same code and by doing that and if they all agree with each other then they must be on the right track. And then as a outsider you can look at that subnet you can say okay well you have all these independent parties that are not related to the subnet owner that are all operating this code and that gives me confidence that oh this subnet might actually be doing what it says it's doing.

Now there's a little bit of a complexity when you have subnets that really don't need 15 validators to run on the subnet operationally. Um and so what do you do there? How do you how do you uh bring trust into that subnet where maybe you really only do need a single validator? Like to Keith's point, he's like you can build a bit tensor any which way you want. Um that's kind of the beauty of it, but it's kind of like the blank canvas problem where you can do anything, but um but also it it can be kind of paralyzing to u to not know exactly which direction to go. But with the single validator system um there needs to be a way to verify and I think that's where we're going to see auditing and things like that um come into play more often where you have validators that maybe aren't running the code of the subnet owners as much but they are verifying the validators who do run the code. Um, and that way we can still have trust in the subnet um, without requiring there to be 15 or 20 validators if a subnet doesn't need that truly.

Host: Keith said a interesting point earlier and that is some subnets don't have miners in the traditional way. Both of you guys have been in the ecosystem for a long time now. How have y'all seen the ethos of Bit Tensor shift from when you first joined to now?

Garrett Oetken: There's a lot there. I think mining has changed a lot partially because of onchain features like minor burn and things like that. But also people are just getting more creative in being more efficient with their alpha. So, kind of the the vanilla bit tensor is that you have weights that rank miners and the miners get emissions on a regular basis based on where they sit in that ranking. The inefficiency there is that what if you are in a certain rank but you're actually not contributing a lot. Should you still get that portion of emissions? And this is kind of where the logic behind minor burn comes in is that hey, if there's a subnet that is inactive or a subnet that is partially active but isn't really taking advantage of its full potential, should those rewards still be distributed to miners? And I think that's a big question that people go back and forth on because one, it makes it highly profitable for miners to put in less work and still get, you know, the percentage of the total alpha emissions for miners. Which is makes which makes bit tensor very attractive to work for. But also it can leave money on the table. It can it can um make capital inefficient and subnet owners are starting to think of alpha as their capital even though it's alpha that's being emitted by the chain they they never uh have possession of it but it's their subnet and so they're starting to think of this as oh this is capital that I can use to motivate people and so if I'm paying someone more than I would pay them otherwise now all of a sudden that capital is inefficiently used and so people are coming up with all these creative ways is to distribute that.

The basic way is to implement some degree of of minor burn which is basically means that a a portion or even an entirety of the emissions is just going to nobody. And so the total amount that can goes around is going to be less. But the other thing that we're starting to see is this idea of a capacitor or a treasury. It's been called both things in the community, but this idea of like diverting funds to a pool that builds up over time and then it can be released to either a group of miners or to even a single minor as in a winner takes all system to where instead of miners receiving rewards on a regular basis, it's a little bit more contractual where you achieve a thing and the capacitor or the treasury will pay you. And that's something that I'm really looking forward to partnering with people in the ecosystem on figuring out exactly what that should look like cuz right now people are doing things very creatively. And some of that is, you know, difficult to see. You know, it's not very transparent and some of it is easy to see. It's uh you know, it's it's as onchain as it can be. But that's where I think a lot of the debate comes in on that on that side of things. Is is a capacitor or a treasury model okay? Is that good for Bit Tensor? And should there be a native Bit Tensor way to do that as opposed to, you know, using smart contracts or even doing something completely off-chain that we've seen happening in the ecosystem.

Keith Singery: Yeah, I agree with all that. And I would also add that I've seen a huge change um in the focus of development of subnets mainly around the end goal. If you go back to prior to dynamic tow and for some of us I think probably the last like 10 months feel like 10 years uh in bit tensor so it's hard to remember but before dynamic tow you know we had the validators determining the emissions for all of the subnets for a year and a half right so it was October of 2023 all the way up until uh you know we launched dynamic towel um now back then and even now, but back then way more. Every subnet was a research subnet because everyone was this whole thing about research subnets. Everyone is and everyone still is a research subnet. Some of them are researching something a little bit different. I'll go into that in a second. But back then, nobody talked about revenue. There was no I swear I never even heard it. Like I might not even saw I I bet you if you go back to the discord and you look for like how many times revenue came up before dynamic tow I don't even think it ever came up the like the thing we were just looking for is users exposure to bit tensor if you go back to and you know the big breakthrough was corel right because you know you could do some of the fun stuff that you could maybe do on some of these centralized models but it was powered by a subnet or two and it was decentralized and you know Mog machine did a great job of sort of showcasing what is possible um on the network with that. But it was none of the subnetss were designed to to like and there was no feedback mechanism because there was no alpha. Everyone was earning TOAO, you know, it was this sort of a it was totally different than it is now.

Now fast forward to dynamic tile launching and now the focus for everybody and maybe rightly so um to a certain extent is how am I going to create revenue so that I can buy back my alpha token or create value for my holders or create this you know internal flywheel mechanism to take this chart up and to the up and to the right while we've got this sort of like dilution mechanism of continuing to emit all of these tokens all the time, right? Um, and that's fine, but I I would tell you that everybody is still a research subnet. Shootes is still figuring out how to optimize its incentive mechanism and its offerings uh to create the best possible decentralized inference, you know, platform uh serverless compute platform on the planet. And it and it's going to continue to evolve. Same thing with Liam on the GPU side. Same thing with uh Hippius on the storage side. Same thing with Targon on the compute side. So there's there's these subnets that have chosen to like focus in uh very very um you know closely on the uh generating value from a commoditization of of their quote intelligence commodity. However, those subnets that I just named aren't really generating an intelligence commodity, right? because the thing that they're offering you is is almost like these, you know, foundational um uh these foundational pieces of of the AI stack and or really not even necessarily AI but just of using the internet, using a computer, creating value from from tools that use this technology and these subnets are going that route.

Then there's a whole, you know, there's almost everybody else that isn't creating that isn't got one of, you know, their teeth into one of these like basic foundational um stores of value or or value creation mechanisms in the form of inference, compute and storage. And those are much more like on the on the hardcore quote research track. But I think what you're going to see then is these re quote research subnetss start to break through and create their own commodities that are true intelligence commodities that are a little bit higher up the stack maybe but that have much more value in them uh for the end user because they're not just a sort of and I don't want to simplify too much but they're not just kind of like using tow emissions to pay the miners and then it's a subsidi subsidization thing for these these commoditized um you know intelligence commodities. And what they're actually creating is something that leverages the ability for miners to contribute intellectually and create creatively to this intelligence commodity that unlocks a much greater level of of value creation than maybe just like you know the pick and shovel method of these top 10 subnets that have that have sort of found theirelves successful at this point. And there it's not really a surprise that those subnets are at the top of of the st of the heap at this point because those are the that's the easiest the quickest way the easiest way to generate value is to find one of these existing things that have demand that have uh you know maybe a little bit easier way to to implement and and like have an open router to plug into, you know, for shoots in Targon that allow them to to create some value for themselves and their holders. But I really think that as as Bit Tensor develops further and you're and and you'll see some of these subnetss that are on maybe quote a research track now, they'll start to unlock their intelligence commodity as this research continues to bear fruit. And you'll see a totally different level of ability of value creation out of those subnets. Way different than we saw before Dynamic Tow and way way different than we've seen in the first 10 months of Dynamic Tow up to this point.

Host: What is that ideal ratio for a subnet for miners to be hardwarebased verse skill-based? What is the ideal balance there?

Keith Singery: It's the that's how like who do you want to attract? What kind of investor do you want to attract? Like are you are you going to just swing for the like are you looking for the home run ball every time? Then you want to go almost like full research if that research is leading towards a commodity that could be uh monetized in a way that could unlock a huge amount of value. This is ridges, right? This is ridges. This is this is where you have a very low cost on the minor side and you take the the um you know the abilities for miners probably what and what they're developing every day in their day job or maybe some other research or other subnets whatever they've done to create themselves uh to the point where they can contribute to this you know 10,000 lines of code that allow um that that agent layer to do the software engineering. That's that's like who knows how much value could come out of that. Nobody nobody really knows. Like it could be monstrous. So that's one way. Uh and it and it really is like a risk. It's an appetite of riskreward that I think as a subnet owner you decide what you want to take on, you know. So that's that's how I think that that that's play played out.

Host: Garrett, what would you say is that ideal balance between being 100% skill-based for miners or 100% hardwarebased?

Garrett Oetken: I mean, it it does depend, but my like excitement comes through a emphasis on skill. The reason being is because if it's hardwarebased exclusively then it just becomes a numbers game of how much does it cost to produce hardware versus how much emissions am I getting as a minor and then there are some optimizations that do require scale like latency and things like that that need to be considered. But what I think is really exciting is when Bit Tensor is producing something intelligent that no one in the world has produced before. And I think when you produce a hardware service that is a precursor to that. And so when I look at shoots and I look at these other subnets that are providing kind of infrastructure I view that as valuable but I view it in a different lens. I don't view that as oh they're going to kill AWS. I view that as they are going to be a further layer of decentralization in that intelligence creation.

So, this is what I'm actually really excited to see is when you see subnets that are actually working together and cross like integrating where you have one subnet that is focused on training a model, but the other subnet might be focused on producing the hardware to train the model and you have another subnet might be produced like producing data or or or what have you. And you have this entire pipeline that's roped together to where no singular subnet is going to destroy the you know the big AI players in the space but collectively you are exponentially increasing your minor pool essentially because you're using the minor pool of other subnets. We're saying hey I got I got miners that are producing hardware. I got miners that are producing intelligence with skill. They are engineers or they're researchers and then I've got uh you know miners that are producing all kinds of other aspects to the ML pipeline. I think that's all valuable, but I think the ultimate excitement of Bit Tensor and what will come out of Bit Tensor is going to be intelligence that no one's ever seen before and that's going to be produced by skill or that's going to be emphasized by skill. Because right now like the the giant AI companies can always throw more money at data centers and produce hardware. But what is that hardware doing? And I think that's the big value of shoes. is a big value of gradients a big value of all these other um you know hardware based subnets Liam etc is what is that hardware being used for and in concert with what and and that's where I really see that going and so I view the infrastructure subnets as being um the foundational layer for the skill-based subnets to achieve terminal velocity and you're seeing this happen already you you just saw what was it a DB get get trained on Templar and then get fine-tuned on gradients and then was hosted on shoots. So like this is how the whole thing works, you know, like you stack them all together and I remember when we had like the first subnet in integration with the other one, it was like they're talking to each other, you know, but now we've got actual models being trained in a decentralized way. We've got them fine-tuned w with one of these mechanisms that's you know was was pioneered by ridges but then uh gradients uh you know adopted that allows for for this fine tuning and then we've got like like Garrett said the infrastructure layer that allows for the hosting and and this is the holy grail and now you know you'll see that continue to expand to where the entire stack of artificial intelligence is represented by subnets on bit tensor and where we don't have subnets those subnets will emerge church because that's what our ecosystem demands. We want like we don't want to be re reliant on any of these sort of outside uh you know sources for intelligence data training fine-tuning whatever uh and all of it will eventually be consumed by the mechanism of generating subnets and and creating value in that way and then linking them all together.

Garrett Oetken: Yeah. And I think that when we're seeing this orchestra of subnets like that are producing value and producing intelligence it is like it's creating this ecosystem that is not um it's it's uh it's not bound by like any one particular subnet that's responsible for making a product that's going to like break out and be the big winner. I think everybody's like, well, you know, how are you productizing your subnet? Well, if we have this orchestra of subnets, you play one instrument. Yeah, you could be a great solo artist, but what if you just played your part in this massive production? And then the product side can come via either at a different time or it could even come from a completely different team. So the it kind of relieves I think a lot of pressure when it comes to the when revenue question of like research subnets for example if you say hey I'm researching this thing I think it's going to change the world how exactly it's going to change the world you know we can collaborate with other people on um but we're going to produce a model or we're going to produce some sort of output that is going to be the very best in its space that's going to be truly innovative. you just laser focus on that, the revenue will come and the opportunities will come.

And so like OpenAI was not a product company, it was a AI research company and I followed OpenAI for a long time prior to Catbt's launch and it was very interesting. It was cool to see their progress. You had to try out the little betas and stuff and then they hit something that actually worked really, really well. They finally got chat GBT. Something that the average user could use and it would work remarkably well. It wouldn't make uh silly little mistakes. I mean, it made some, but relative to the previous models, it was it was pretty good. And once it got to that point, then all of a sudden they released products and apps and websites and things like that. But before that, OpenAI's chat GBT interface was pretty rough. It was it was not it was not super sexy because they weren't product people. And now we're starting to see that they've taken this massive lion share of of product market because of the what they've produced. And I think the bit tensor is no different. It's like focus on building the things that you're really good at building and collaborate with other people to create the final product. So that way it's not like okay, we have a model and now we've got to build a website, we got to build infrastructure, we got to build all these things. It's like there's a subnet for that, you know. And that's what I think is really key here is that as subnets start to collaborate more and more, the pressure of any one particular subnet gaining revenue decreases in my opinion.

Host: How do we incentivize or cultivate this environment to where it's almost encouraged for subnets to collaborate? Because recently we've been seeing heightened tensions within just bit tensor in general. There's there's a lot of self-fighting within this orchestra currently. How can we avoid that self-fighting internally?

Keith Singery: The self-fighting is more of just like uh you know the water hole is getting smaller because the price is going down. So everybody's like a little bit closer than there used to be.