Bittensor Roadmap :: SN7 GitTensor Software Engineering + SN7 Peer to peer BTC transfers

Añadido: 2026-05-23

[00:00:07]Okay. Well, there's a lot going on in Bit Tensor in general, but I think that what people um probably want to hear us talk about on the stage today is a bit of a road map for F. um the project. Some of the things that I spoke about last week, we um mentioned the potential for a proposal to remove root root yield. A lot of people were upset about that.

[00:00:41]They had some concerns, a conversation.

[00:00:44]Our our our intention with bringing that up was actually to stimulate a conversation, an economic conversation in the Bitensor universe about what the best thing to do. And one of the people that reached out to me was actually Sammy um from from Unsupervised.

[00:01:04]Uh and he he said to me, you know, like I think that I think that uh what people are concerned about when it comes to rem removing root yield um what they're concerned about is is that there's no other option. And so a lot of the the voices in against the idea of removing root root yield were concerned with okay where are we going to what what are going to what are root stakers going to do. So I want to talk a little bit about that today.

[00:01:32]Um I want to talk about conviction because that's going out in in in about 4 days. We're going to be launching conviction. There's been some small changes. I want to talk about governance in vitensor because there's a lot a lot of moving po part parts here. We you know some people have said that things have like centralized around const and you know part of the reason why things have centralized around myself is because we're actually in a process of decentralization where uh F is being dismantled and moved into a set of other organizations actually we're going through the process of decentralization and I just happen to be still here making sure that the the ship goes down properly. So, it's an it's a it's actually like a unique moment in the the story of Bit Tensor.

[00:02:23]Um, for me to be up on stage here kind of uniquely alone. Uh, Alice no longer here. He's with Crucible. Um, Medulla who who run have run the infrastructure for the chain are in the process of of moving outside of of F.

[00:02:39]Um, we no longer have uh Macrocosmos which was our Synapse team back in the day. um some of our nucleus team members are being moved to different companies.

[00:02:52]This is all in preparation for F to eventually die. Um which is a strange goal for a company. And I will say actually um that when you go to when we went to the Canadian government and basically explained uh that we were trying to shut down uh all of our lawyers said that that was that was basically illegal. Um uh because there are basically as a as a CEO of a company uh you're not supposed to want to destroy the company that you're inside.

[00:03:21]Um but we have had this intention from the get-go to construct this entity for the purpose of eventually disappearing and giving more of the governance and um sense making and decision-m of the protocol back to the community in an intelligent fashion. Now, that's very difficult to do. Um, and uh, I've been trying my best um, over the last couple of months to make sure that this ship is aligned for that that change.

[00:03:55]And one of those changes is conviction.

[00:03:59]One of the most important things um that we can get out while we still have F is the ownership of subnets and also the decentralization of the chain and also the governance of the chain.

[00:04:15]All three of those things are coming this year in Bit Tensor. A huge a huge endeavor to release full control of this project into the mechanism itself. Um and uh we conviction is a is a massive part of that and that's coming in three days. So that's one additional pillar of of control uh that we're building into the system which is protocol based something that is malleable and liquid owner liquid ownership at the subnet level. That's in 3 days coming this summer. This summer we want to have the governance system in play and I have actually um some good news that we do have um the initial implementation and the MVP of the governance system in Bit Tensor done. Um if people are interested they can follow this link uh which was developed by Loris at F to explain um core mechanistic structure of the eventual chain governance of of the Bitensor blockchain.

[00:05:29]I've spoken about it before and I will speak about it in the future and I don't want to spend too much time on this call talking about techn technicalities but for those that are interested you can visit this link and and read through some of those slides to see about how we're thinking about how to properly build governance.

[00:05:44]Well, it's basically a a trackbased system that's been designed to move fast um um but also slow when it needs to move slow.

[00:05:55]things like can we change the nature of root yield.

[00:06:03]It's not clear if this is something that there will be unanimous community consensus about and if there isn't this is why something like governance will be very helpful.

[00:06:20]It's I actually believe that we'll it will allow us to move faster.

[00:06:25]um and that governance will help us reach consensus around issues like this.

[00:06:29]But um one thing that I will say on that topic about the concept of root yield is we want to replace it with something else. And um what we're talking about internally right now is actually bringing the concept of um root funds as an alternative alternative where it's possible for t for for root validators to actually participate as um in to to try to create a risk-free investment option or a a safe and easy investment option into the subnets so that root stakers can um have access to the the tow ecosystem with without having to take as much risk as you let's say picking and choosing in the southern ecosystem which is which is very difficult and we've known this from the very beginning when we when we started off with DTA we needed to give people that initial risk-free um option for for staking Tao we want to improve that we want to advance it and evolve staking so that it is something that is not just let's call it passive passive um behavior inside the ecosystem but there is some more intelligent selectivity on behalf of of tow holders about where they put their stake who are they looking to invest um with in the ecosystem and our solution for that is to actually turn validators on root network into funds so more on that to come but I wanted to speak to that because there was some concern that we were just going to release this um willy-nilly with without people's consent, let's say, and without getting, you know, feedback and and and really approval from the community. So, there's there's more information coming about like how we're going to we're going to roll this out in a way that doesn't that doesn't just rip people's uh business models out from underneath their feet.

[00:08:26]conviction. There's one update to the proposal or to to the to the upgrade on Tuesday is that and very specific is that on Tuesday it will not be possible for any ownership to change that that's going to be a further um uh execu executed transaction that will from pseudo that will turn on the ability for subnets to be um controlled via conviction. So so on Tuesday with the upgrade that will not be in play. um and we will let people know when that can happen. However, by immediately on Tuesday, it will be possible for people to to build conviction in subnets and because the secondary trigger for a subnet is being older than one year and there being 10% conviction in the subnet which means that the lock has matured.

[00:09:21]The hyperparameters for that the specifics are in are in the proposal if if people are interested. You can see Mache Koula has done um a nice job of building a I'm going to actually bring this here and give it to the community um a conviction design tutorial. Learn Bit Tensor has done a fantastic job of explaining um conviction via their their documentation. And I I recommend all seven owners go go and take a look at this has the final hyperparameters of the upgrade. So to repeat that again um on on Tuesday we'll be upgrading the chain. Uh there it will not be possible for your subnet to be taken over until there is the executable which we're going to we're going to tell you when that's going to happen um in that week or the following week. We'll give you some fair warning. But you can actually acrue conviction during that time even though conviction is not let's say fully on uh inside the update. There's also some other small changes that are important to to submit owners that I think people should pay attention to. One is that registrations into a subnet are going to be counting as inflow. So currently they are buy and burns. They will now count as inflow as buy and burns. Uh which means that if you're a subnet that's actually in the um has a mining process as in there's there's people registering into your subnet continuous mining um you will act you will see that as as something that enhances your um emission in bit tensor. So it'll it'll help teams that are are actually um building a mining network. It's very positive feature. I hope people um are are excited for that.

[00:11:02]coming in to the week after that after next after next week. Uh we also have a change that is going to allow subet owners to to have much longer longer tempos. It's something that we've not allowed for since the beginning of of subnets the ability to change your tempo. Uh it'll now be possible. You can change your tempo up to a week. Um um so so if you have a subnet where you want to slash miners where they have to build trust in the weights dimension for some time before um they're considered to before you pay them out. A lot of teams uh need this functionality u we will have that by um next the week after next. So these types of of building primitives we will continue to to build into Bit Tensor to make it much more easy for teams. And and something else that that we want to do over the next coming months, governance or not, um is is build in the primitive slashing. Uh um so not just longer tempos, buildups um and and the ability to to to earn trust over a longer period of time, but also slashing uh that will be coming out um in June. Uh yes, that's right. Um the ability to turn off emissions in subnets. We're building we're bringing this primitive. so that subnet owners can actually turn off um emission in a subnet, their own subnet.

[00:12:26]Um and there's a possibility that we will be in conjunction with the governance process proposing uh votes for turning off emissions in subnets.

[00:12:37]Once we have the governmental process in play, it will be possible for actual proposers proposers of um legislation, let's say, into the ecosystem to turn off emission for subnets that are that are clearly abusing um the system. Now that we have the the actual technology to to do this, I missed a very important one. Um we still plan to push forward with shorting. Um, it currently has the the the terminology we're calling it shorting, but it's also longing. It's the ability to to borrow from the pool because both shorting and longing are actually asymmetric operations. We have uh thanks to Yuma, we've we've built uh Rapido does not like that, but I think that many people in in this group will um understand the the need for this operation.

[00:13:34]uh we will be rolling that out uh this year in in in in very slowly and carefully. Now some people are concerned that we don't know what the the consequences of this are. I think that the we know how to bring this into the ecosystem in such a way that we can watch how it behaves um and then increase the the amount of leverage that's possible on the pools. Um and this will allow like the core primitive is just the ability to allow people to take the opposite side of a trade. Um to make the market in DTO more efficient and more effective and uh some people may disagree with that and I think we will see that it does work. Um if it doesn't look we're going to smooth we're going to we're going to draw this in in in smoothly and and carefully to make sure that um it it has just a positive um effect on the ecosystem and DTO in general. There's also uh the update to TFlow uh Tal Taflow V2 that uh has come from um uh Tal 5. The TO 5 team um helped us upgrade TFlow to V2 which includes the change where we subtract uh we intelligently subtract the subsidy into subnets from the emission.

[00:14:52]We're continuing to explore and optimize the way in which we we measure TOW flow so it's harder to to exploit so that teams like 104 for instance they can't continue to to apply their their sort of like solistic uh strategy um of of purchasing their own subnets and and getting small amount of emission over time. So all of these all of these changes we think are going to improve the the internal market of of of of Bit Tensor and and help um us you know sense make about where a mission should be going.

[00:15:27]Okay.

[00:15:30]Plus 1% it's allistic. Yeah. It's true, right? Um yes. So okay, there's a lot there and when when I was before this call, I was thinking about road maps um for for Bit Tensor and and how we don't we don't actually have the clearest road map. We know that we're going in the direction of decentralization. We know that we're going to ch we're decentralized the chain. We know that we're going to build in the governance. We know that our intention is to improve um the market dynamics of DTA. We know that our our intention is to is to shore up any exploits in the ecosystem and push teams towards creating value. We know that the economic system in Bit Tensor is is aligned with actually um eventually being deflationary and and being capable of actually having very very sound economics um um systemwide. A little asterisk here for those that are interested about um burning tow yield.

[00:16:36]very cool quality of this design is that as the sum of subnets increases um it's very easy to understand how the economic model of the bitensor universe worksensor is an optimization machine um dao is effectively a subnet applied to itself at a meta level that optimizes for teams that are capable of sequestering towo but not only sequestering towo but also paying paying in terms of root yield more and more to dividends of root yield root root sakers. If we burn the that root yield and we incentivize teams to participate more in detail, we also will increase some of subnets. We will also eventually move bitensor to being net deflationary at a sum of subnets of five. How is net net deflationary? A fully deflationary protocol. And currently um when we measure the amount that's burned if instead of root yield, if we if we measure just purely the amount of TA that's burned via uh root yield, we have a revenue that is that is more than both Salena and Ethereum.

[00:17:53]Um protocols that have uh market caps many times larger than ours. And that's because we we've actually built products here and paired them with towel that have effective value, have real value, that have real revenue. And so I'm very excited for how these things will come in um over the next year and and really shore up the economics and push us forward as a community.

[00:18:18]But let's stop there. Part of this call is to talk about git tensor. So, Git Tensor is uh one of the most unique subnets on Bit Tensor that has focused their mechanism on incentivizing people to to contribute software updates. They've as a consequence of having a permissionless system that that works in this way have attracted a large set of really really cracked developers that they've pointed at um the web in general. Um, so I'm going to bring up uh the the founder of Get Tensor on the call and have a call with him. So yeah, hey Kimbo.

[00:19:02]>> Hey, what's up guys?

[00:19:04]>> I'm happy to have you guys up and you know talk about Git Tensor like uh as you guys know I'm I'm a happy investor.

[00:19:12]Um, uh, when you guys reached out the other day, you wanted to come up on stage, I was I was I was thinking like we could you could bring you up last week, but then I I wanted you to have a full a full like uh stage to yourselves.

[00:19:26]Um, and let's get into it.

[00:19:29]>> Yeah, thank you. Uh, so I mean, I guess just to start, hey, Kimble Logic in the group chat, my actual name is Alexander.

[00:19:36]I just go by Ander. And this is Mango aka Grant and Lando aka Landon, their actual names. And we're actually the team at Ventura Labs. This is us, us three. And I'll let them say a few words.

[00:19:52]>> Uh yeah, we're Grant. I'm sure a good amount of people have seen my face by now. And yeah, we uh we operate Git Tensor and always.

[00:20:07]Yeah, that's really it. I don't have much to add besides I'm just developer.

[00:20:10]You probably see me in the Discords a lot. So, just building with these guys.

[00:20:19]>> Hell yeah. So, >> should we talk about I mean I guess I mean should we start with the the original subnet or the new one? Um I I don't know much about always that that's interesting to me. But before we go there, like I I when you came to me with the idea of Git, you didn't come to me with the idea, you you told me about the idea that you've already created with Git Tensor. Um and you know, it seemed kind of crazy because you were going to be validating all of these PRs from from miners on Bitensor, which is going to be problematic, right? Because they're going to cheat you a million different ways. And um but then you you had some some unbelievable developers that actually join joined your your ecosystem and um I forget some of the names but like who who are they? Who who are like the the knights of uh of git tenser >> the knights of >> I don't know.

[00:21:16]I mean maybe there's some in the audience if they want to come talk they can come talk but >> yeah I mean when when it started we were just contributing to essentially a random list of open source uh repositories and we were forcing them to like kind of market git tensor for us but saying this contribution was from git tensor and I guess just naturally you know if you're contributing to open source you're probably a good developer and then if you start seeing that on some of the repos you contribute too.

[00:21:47]You'll look into it.

[00:21:50]>> What happened? What happened April 6th that just accelerated the number of merge PRs and issues resolved from the team? Is that is that just everyone started vibe coding?

[00:22:05]>> Uh April 6th was actually well so this graph we took out a bunch of uh repos on the list. So the list used to be like 4,000 repos. Now it's down to 10 and that's because you know the value isn't be able to direct the minor talent. So that graph looks crazy like that because uh on April 6th we basically just increased our maintainer speed on a couple of the repositories we um own. And yeah, so miners uh they just go crazy when you let them.

[00:22:45]Do you think that the future of software development is just, you know, like super minors? Um, the the AI agents that are just working across all of GitHub, you know, solving every PR and issue that they can find continuously, non-stop, like it's not even individuals. It's just, you know, a swarm of of agents, arboses, say, whatever, what have you.

[00:23:12]Uh my answer will probably vary from Landon and Anders, but agents have like a very different like economic rationale than humans. So it's like for an agent to do something, they just have to be able to make more than it cost in their tokens to do the thing. So I would say yeah, I think the future is pretty much all agentic software development.

[00:23:35]>> And h how does Git tensor play into that? like how do you guys see this intersection of incentivized software development and agentic software development?

[00:23:49]>> Uh I I do want to step back for a second and say I my answer is a little different from Grant on that. Is it just going to be like mega agents all over GitHub? My answer is like yes and no. I think they all still need to be basically have some human in the loop or some kind of like human driver because I mean as we've noticed with like a bunch of the repositories that we've taken out and added throughout the history of Git Tensor so far each repository has its own story. It's it has its own set of maintainers and an agent for one repo actually might not work for an agent for another repo. they all need to be like tuned a bit to fit the contribution quality, the the style and just the story that the repo has has held uh for various fixes and changes in the past and it what it will do in the future.

[00:24:42]So I just wanted to back up on that question a little bit. We sometimes have different views within the team on things but we generally find >> so like when you guys look at the the contributions that miners on get tensor are making as I understand it like you you pin some repos um but you have a lot of repos there and the these are anonymous it's it's a permissionless system they may not have reputation at all. How the how the hell do you get any PR sub like submitted? Like surely these these these repos are just getting inundated with with you know AI generated PRs all the time like you know where's the steady state of of it's very easy to write a PR and there's a million repos and there's a million agents writing them. Um yeah like what does what does it look like? What does this future look like with with Get Tensor and and and Agent Glow?

[00:25:45]>> Yeah, I'd probably let Landon or or Kimbo really take this one, but I think this starts with like the initial version of Git Tensor is significantly different than the current version of Git Tensor. So it started as just like proof of the incentive mechanism like it can work and then we were doing thousands of PRs and just overwhelming maintainers and then we we created you know a bunch of things for the um incentive mechanism to help sharpen the the PRs that actually come in make them all just higher quality.

[00:26:23]And uh I would let Landon actually probably explain how he viewed like the start to to now.

[00:26:30]>> Yeah. Yeah. Yeah. Tell us the story like what what what happened.

[00:26:34]>> Yeah. I think you touched on it a bit with the agents um and how they it's a permissionless environment and then they have essentially infinite tokens. So it pretty much is infinite spam. And of course, we just went to the incentive mechanism like the drawing board and went and tuned things like you can only have so many open PRs or this credibility like your ratio of merged PRs versus closed will be applied to your um emission reward. So then it really matters and things like this are what we implemented to tune that. Now to touch on the other side of it where it's kind of like this infinite token spam that is available and honestly it can be quite powerful if you harness that. I view that as best captured not through necessarily like typical open-source contributions but more like um objective benchmarks or something like this where you can throw as many tokens as you want but I can objectively decide if that improved something. So through a benchmark, that's where I think that's best harnessed.

[00:27:38]>> And and so this is the direction you're going. You're you're thinking about like it's going to be eval based agentic submission evaluation.

[00:27:49]>> To be honest, we're going the direction that supports both. Um and when I say that, that means that we'd have like configurable hyperparameters for each repository. So specific code bases like subtensor might be a little more of like the social aspect where you want to actually understand what you're changing and not just token max and spam it where those guys won't like that versus um other repos might be different needs like let's just say benchmark maxing and um we're we're trying to set it up to where all repos of all types can come and be empowered by get sensor.

[00:28:24]>> Okay. So I get it. So, so you know, you guys are thinking of repos being like auto researches in themselves and and and where there's an online continuous non-stop collaborative um competition to make PRs to those repos that actually improve specific evals. Is that correct?

[00:28:50]that is a a potential like way to do it because that's just how you maintain your repo is you accept a PR that improves on that benchmark. So yeah, absolutely you could do the the auto research as a repo and right now we actually incentivize both sides. So like the code generation and the code evaluation of uh does this PR improve this code base? Yes or no? because we added a maintainer cut for each repo.

[00:29:21]>> How do you evaluate that? How how do you how do you evaluate whether or not PR improved a repo?

[00:29:29]>> That's that's the main up to the maintainer. So in the objective case, it's did it improve on this benchmark?

[00:29:36]>> And so the maintainer basically defines that beforehand. Is that correct?

[00:29:43]They say like this is how a PR would be evaluated.

[00:29:49]>> Yeah, they could. I mean, we haven't seen it yet, but yeah, they absolutely could.

[00:29:54]>> To add some context on all this, so recently we added per repository called hyperparameters that allows you to adjust how you want PRs to be scored. And so the name of the game is emerge PR is is prize. And you can adjust per repository hyperparameters to say maybe just the most recent PR gets all the emissions of this repository.

[00:30:20]Or maybe the last five PRs that were merged get the emissions or you know you just have like a credibility system and the like the people with at least three merge PRs and two out of three merge.

[00:30:33]You can get kind of nuts with it.

[00:30:35]And what we're trying to say is that you some repositories can adjust their hyperparameters such that the most recent PR was merged. And if you have an agent maintainer kind of guarding the merge, what gets merged and what gets closed, you can effectively have a benchmarking repo that just rewards the most recent PR. Or you can just have like a normal repo where you just kind of put a bunch of issues as a maintainer and the maintainer is the one kind of going through and reviewing PRs and just kind of doing their thing and more than one person gets rewarded.

[00:31:07]So we're trying to allow that to happen like all like a whole world of repos that reward how they want to to get what they want out of you know their repository.

[00:31:18]And so so you you guys are building the tools to to to like what is that like a GitHub CI style thing? Like is it a um something they add to their website? Is it something that's actually added into the codebase specifically? Is it on GitHub? Like where where is that defined and like how is that written?

[00:31:35]>> Yeah. So we're like in early stages of that pretty much where we just added all the hyperparameters and add encoded it into the uh the repository itself, the subnet itself. And right now it's it's just a JSON file and all the whitelist repositories are JSON. Not the prettiest of course. We do want to in the future and I I do ask for some advice here from the team from you or someone else out there with the ideas as well. My thought was it's a smart contract repository lists. Someone goes on, registers their repository and they commit x amount of stake and out of all the repositories registered on that smart contract, your normalized committed stake to your repository is the emissions your repository will get. And that's also where hyperparameters would live.

[00:32:29]Grant takes the approach.

[00:32:30]>> Say that again. Say that again. your normalized committed stake is the amount of emission.

[00:32:36]>> Let me let me uh let me let me like rephrase that so it's a little more clear. Um right now we can like basically build anything autonomously and it's about deciding like what to build and so he's talking about giving the admission share the ability to say which repos deserve more admissions on the subnet over to the alpha holders.

[00:32:58]That's what he's talking about.

[00:33:00]>> Okay. So they can direct, they can say, "Hey, like we want to push.

[00:33:04]>> We're going to try."

[00:33:05]>> Yeah.

[00:33:05]>> The alpha becomes the steering wheel of what gets built. That's what he's expressing. But just like super technical and implementation, >> where have you guys been um like uh when you look at the the PRs that miners have made on Git Tensor, you know, what are what are some examples? I guess you just put the link in there right now, Lando. um it you know what are some what are some like headline PRs that you've seen like like sophisticated um improvements on code bases that would justify like somebody wanting to go and connect into the system and say like get get g get tensor to to make PRs on their on their code bases >> probably Kimbo could touch on on some of the ones he's seen to get tensor itself because he would just write issues and miners would build the subnet essentially Yeah. So, one is is get tender itself, but that's kind of a copout answer. But I do like sometimes I'll just seed an issue. I don't want to build it. I'll get like 10 PRs, have my agent review them all, choose the best one.

[00:34:10]You know, I just spent >> a specific example that's coming to mind was uh one of the old subnets that got dregistered now, but they made an issue using a get sensor issue bounty to like move their their storage over to Hippius and a minor completed that.

[00:34:29]>> I will add too that I saw some qu there was one minor that was had a lot of PRs on openclaw. So, he was going crazy on openclaw.

[00:34:37]I guess I'm also we also uh get tensor miners won the synth hackathon that we got first and second place without uh writing a single line of code. It was all done by miners.

[00:34:53]>> Wow. No way. So and and and you're okay.

[00:34:56]And also like the openclaw repository interestingly right he it's um is it Steve? Is his name Steve? um he designed it so that they take they take PRs all over the pl from all over the place, but they he has just agents that spin out and check those PRs. Um and he really believes in like agentic workflows.

[00:35:22]um you do you guys see a future where where every repo out there or like maybe 80% of repositories out there on the web are just managed by um agents that that you you know maybe one of the reasons why you can't actually submit PRs to a lot of repos is because they're humans and they just don't give a [ __ ] about all of your contributions. But the the speed at which things move will be moving in the future. Are lots of repos going to have this quality where they're they're looking for um submissions from agents from across the the web and they're willing and able to to you know rank and filter them um without the sort of like annoyance that you would get if you just got flooded by a thousand PRs.

[00:36:05]like are we moving into a world where where there's like the the inter agent contribution across the the like the web of of gits is much faster paced and does git tensor play into that future?

[00:36:21]I would say we started out with maintainers being like for sure the the limiting factor in essentially how fast a repo can go because miners would send 200 and we faced this oursel we got to like 200 PRs and like 200 issues on our repo and as a maintainer you have to handle that agentically like you're just unless it's going to be your full-time job and that's all you do you have to use some sort of agents to at least filter And uh Landon and Kimbo both have like made agents for that. So they could probably talk more on that aspect of it.

[00:37:00]>> Yeah. I'll just say it's not crazy. I think there is a future that trends towards that. I don't think it'll ever be 100%. But I think it will probably be 100% at a minimum agent aided.

[00:37:14]So in the chat um Leo posted here is an open here is an openclaw top minor PR PRs plural. So in in the chat these are all linked to PRs that have been merged into openclaw via a minor on git tensor.

[00:37:34]Is that correct?

[00:37:35]>> Yes.

[00:37:36]>> Yeah. our our top miner we saw had like I don't know 50 PRs to open claw in like couple weeks.

[00:37:44]>> Wow, that's incredible guys. This is this is like >> this is like 30 p. Like the obviously um openclaw is is is a very fast-paced you know it has like who knows how many PRs. Um I it doesn't say on the on on on on the page. I can't see it. Um, it has currently 3,000 open at at this particular moment in time.

[00:38:11]Um, but it's still really impressive that that you know you guys were able to get that and that's just from one minor on the on the network. Um, I mean like that would seem to play into like that seems would seem to be evidence that that this is working quite well and and like you guys have a machine that you can basically point at a variety of different projects and get real improvements. Um, you know what's interesting to me, I think, is that like you have the speed at which the miners can work because they're obviously running this with agents. Like I I'm I'm assuming, right?

[00:38:45]There's no way that they're doing this individually, are they? Right. So, they just have really sophisticated agents that are capable of of going searching the web and finding PRs. Is that correct? like >> I would I would add there it's I would probably say most aren't fully agentic but they'd be like a human agent hybrid right now.

[00:39:08]>> Okay. And that's because and that is is that because you you guys also measure the quality of submissions like if you're well they have to have 80% like merged. So that means if they put up 10 PRs eight have to actually get accepted.

[00:39:23]>> That's a very high bar.

[00:39:26]Yeah, >> to be eligible for the rewards for that repository. And I will add it, they might not be fully agentic because sometimes I'll I'll leave a review on an agent uh PR, which dings them penalty-wise if you leave a change requested, funny enough. But I'll leave review and if they don't respond, you know, quick, um, I'll just close the PR, dings their credibility. So, they have to have some kind of agent to help them. Hey, I just got a review requested or hey, I've got conflicts I got to fix. Um, but that depending on my review that I give them, they also need perhaps some human in the loop to tell them, okay, what fix is he asking for and how am I going to do it optimally so this gets merged.

[00:40:06]>> Wow.

[00:40:08]>> And I I would just add to that that point that you were saying like to actually get PRs into open claw is like impressive. That was that's we removed that from the repo list because that was the stage of like proving that the system can actually build things of value and it's like open call is the most popular repo right now. If you can write code good enough for open call you can essentially write code good enough for anything and then now it's just about what do we actually direct the system to build is how I view the stage we're currently at.

[00:40:42]>> You're okay that makes sense. So you're you're at you where you guys are sitting it's like okay we can direct this super cannon wherever we want it to to be directed. How are we going to direct it in a way that I guess draws revenue into the subnet like that that justifies the revenue? Um, could you direct I mean like there are there are subnets on Bit Tensor for instance that uh are agentic subnets like Ninja. Could you direct it at Ninja and then sell the mining emission and buy back into your own subnet?

[00:41:24]>> Hey, go ahead. You could you could um I personally like the idea of actually like making products for other subnets, but I don't know how much other subnet owners would like that because that would require us to have like a large validator on their subnet. So that and then they they talk about the ideas of like the optimization sub uh repos to where we just build these objective benchmarking things like you see subnets doing. So that's we're at that like inflection point of what to actually build. Now >> there's an interesting intersection with what you guys are doing and let's say a sub like 66 Ninja right their system is PR based right so they have an agent you can you can submit PRs as a minor to the agent as a hotkey on the subnet that um and and then it goes through the CI of that codebase. So it's it's a coding agent and there's a CI and the CI says this PR will only be submitted if you know Claude Code thinks it's good and with this with this review and that review and it adds these many lines and you can't copy somebody else and it can't be gibberish and it has to look really good. It has to be a legitimate PR to this coding agent and then on top of that we're going to run this evaluation to determine whether or not it's actually a good PR and I I reckon that there'll be a bunch of subnets on bits into that follow this structure because it it it works really well. Um, basically if you're building a coding agent, you should do it this way. If you're building any type of software, if you're building any type of incentivized software, you want to build it that way. And and this is like um the symbiosis of what you guys are building and those subnets is is obviously quite interesting, right? you you can direct the the engineers to to mine into the subnetss that have their mining emission as um uh as as GitHub repos that you can you want to get a PR into? And do do you think that like just outside of Bitensor itself like I mean you can mine in Bitensor but like is is the future of software um repos like OpenClaw repos like Ninja that are agentic run um very clear instructions really clear evaluations uh for for like how you can make a PR into the into the system with bounties.

[00:43:53]>> Um short answer yeah. Um, but with the with the point to the the ninja subnet, like kind of how we're thinking about how we're building our system is it should become unnecessary to have a whole subnet to do any type of software optimization. We're trying to build it to where you can do any software optimization on git tensor instead.

[00:44:18]>> Wow. Okay. So you guys are actually inverting the mechanism. So you you are going to make it that you're going to abstract the concept of something like Ninja and build any type of repo and feed the miners into that. Is that correct?

[00:44:34]>> Well, an optimization subnet is just a like super sophisticated way of deciding should this code be merged? Yes or no?

[00:44:43]>> Right.

[00:44:44]>> Exactly.

[00:44:46]And that's what Leo says, "One repo of git tensor is like a bit tensor subnet."

[00:44:52]>> Yeah, that's that's actually a lot closer to the vision of where we're going in my opinion. I it's it's very similar to the design of bit sensor pre-dynamic towel because we plan on having it be alpha like stake weighted through validators basically the same as pre-detail.

[00:45:15]Wow. Okay. So, so you on git tensor you guys are going to build a platform where there's going to be a set of different optimization problems that you are you going to define them or like people can come in and define their own but like initially there's going to be your your structures and you're going to run the evals and they're all going to have the same structure as a GitHub with a specific type of bounty and the the miners on the sub on the subnet are going to commit into those specific repos.

[00:45:42]See that that's the exact point we're currently at is deciding like what direction of how to do this. So Kimble actually came up with like an initial template for uh optimization that we may or may not do and it's uh he he can go into more detail about it but yeah >> I want to give credit to Landon on it because he he was like he was the first one that thought of it and then we've just been plowing through updates and get tensor to enable it and this kind of goes back to the repository hyperparameters that you can set and the merge PR is the golden than thing you want. Um, and yeah, like like we're talking about, you can effectively adjust your repositories hyperparameters to only reward the person that last merged PR. Why would that matter? Well, for the optimization repos that we're talking about, you build the benchmark, you build the agent that gates the PRs, and then you just set your repository hyperparameters to say, "Hey, just reward the last person that actually improved this benchmark or this optimization."

[00:46:46]So yes, >> do you do that and do you build that in directly into the CI of the GitHub? So it's just like this is like a GitHub action that uses your compute backend or and or like compute from from Bit Tensor to to run like because like some of the some of these eval processes may be like really complex, right? You know, Ninja is actually to to run out like a synthetic evaluation.

[00:47:12]uh you know you need to run with an open router key, you gota maybe you need to use have use a GPU. Is it are your your is your plan to to build git tensor as like a custom git github action that you can add to your >> well is >> because evaluate right >> here's where it got a little tricky is I was building it out and scaffolding everything I was like where where does that compute come from and I was playing a lot with shoots but I didn't have a lot of time to play with it. I've maybe someone in the chat can help me um figure out shoots a little better as a user because I was trying to deploy a shoot and then it wasn't warming up and then we're not integrated subnet and this and that. A bunch of the shoots lingo was was tripping me up a little.

[00:47:58]But in theory, a minor would, this is my first like V1 idea. And Landon, again, I want to give him credits on it, too. He thought of it, is they put it up on shoots. They grab at a station and metadata from the shoot. They put in their PR description, hey, this is what I ran it on. This is how it worked. And if the agent deems it well valid, it would then check itself using shoots or similar. Um I I don't know if that's maybe you would know better as well or again someone in chat but that was the first kind of route we'd go along as to the compute for evaluating the optimizations.

[00:48:33]>> Yeah there's there's there's a million different ways like um you you need to come up with a language for just describing how you would run those eval but it it sounds like it's it's fairly simple. you're probably looking for compute uh you know who somebody on beter will will build something like this like obviously you could use targon's like serverless as an example um or shoots probably has something or maybe Liam has it um where you know you take the repo with the p with the code in it um and you build a micros service couples CPU instances with that repo in it um and run an evaluation like Day Daytona as an example provides this um feature where you can create a sandbox with with your repo in it and then run the the evaluation in that inside the CI, right?

[00:49:28]>> Um >> just figuring out the exact nitty-gritty devil's in the details, but exactly how it >> when you guys were first talking I I I was really confused, but now I totally get it.

[00:49:41]>> So yeah, sorry about that.

[00:49:45]Well, it's good. Um, this is brilliant.

[00:49:51]>> Yeah, I I don't know how much time there is, but I I do actually also want to touch on seven is that's the new subnet.

[00:49:57]I don't know how this this works or how we should go about that.

[00:50:02]>> Well, yeah. I mean, it you don't have to split it up into into two sections. You could just stop on get tensor, but if you want to, yeah, for sure we can go.

[00:50:10]Let's go in.

[00:50:11]>> Oh, sorry. if you had any more questions. I I love talking on Git Tensor. So, if you have more questions, well, I just didn't know if there was like a time limit or something or this is like I don't know.

[00:50:19]>> Well, my first question would be, you know, and this is a bit of a a question with the dagger would be like why two subnetss like if you're still kind of doing git tensor, do you need another one? Is that the right move?

[00:50:34]>> Well, they're completely like different things. So, it wouldn't really make sense to have them as the same subnet, but I think your question is more so like why two subnets now.

[00:50:48]>> Mhm.

[00:50:50]>> Uh I wanted I wanted subnet seven. I like the I like the number and uh the idea.

[00:50:58]>> Yeah, the uh the idea was something that I I knew for sure I would do. So, I was like there's no point in waiting. And I saw a lot of people do like relatively similar ideas just completely like wrong. Um but yeah so the idea of seven is you know as the tagline says universal transaction layer. A lot of people are just like hey is is this an exchange? Yeah transaction is a subset or an exchange is a subset of what a transaction is. So you can come to always and you can okay let me let me actually step back in that explanation.

[00:51:38]Always is you exchange or you can you have an input and you have an output and the input can be any asset and the output can be any asset. You just specify the location of the output. So it turns into a universal transaction layer when you add essentially infinite trade pairs like that.

[00:52:04]>> Wait, so the miners are are are optimizing trade paths.

[00:52:10]>> That's one way to think about it. Yeah.

[00:52:12]So right now it's it's currently live with Bitcoin and Tao. So it's like you can go from Bitcoin to Tao or to Tao to Bitcoin. And we'll do that for for any assets >> and the miners make this transaction on behalf of you.

[00:52:26]>> Yeah, they they optimize the completion of that. So you would send a minor towel and they deliver your Bitcoin or whatever asset you specify.

[00:52:35]>> And what if they just don't deliver it?

[00:52:36]You know, like I take >> every every transaction is over collateralized. So it's game theory optimal for them to actually like deliver otherwise they would lose more money.

[00:52:50]>> Oh, so wait how is it collateralized? So like you actually construct a coll yeah they have collateral in a smart contract. Landon can talk more about that. He had to he's he's done some subtensor work. Um but yeah so you have collateral in a smart contract.

[00:53:08]That's essentially like the upper bound of the amount for a transaction. they they can complete and if they fail the transaction the collateral gets slashed and reimbured the user.

[00:53:22]>> Interesting. And and and can you even maybe put like uh some some like uh slippage guarantees on the transaction?

[00:53:30]Is that possible to to to stop a faulty transaction from burning all your money?

[00:53:37]>> I'll probably let Landon take that one.

[00:53:39]I'm not 100% sure. I'm clear on what the question is.

[00:53:45]>> Yeah, we actually just added slippage.

[00:53:47]So, like when you agree upon a rate between the assets, um then like you could have a 2% slippage or something in case that rate gets updated by the minor, but um you get your rate locked in for that that exchange. So, you won't get like rate changed and attacked as a user.

[00:54:06]>> And so, so why why is this, you know, I'm not the biggest DeFi guy. I don't really know my T5 well enough, but like what why is this such a big problem? I mean, like, hey, I'm sure somewhere out there there's an algorithm that just basically solves this within some percentage points um of the optimal solution just about every time. Is is that not just a pretty fixed solution?

[00:54:29]It's not just a canonical solution to this that you can just run an algorithm.

[00:54:33]And do you you see what I'm asking? like like how much better can we get than just the routing on MetaMask or some other platform that does this?

[00:54:45]>> Well, the the problem I think it addresses is actually the need to use like a a centralized exchange payment processor like a bank. So, it's it's more so about the the freedom to actually transact because in the future we plan on going to like TA to USD and one of the goals with it is like I want my phone and I can tap to pay with TA and the the whatever I'm buying all they need is to get USD.

[00:55:16]>> Wait, so I I just send a transaction to 77 and then some guy shows up at my house with a bag of cash.

[00:55:24]Um, maybe that would be possible. We'd have to be able to I would I would probably say for now it's going to have to stick to onchain things, but theoretically we we may be able to do the inperson stuff as well. It we just have to be able to verify that ownership changed hands. That's all we have to be able to do.

[00:55:46]>> Can you KYC entered the chat?

[00:55:50]Um, so, so like where's the where's the pain point? Like you said you can do tow to to to Bitcoin. Um, like I can do to Bitcoin.

[00:56:01]>> Yeah. No, no, no. The the pain point is um I I describe it well I guess the simple pain point is a bank can freeze your your accounts. A centralized exchange can freeze your arm ramps. people can tell you not to buy things, but then if you can go like directly peer-to-peer, there is no intermediary like that that can deny you the ability to transact.

[00:56:28]So, I would say that's the pain point, but like personally, it's because I want to hold all my essentially wealth in TA and just be able to transact directly in TA. So I only operate in TA and you can operate in whatever you want to operate in and I never have to change from TA because I can pay with TA and you receive your USD or Bitcoin or whatever asset you want to receive.

[00:56:55]>> And and and so like you said that there's like a a US dollar aspect to this. Okay. So I want to pay in tow for my coffee. um the eventual product.

[00:57:07]>> I would I I want to preface that by saying that is like future. We're six days into actual operation of the subnet. So it's just Bitcoin to tap right now.

[00:57:19]>> But but the the the goal is to is to have the miners find these optimal route routes. So you you say here's the end point, Bitcoin. Here's where I am right now.

[00:57:31]this amount of money needs to show up in this account and we're going to pay but and and so so the miners are going to be paid for making that transaction. Is that correct?

[00:57:44]>> Yeah.

[00:57:44]>> Efficiently and and so initially you're just going to be sending these transactions over to the miners. The miners do this, do this, do this and and you're going to measure how well they were able to do it. And you're also going to measure if they don't do it. So if they just walk away way away with the money that you just slash their collateral. So, you know, when it comes to the microtransactions that that you're doing, um they have no uh let's say like opportunity to to to mess with the the system because everything's verifiable. You know that it showed up in Bitcoin. You know that it showed up in Ethereum. You know that it showed up here and and so they can't cheat. Um um and you can kind of fire and forget because you know that it's collateralized. So if I want to pay 10 bucks, like if I don't get my 10 bucks on in Bitcoin or I get my 10 bucks in in Ethereum, it's fully collateralized and and I'm guaranteed to get some money back from that transaction. So is is that is that like in some sense it's very simple like there's no trust, you know, because they've already put money up at stake.

[00:58:42]>> Yeah, exactly.

[00:58:44]>> And and as as the collateral grows, the more there's more and more money that can be transferred through this through this network. It's it's kind of like a blackbox network is is what you're saying. Like you you don't care about bridges. You don't care about anything that it's just fire or forget as I said and and you're going to give me >> who likes a bridge. I don't I hate a like personally when I first got into crypto I I never understood why bridges were like a thing like to where you get like wrap towel. I I never personally liked that.

[00:59:13]>> So it's this is the the the omni bridge for crypto to begin. Um, and you'll be able to build an application on top of this that allows you to connect any wallet um, from any crypto in the world. And you don't even need to think about the back end. You don't need to think about it's on arbitum, it's on Ethereum, it's on Salana, it's on whatever the other blockchains are. I don't even know them all. Um, >> okay. And okay, that's that's cool. And would you do you guys have a you know what are you imagining the miners to actually do in the back end there?

[00:59:54]>> Lord Lord knows what a minor does. It's it's impossible to know exactly what a minor does. Um so that question I is impossible to answer right because like it's just speculation but what I can say is what we will reward for and that is the like credibility so the amount that they complete the transactions like how much you can trust that they'll put transactions through like 99% of the time they complete and uh speed so speed of transactions oh and also best rate.

[01:00:29]So essentially, we're it's going to be the place that's always going to have the best rate for every trade pair because it's it's a level on top of everything. So if Coinbase has one rate, Binance has another rate, the miner is going to use whatever rate's best for them, and then speed of transaction. So it's going to be a place where you can have the best transaction like rate and the best speed to completion.

[01:00:55]>> Okay. So you you do actually have an understanding of what you think the miners are going to be doing. you they're going to yeah as Capricia said uh halala subnet is actually a pretty good idea because because what the miners will likely do is have um exchange accounts on those different platforms in order so they'll they'll be managing money across these different platforms so that they can do the transaction fast enough correct right because they don't actually need to perform the bridge they can just you know if I want to move to let's say uh USDT on Tron, they can just have some USDT on Tron, make a transfer from USDT to Tron, and then just pocket the difference themselves. Is is that and then that's that's insanely fast.

[01:01:41]And so um I believe that's how the whole Wallace system works. Forget me, but it's it is underneath the hood. This is a peer-to-peer marketplace um for yet the delivery of funds as fast as you can. And it doesn't need to be it doesn't need to be the transfer like the miners are going to be a be the ones that are allocating the resources across the let's say the finance web. Uh they don't even need to do the the the transactions.

[01:02:13]>> Yeah. I mean you described it correctly.

[01:02:18]>> Cool. I have a good habit of doing that.

[01:02:21]Um uh I mean I I I mean when I first spoke to the get tensor guys I I was really really stoked about what what you guys are building and and like both of these these projects are you know they have have um great visions at their core. Um I am super excited to see what you're going to be doing with the git tensor like SDK you know abstraction on top of bitensor. I think it's a brilliant idea.

[01:02:49]Um obviously like I'm trying to build something very similar on Ninja but just for a single project that you guys have gone ahead and abstracted me to uh the nth degree which is exciting and then you know what you're doing with all always is I'm I'm happily surprised that actually makes a lot of sense um Mango and and and good luck in in taking this live.

[01:03:12]>> Thank you man. I uh that so that's the answer for two subnets. It was something I knew I would do and there was no point in waiting because, you know, you'll get a lot more progress just trying to to make it happen.

[01:03:26]>> Thanks for coming up on stage, guys. And and you know, usually teams have have powerpoints. We just winged it. I I I like that. I think that was fun. Um we got to get to the the core of it without having to meet having me interrupt uh the slide decks. Let's end it here, though. Thanks. Thanks to everyone for coming. Thanks for you guys getting up on stage and having this call.

[01:03:50]>> Okay, good night.