Bittensor Founder LIVE: Conviction + TAO Flow V2 Explained

本站添加: 2026-05-15

[00:00:00]ing.

[00:00:04]>> All right, ladies and gentlemen, if you're watching here, we have >> Hello everyone. Thank you so much, >> the founder of Bitensor Con.

[00:00:15]>> They do this every week on Thursdays at 5:00 p.m. Eastern time. Usually >> we are recording >> and we'll just wait what people come in.

[00:00:26]Um, they're going to be talking about two protocol upgrades to subtensor.

[00:00:30]>> AF indeed.

[00:00:37]>> I was like, why record it when I could just like live stream it? So, that's what we're going to be doing.

[00:00:40]>> Last time I was trying to get my music to play through the mics, but it was really bad, wasn't it, Mars? It was really It was really terrible.

[00:00:52]So, we won't do that again.

[00:00:55]Yeah, exactly.

[00:00:58]It's hit and miss. This court's really, really hit and miss, but is our home. It is our home.

[00:01:14]It should start in like two minutes. So, if you're watching after, >> just trigger the event um so that people come in.

[00:01:21]>> You can skip like two minutes and it'll start.

[00:01:26]people will show up.

[00:01:28]Lovely.

[00:01:30]Okay. Well, this is going to be um a slightly unique novelty search uh because we have two teams presenting um as opposed to one two for one.

[00:01:43]And uh that would be sub 14 and swarm.

[00:01:51]So, we're bringing Chachion and we're bringing Swarmup uh on the stage to present um some of the work they've been doing.

[00:02:02]Excited for that. And then there'll be a brief uh moment at the end where we can talk a little bit about some of the changes in the network um over the last week uh going into the upgrade for next week.

[00:02:16]So, one of the interesting things that's going on right now is that um the validators have organized themselves and are super burning um I don't know if people know about this but this is a it's actually a promising direction um spearheaded by Dan Talbot actually um and others in the validator community that that are just not having it with subnets that are self mining especially when they come straight out of the gate with if they have no code. Um, and I'm actually really happy to see that sort of like uh that those individuals coming in and and protecting miners honestly um by ensuring that that teams are actually running legitimate and fair mechanisms on Bit Tensor. So like if you're uh a team that's registering a subnet and you're looking to just do a premine and do nothing and and and send all of the emissions to yourself, um you're going to get super burned. And uh yeah, I'm happy to see that. like I I it's cool to see that step up from um from the validators to do that. Superb burn is basically when the validators u which have stake weight in initially on a subnet they um they send to a key which has been registered by by a smart contract uh and it actually sells the token um creating outflow. So you can't get emission from the protocol um if there's superb burn in place.

[00:03:41]Sir, it's 14122 and 114. That's right.

[00:03:43]114. You guys are also coming up. Hell yes, Oscar.

[00:03:46]>> Um, uh, thank you. So, let's get everyone up on stage. Um, >> let's start with Miguel. We got Max.

[00:04:01]>> Hey, what's up everyone?

[00:04:02]>> A bit of an echo.

[00:04:05]>> There's echo. Sorry, my dei. One second.

[00:04:09]It's a mirror.

[00:04:11]Wait.

[00:04:19]>> Hi, G.

[00:04:21]>> Hey, Miguel.

[00:04:24]>> How are you doing?

[00:04:26]>> Very well.

[00:04:28]Xavier, are we gonna are we gonna get you doxed as well?

[00:04:34]>> Why not?

[00:04:37]I'm going to be on my phone because I think the code is coming from my laptop.

[00:04:42]Ah, and then Max, are you coming up?

[00:04:50]Hey, could you bring Dei also?

[00:04:54]Uh, if he can raise his hand.

[00:05:02]I don't see him. Uh if can he can he raise his hand in the in the gentle then I can bring him up.

[00:05:10]>> I see his hand.

[00:05:13]White background, black square.

[00:05:15]>> There it is.

[00:05:19]>> Hey everyone.

[00:05:21]>> Okay. Well, this is this is giving me a great novelty search because we got we got three awesome teams. Um and uh we only have an hour. So I think we should go for it. And um well actually why don't we why don't we we all go off camera and and honestly let's start with Xavier right well let's go with 14 um and you can take the the spotlight here nice awesome yeah uh so 14 is launching on Tuesday uh what we're doing is uh inference optimizations uh so the analogy I've been using is that training a model is like uh designing F1 car uh And uh inference is like maintaining the pick crew and the race strategy. So you might have a really good car but if your pit crew is not that good you might not win the race. Um so so yeah so inference is pretty much you know you take a model and then uh the software uh figure out like the caching the memory and then serves you via a API endpoint. Uh so what the miners are doing basically is that they're submitting a docker container uh with the implementation of this inference server. uh by inference server is it's a software some people thought it's it's a hardware uh but it's auh it's a software uh you can write it in Python you can write it uh in rust you can write it there's also like sglang uh so yeah you can use like any framework any uh any technique although some technique might not be allowed in v1 because we want to get things done one step at a time uh yeah so uh what the validators do is that they take all the challengers or all the miners then they compare it against a baseline which is the latest VLOM uh and they're given the the same model the same cononical model which is Quinn for now but we'll expand it later uh the same GPUs uh and the same prompt uh across like everyone so everyone gets uh it's like a you know uh a fair fair platform for everyone to compete on uh and then we measure uh the time it takes for each miners to generates the first tokens and also your throughput which is uh token tokens per Uh so this measures like how fast your your inference your inference is. Uh yeah and we score everyone uh winner takes all because it doesn't really make sense for you to have the second fastest. We know we want the fastest for uh out of here. Uh yeah, really excited because even like uh it's been like four days into the announcement into like launching on test nets and we already have 26 uh UID signed up already has some 26 uh miners uh eager to uh to do this. Uh we even got one minor like made a dashboard for us. I you know our team hadn't had the time to uh make the front end yet, but the API is out and then uh that minor took the time to make us a dashboard which we're very thankful. Uh yeah, really excited to see uh how far this goes when we open up uh on the mainet. Uh yeah, so the road map uh right now uh the opt optimization surface is it's still quite big. You know, you can use any language, any framework. Uh but some advanced technique like uh spectative decoding um or custom kernels are are quite not allowed yet because we want to get it done uh one step at a time. Uh but later on you know the product is pretty much just selling uh selling the software you know we can plug into because it's a docking container we can plug into any backend um I think you know John from shoots uh mentioned us mentioned in our channel saying that oh we would love to work with you guys once you have our products you know of course right now we're pre-products but uh they thought about doing this doing this themsel um and I think he said that it makes a lot more sense if you guys are doing it so we can focus on shoots um Yeah. So, um, Mu has a really good piece of advice as Mu from the chat. Um, as a minor that has worked on a number of King of the Hill subnets in his time. Um, it I I've seen that that keeping the previous couple kings in the subnet is actually a good incentive landscape because you if they have to they can dethrone themselves. If miners can dethrone themselves, then they will push out their own submissions and continue continue to work. Um, if you just have one king, then the the top miner will basically stop. Um, and they stop working because why would they try to dethrone themselves? But if you have like say three or four or five slots that they can win all five of them and they have to improve their performance, they have to beat themselves. Then you keep that that engineer working. Um, so yeah, Muse points is well taken.

[00:09:58]actually is a good adjustment to a K koh k um instead >> awesome yeah I'll note that down yeah that's a good idea >> with the mechanism Xavier um you know pushing you're looking to find inference servers that can optimize the the tokens per second and and and tokens and time to first token for specific models. Um the goal would be that you optimize for any particular model that we're we're inferencing be Kimmy GLM etc. uh and if you can move the needle there like the intention for the subet is to is what transition into being an inference subnet that is uses these these optimized kernels or is it just going to be an optimization system?

[00:10:45]Uh so for now it's more of a I guess for now it's more of R&D uh you know it's going to be a arena where we surface uh the the best configurations uh down the line we want to deploy this configurations you know to to like enterprises to other people's back end uh in terms of kernel um so we we haven't really thought about it that much because we want to focus just on the software part for now uh but is it is optimizing surface surface that I think we can potentially open How do you beat VLM?

[00:11:20]>> How do you beat VLM?

[00:11:22]>> Yeah, exactly. I mean like and when when it you guys want to optimize beyond VLM which is you know like they obviously are some of the best software in the world there in in in terms of inference speed. There's SG Lang as well. um like for custom inference servers where where do they where do they compete and like in what and and how well can they compete with just like out of the box things like VLM?

[00:11:47]Yeah, I think I mean what I would do is probably search up the the current literature the frontier literature's uh so for example Turboquant had a paper uh back in March um you know I would love to read their paper read their codes and see if there's any way where I can strip out like the KV cache management from EVLM and then and then try to plug in turbo quant uh obviously you know I want to have a good understanding of how VLM work or how SVL work uh a good understanding of how turboquin Um but good thing is that they publish the paper, they have some code, they're available uh and see if you can yeah like strip it out and plug it in there.

[00:12:24]Um and yeah, submit it and see where where it goes.

[00:12:29]>> Yeah, I'm sure there's like other paper that are doing similar stuff but maybe in like the other part of the stack in the VLM or or inference server. Um, yeah, if I were minor, I'll probably I'll probably set up a agent that scans the current literature and see how we can piece this different different uh different configs to each other.

[00:12:49]>> And so this comes down to like the you're doing king of the hill. Um, you might do like glass five kings kind of king of the hill. They're pro providing a docker container. That docker container has inputs. Uh, they accept a machine learning model. they you load it into the the the server and then they run inference on that on that box or you run that inference on that machine um given some GPU. So like you give them a B300 or B200 or whatever and and and then they have to inference that model.

[00:13:16]How do you verify that they're not degrading the performance of that model by improving the by improving the the inference speed? Like I mean you could just trim out all the weights and just not run the model and >> and garbage. Yeah.

[00:13:29]>> Yeah. So, so that seems to be like the the difficult problem for all inference subnets is is how to know whether or not they they've actually run the thing. Um, you know, how are you approaching that?

[00:13:42]>> Yeah, so we thought about just doing the full KL distrib KL divergence. Uh, but then it turns out that it's uh it takes too long. Uh, it takes way too much resources as well. Uh, so we're doing a very greedy approach. First of all, we turn the temperature down to zero for now. uh and then we compare the first token where it diverges. Uh like you know you have the minor uh list of outputs tokens and then you have the baseline list of tokens. Uh for the first positions where it diverges we compare the top five uh token probability distribution and see how far uh they diverge from each other. Uh there could be some noise. So I think we have we have a small like margin of error uh just in case uh because you know we have um we have like multiple GPUs and when the GPUs communicate to each other there's there's floatingoint errors. Uh so this is another reason why we're fixing that that swap for V1. Uh and when we figure out a better solution we're going to open it up for V2. Uh so yeah so it's a it's a greedy approach because it's fast. Uh it's a it's a good proof of concepts. uh and for later on we're going to figure out if there's the best way to uh to do like a full distribution but uh I mean I don't think it's I don't think it's really needed to do a full uh KL divergence if if your token if your tokens are matching uh one to one for majority of it and then for the part where it diverges it's uh the top five token distribution between the two is it's quite similar anyway I think I think that's still pretty acceptable uh obviously I think >> I think you're going to see the miners just break that apart pretty quick. Um, but like this this is this is how Betensor works, right? It's it's a continuous process of that they will rapido and knows all about it. In fact, you should definitely be working with the team from shoots um on on on on taking that apart because they know all about Kale divergences and how miners can can manipulate things. Um the the the advantage that you have is that it's not blackbox, right? So earlier earlier inference subnets failed because it it's very difficult to do um uh it's very difficult to do blackbox inference verification in the the the first thing that miners try to do is they they they look up um a data set of rollouts from that same model and they run um they obviously can't do this in your specific case because it's a docker container right so they don't have a data set but uh in they look up if it's blackbox they can a data set of all the rollouts and then they just they just find the first roll out that does match the kale from the model and and then they just plug it in, right? So they actually just swap out a previously um a previously run inference from that same model with the with the one from the data set and they actually get a really low divergence except for exactly one token.

[00:16:30]So perhaps your structure is um is the way to go, right? you you you you're going to do this like look for the first first divergence in KL. But yeah, this is going to be very exciting to see.

[00:16:41]>> Yeah. Yeah. Uh we thought about this as well because you know we we've seen so many subnets where the minor figure out figured out like a very clever way to to game the system. Uh so that's why we're having a lot of you know work across that bridge when we come to it moments uh for V2. Uh I think we're going to see a lot of things that that may or may not surprise us. Uh and yeah, you know, just iterate from there getting feedback from the miners. We're we're logging everything uh the docker the docker uh logs from the miners the docker logs from the validators just so that we can go back and you know review them. So, when is this going to main?

[00:17:16]>> Sorry.

[00:17:17]>> When does this go to mainet? When do >> Yeah. Tuesday. Tuesday. Next Tuesday.

[00:17:21]May 19th. Very excited. Very pumped.

[00:17:24]>> Great. Okay. Well, let's let's pass the mic over to um uh Miguel because Miguel, you you've been battling miners for a long time. Um thank thanks. Thanks, Xavier. Um, and and really really you should have had your own novelty search, Miguel, but I'm glad you came up um on stage today and shared it. And you you also were on Mark Jeffy's show. He he pointed out that he had you first, which is well well done, Mark. Um, so I mean, for people that want more time to to learn about what what Miguel's been doing with Swarm, um, obviously go look at that that podcast from Mark. Um, but that's we have you here on stage. you know the there's no other subnet competitive that's working with drones and um like a little bit about like how you came to this design for the subnet and like and why you did it as you did.

[00:18:17]>> Well the subnet has gradually been improved. So the current design which is much better than what we started around 9 months ago is not what we started with right but basically the subnet was launched started because there was a a clear edge for bit tensor to solve this autonomous AI problems right now and I'm going to talk about the current state of the subnet we have several maps environment which we've developed and if you go anyone gets into Twitter they can actually see the the drones flying there and stuff and the idea is that miners upload the full code the full solution including code and including models. So the models even that they train for this they're miners which implement several models for one for one autonomous pilot and they go from the original platform till searching the other platform which of which they have an approximate coordinate but they don't have the exact GPS they use for this depth depth depth cameras and they use the sensors of the drones so that the reinforced learning algorithm that navigates the drones is able to are able to to actually get to these platforms. We have several maps mainly focused on snow and mainly focus on mountains. If you see them, see them basically because we want to close here in swarm the sim to real gap which is the holy grail of robotics because on on simulation you can say whatever you want and you can get to the results you want but you need to have these things actually in real robots. That's why we're building a lab here in Andor a physical lab. will buy drones out of basically every vendor who who sells them and we on board our AI with this with this drone with this on onto these drones. Why? To solve real issues that which can only be solved through autonomous agents because there are there are two two branches no of piloting drones. the ones the the scenarios which don't change for example inspection agriculture there how would you say the map is static right so it doesn't change through the course so there is not that much need for an AI to be constantly checking the environment but there are others that yes and that can only be done with autonomous agents for example and this is a project we are running now in Pandora and we are closing the the partnership with the largest ski resort here rescuing people in the mountain. This year it was a very snowy year in Andor. It snowed a lot and Pandora has a lot of of tourists tourists based on a snow and on the ski resorts. There was the issue that this year a lot of people got buried under the snow due to the avalanches.

[00:21:17]In order to search and find these people currently they do it manually. They get to the place they are called. Hey 911, there was an avalanche. Okay, the problem with this is that the time the time you have to rescue someone is very limited because if you dug bury them out, dug them out in under 10 minutes, the most likely outcome is that they survive because because it was quick enough. But if you wait for 30 minutes, the most likely outcome is that the same person dies.

[00:21:51]And it's a it's a matter of 20 minutes.

[00:21:53]Why? because they suffocate when the snow settles is in is like acts as an insulator. So you get no more oxygen. So you you have some minutes of the bubble you created yourself but then you suffocate. And in case you were lucky and didn't suffocate then you you freeze because you're under snow. You are not equipped for being under snow. The snow is on your clothes so your body doesn't hold your temperature. You freeze. This is basically a time trial test of searching these people the quickest the quick in the quickest way possible. For that, what we propose to the ski resort is to have several drones across the ski resort with the this AI that will be trained by the subnet and that once the avalanche occurs, they have kind of a red bottom which they can, you know, emergency and the drones start searching across the area directly quickly. So that once the the the team, the rescue team arrives, they already know the coordinates. If the the snow is safe, so you know after an avalanche another one can happen just like with earthquakes they need to verify that you know the snow is is is is settled. The drones can also do that because they can take images and they can do this this report and this is basically the project we are trying to to to do right now. um we'll do it from for the largest ski resort here in Andor and the idea is the lab so swarm is a laboratory which produces this intelligence and then we are going to create this use case but we're going to go for many more use cases and we'll expand as Pandora is a very tiny country if someone hasn't doesn't know where it is is the moment to to go to Google maps is between Spain and France we create this intelligence and then we sell it to the world basically if If the if this autonomous rescue works well then there are other 500 rescue resorts that might be interested potentially but not only limited to this but this is our first use case because in Pandora last year was hard one in this case and they are very willing to let us test to let us you know they are giving us regulatory freedom or you know regulatory willingness to allow us to test these things which in other parts of of Europe would be impossible because you will be talking to a governmental agency with 100 employees and it's impossible.

[00:24:29]>> Here you go. You can actually put these these drones to test. I I have some questions about like the the technical side of things. So you you you know specifically obviously you're going to test this, you know, in the wild, so to speak, right? So, you're going to put these out. Um, you have a partnership with with the Andorian Authority, um, to run these drones in, let's say, uh, in in real life. And, you know, but the way that you're training uh, the the um the drones right now is that you build synthetic environments.

[00:24:58]That's right. So, you you you have um RL environments. Those RL environments are basically like synthetic or like simulated um, worlds for drones to to fly in. And you know how do you generate those those environments and like is that scalable like you know beyond beyond just Pandora right I understand that you live in Pandora and that's interesting to you but like to to like the world at large what you know can you scale um the performance of these drones to work in any type of environment can they work in a factory can they work in a field of of wheat um and and like what what is the structure of the subnet that allows you to do that?

[00:25:35]>> Yes. Okay. So we have several maps currently. For example, we have a warehouse. So the map of a warehouse so that the drones doesn't don't get crazy once they are in. And we it's as easy as generating more environments. The current environments are kind of focused on the first problem we're tackling. But it's just basically like generating a video game map, something like that, where you can basically create the textures, create how do you want everything to be, and then design the challenge. right now is going from platform A to platform B. But the next the next challenge is going to be something along the lines of finding a mannequin a mannequin. So so that the miners start to create algorithms to detect human forms for example. No, the idea is that creating these exams is rather easy and and it's very fair and valid. So we basically measure three things. Success if they got if they get to the platform that's very easy. time.

[00:26:32]So we know the max speed of the drone and we know the coordinates at the start of and the finish. So there is a perfect time you can do and then from there on and safety it's not the same to fly safely without risks than to go for example 1 meter away of of a wall of a tree or things like that. That's how we validate and the idea for generalizing this to generic autopilot for for every aspect is creating more of of these maps so that the benchmark is more complete so more different use cases and where where has swarm you know if you take us like on a journey of the subnet like from from when you guys started to where you are now like you know how much better is the are these drones and like where do they stack up um against like other software that's doing the same thing? Like if I just go and get like an autonomous general drone um for this task like is there open source software for that and like do we compete do we compare do we outperform um on performance, speed, safety?

[00:27:45]The thing is that this industry is very closed. So there there is there is not much open source there is because once they get anything so I saw this in Poland when I went to a drone fair every business were was creating their own drone for their own problem with their own software and wanted to sell the whole package you know so there is not that much benchmark around this swarm has changed from the since the start massively maybe it's even embarrassing to say where we started but you need to start somewhere no at the start what we were testing is if miners could actually produce flight plans. So not models but rather how to to go how to go from point A to point B remotely do it however you want and just test that we can actually get the simulator going you know it was the first subnet of robotics at the moment so this wasn't as clear now there have been many more h then we went uh maybe at the end of last summer we started allowing the miners to upload a model so just a model and a refor learning model to solve everything. This was it was a a good step in a step in a good direction but it was too limited because then basically the manas could only create a model a small model by the way because this has to go in real drones. So you cannot have a massive LLM and this trillions model our models are small. This was a bit limited because there is there are many things inside the autonomous agents which do not actually need to be a model. They can just be code you know they can just be intelligent state machine. So the next the next big update which is the one where we are now is basically allowing the miners to allow to upload a GitHub.

[00:29:34]They just upload a GitHub. They publish it to the chain. Then validators unload it inside a docker which is isolated.

[00:29:41]They run it. Each validator publishes the the the the average score they got and then based on weight well the the winners decided and and that's all this is how the models have been improving but we've also run like more sophisticated it's a model plus software in addition to and and so so like that model software like when you when you peer inside these aer containers like what do you see like what what type of innovation do you see inside them that that justifies the mechanism.

[00:30:15]>> There was a big bump in that one miner discovered a way or created a model to detect the platform. So basically because when miners did some hard coding for example there was a funny part that we did the platform green so miners did if green go to go towards that. Okay so we fixed that. No. So now the color is random. Okay but by size. Okay. So the size is random. We went through this. So finally a minor created an actual algorithm which takes the the depth image analyzes it and tells you if the plasma is there or not. That's that that's actual AI and not hardcoded. Also they've created a very good state machine whether the for example the drone needs to go in the general direction of the search area and it's wondering and once it gets to the search area they have another step which is to rotate to rotate to see to see and if it rotates and it doesn't see anything then it wanders and it rotates again that's the way that's the current of the best model and also They've done a lot of hyperparameter tuning and stuff so that the model so that the drone flies almost at at max velocity right now in the start it was very slow. So yeah miners are miners are building on top of of each other. So yeah also the the good part of this is that with the with the agents that miners can on board more than one model. It's not required to have one model to do everything. They can have one model to detect whether the platform where the platform is one model of reinforcement learning to pilot the drone for example. they can they have much more flexibility >> in in industries beyond just drones like let's say robotics you know you have a robotic um you have a a humanoid robot inside of a a factory you want to make it a generally intelligent humanoid robot like this this field is moving very quickly um towards general robotics those are being run by LLMs rather than like let's say specific speific like um logic for determining like how to deal with a very specific type of problem where you're searching in an environment. Instead, it's like, hey, we're looking for someone under the snow and it and it deduces or actually reasons how to behave in that environment. Um like how how do we get to that general type of of um intelligence with these drones? So they're not just specific like let's say hardcode rules about like what you do when you get to a plot of land that you're looking to find someone.

[00:32:57]>> We certainly need to do more developments mainly in the amount of compute that can go on the drone that or if the of or if the place is very you know fenced. So for example is only this only work inside the ski resort then you could potentially have the the compute outside of the drone. That's also a uh that's also an option um in the for for now as we are trying to do an autonomous drone which does not depend on third parties that basically can fly alone. We haven't focused that much on that the the humanoids and basically any other robot which doesn't need to fly it's much easier because you you are not counting every gram you add to the drone you know. But here we we need to take we need to be very careful which what's the bare minimum of weight which can solve the problem because otherwise the drones lose lose range basically or or don't fly at all. So it's it's a an challenge. No, we don't have as much we don't have at the moment infinite compute because yes, obviously if if we had all the comput we wanted then we can put LLMs to analyze images and to do everything basically but also these things need to work remotely far away from >> so is this a unique quality of drones that they have to work remotely like they can't work with access to the internet like it's expected that they will have to be able to to like host their their their intelligence like to run their neural models on on on the on the drone is as opposed to say just access the internet and then using your reasoning model over there to determine whether or not they're going to move left or right.

[00:34:39]>> It depends. It depends. For example, inside a factory probably you have a a stable signal and and you can externalize it. But on the mountain, you're not going to have a a reliable signal. Or maybe you have the phone signal, but you're not you're going to have Wi-Fi. you're not going to have me multiple megabytes of connection and maybe the connection goes out and the problem here is that it needs to work on harsh conditions. This is a this is a problem we we we have right now it needs to work because people get lost right now. Okay, for this specific use case, not generic for everything, but for this use case, people get lost when there is bad weather. Nobody get lost, you know, with the sun out in the middle of the day. you know, they get lost on on the worst possible outcome. When there is bad weather, there is also in general worse connection. And when there is bad weather here, there is also very low temperatures. So, we need to get things to work in -20° C, for example, otherwise it won't work when we need to.

[00:35:41]We can test now in the summer and everything is very nice. So, we are trying right now to tackle this.

[00:35:47]However, it's true that for example for other use cases more confined more inside a warehouse then yes you you can do these things and also I think I've seen some X videos that people have already done it but that assumes very specific conditions and and it's not for every drone flight. If you want drones to actually fly somewhere, you cannot depend on having a stable connection and a powerful one because that otherwise it will crash when when you eventually lose connection.

[00:36:23]>> Miguel, thanks for coming up on stage and you know for people that want more information about that like go talk to well go listen to Mark's podcast. I think he probably went even deeper in all these ideas. Yeah. Um uh so the third third and and not least um Bitrex.

[00:36:40]Um thanks Miguel. Um so Demi you speaking or Max or both.

[00:36:51]>> Oh well let's get started I guess.

[00:36:54]>> Oh one sec. Here we go. There I fixed my mic. How's my audio?

[00:36:57]>> Nice.

[00:36:59]>> Great.

[00:37:01]>> So Bitrex um what is it? We're a product recommener built on Bit Tensor. What is a product recommendation? Well, they touch everyone every day. If you think about when you do a Google search, if you're open up your Netflix, you know, how does the system know what to suggest to you? If you're on Amazon shopping for your favorite items, you may not like the item that you're looking at. And as you scroll down the page, you'll see a bunch of recommendations, right? And so, this idea is is not new. Recommendations is a massive massive market. It's been around in e-commerce for for many decades now. There's a lot of players.

[00:37:36]There's a lot of uh total addressable market and so that's kind of where we're playing and we're operating in. Our background is in e-commerce. We have some experience there. So, you know, we thought this would be an interesting subnet idea on Bit Tensor. Up with the idea last year and we're on version two now. So version one was originally let's get the miners to actually you know answer these these live recommendation requests from e-commerce sites and you know we we managed to make it work um but obviously with Betensor you know there's exploits and all kinds of nuance there so we had to kind of wrap that idea up and um with the help of Yukon thanks we rebuilt the subnet to a winner take all competition and we also bifurcated it so we split this idea of live inference and intelligence, we we split we split it. So now we serve the live inference from kind of a traditional web 2 infrastructure that we have and then we have this uh winner take all competition running 24/7 where miners are submitting not agents they're submitting prompts into the system and then we run those prompts against an evaluation suite which contains both synthetic benchmarks you know industry standard benchmarks and some live holdout data from our e-commerce sites And so over time, yeah, >> you guys get you guys get a constant stream of of actual recommendations and the the selections, right, of what people click on. And so you you have a natural hold out set that you can use for this competition where people, you know, like you have the prompts. The prompts, they go through LLM. They go through cheap LLMs, I'm assuming, like, you know, Gemini Fast or or what, Quen 32. Uh and and you basically suggest specific um products for people on websites.

[00:39:28]Correct.

[00:39:29]>> Exactly.

[00:39:30]>> And so the the the it's actually quite easy to evaluate whether or not uh someone's prompt is is good because you just eva you you you take this hold out set and you run it through the prompt and um do do they recommend the thing that gets selected or do they recommend that high up in the ranking? Is that correct?

[00:39:48]>> Yeah. Exactly. Exactly. It's a pretty pretty straightforward idea. Um it's super effective and um it's the businesses love us, right? So our customers are businesses. We're selling to, you know, merchants, store owners, e-commerce businesses that aren't necessarily super technical and they they don't have, you know, the chops to install the latest LLMs and do all this stuff locally. So we made it super easy for them. It's like a one-click install.

[00:40:14]And right now, we're only targeting Shopify stores. There's 5 million of them, by the way. There's so many and um it's really easy for a merchant to get up and running. The cool part though is that this prompt evolves over time and it gets better and better. You know, we just launched V2 and we're on the early evaluation sets. They're really simple right now. We have, you know, over a hundred of these tests that we can throw out >> because it's constantly evolving because the benchmark is also constantly evolving. It could it could even take advantage of trends.

[00:40:44]>> Yeah, exactly. I think when you think about e-commerce, it's it is very seasonal and there's also, you know, certain events of the year that are really pivotal, right? Obviously, the biggest one is like Boxing Day and Black Friday. When it comes to e-commerce, those are like where 80% of the sales are done for a lot of these stores. So, having certain events, you know, there's Mother's Day, there's Father's Day, there's all these kind of worldwide events that happen. Having the LLMs be able to understand the context. Oh, we're actually in the time frame of, you know, Mother's Day, so maybe I can recommend flowers, for example. That's kind of an idea. And so, we get that naturally a lot from the LLMs. It all comes down to the context for us. And this idea has opened up because we got large contexts now, right? A lot of these models, 128 is kind of the baseline. 256K context. We're seeing a million contexts now, I think, with DBC4.

[00:41:33]That really opens up a lot of interesting use cases for this stuff.

[00:41:36]So, we're excited.

[00:41:39]What's the like headline number here?

[00:41:40]Like when you guys I mean what you're optimizing really is a metric that can be easily converted into dollars and you know if you go an extra percentage point in terms of accuracy that can mean you know for your clients tens of millions or hundreds of millions of dollars like especially in terms in advertisements.

[00:42:02]What exactly is that number that you guys are looking at? Is it is it like it's it's not clickthrough rate right?

[00:42:07]Um >> it's yeah it's it's lift. So we're just trying to generate something called lift which is kind of a general term that you can apply to any online business but in the context of e-commerce it's it's really are you increasing first of all are you having the users click more and engage more? Are they spending more time on the site because of your technology?

[00:42:24]So that's more engagement. Are they adding stuff to their basket? So that's you know you're increasing their basket size which ultimately leads to something called AOV increase right? Their average order value increases. And those are the metrics that e-commerce sites really care about, right? Ultimately, they want to know is is this thing generating me more engagement, more clicks, am I selling more with this technology and it is it's all measurable. Everything's like uh it's a very mature industry. So there's a lot of tooling around this.

[00:42:52]>> Is it measurable via your subnet? Like if I came to you, you know, two months ago and you said, "Hey, this metric is at 20%." And or like it's 1 in 10 million and now it's two in 10 million.

[00:43:02]um you know can you translate that into real dollars for your customers?

[00:43:10]Um yeah, I mean we'll as we have these prompts, so yeah, we evolve the prompt over time. We take that prompt and we actually have we have an AB test running on our production site. So we'll in we'll inject the latest prompt from the subnet into that and measure it. And this is all done through Google Analytics and all that. So it's not actually directly tied into the subnet.

[00:43:29]So our customers will see it and our the extension in Shopify. So when you install our our product into Shopify, the admin of the store or the store owner sees all those metrics. We're not pulling that data back into the subnet right now. It's something that we thought about. Um the first version actually had that built in, but again, we didn't uh we didn't roll that out.

[00:43:49]So, and because that would be the the headline number, right? If you could say, "Hey, look, we're you know, it's an obvious thing. You pay us $10 a month and we make you$10,000." Um it's it's an obvious sell selling point. Like what what is the like when you guys are measuring the performance right now?

[00:44:04]like what is the core metric that you you measure and h how much has it improved since you started optimizing with Bit Tensor?

[00:44:12]>> Yeah. So, we could talk there's two parts to that. The first part is the actual e-commerce metrics which is just straight lift, >> right? And ultimately, if you're over one to 5%, you're doing good because for the e-commerce owner, all that traffic is already there. It's no cost to them.

[00:44:28]It's just it's an added bonus to their business, right? It's like stuffing more people into your store basically. And so we show them Lyft. Oh, this month we generated 2% more of your sales, for example, and that's what they key off of. As far as the actual subnet goes though, we're using kind of standard Rexus benchmarks.

[00:44:44]>> I think Max, if you're online, you want to talk through some of that?

[00:44:47]>> Yeah. So, the industry uh commonly uses this one called recall at K where that K is just like the number of Rex being returned. So for instance, say we're asking for 10 rex, we would have recall at K and we ask in the old in the old V1 system, if that minor returned our ground truth in that K, we say great and that recall was a binary val. And then the other interesting one is NDGC at K.

[00:45:21]And then we're basically trying to get most relevant uh product returned at like that index zero slot, >> right?

[00:45:30]>> So, um the Rexus there's actually a new conference coming out soon. So, there might be some new um methods, but yeah, we're seeing over time that the the base value is going up. So on our dashboard um you can actually track that value trending over time >> and how do I go there?

[00:45:59]>> Did you say how?

[00:46:00]>> Uh dashboard.ai >> dashboard dot >> Yeah.

[00:46:13]Someone else help me on the chat.

[00:46:23]>> I had it spelled wrong.

[00:46:27]>> Launch dashboard. Nice website.

[00:46:30]>> So on the dashboard, you you you have a nice radial graph here. Um there's a whole this is uh Garfield, Monroe, Ulyses, Draft, Fish, Preliminary, PSD, TST. These are all minors and their performance on different >> different valves of the current evaluation set.

[00:46:52]>> Okay. And so you you actually maintain um like a covering of different prompts and one not there's there's not one that has performed well on all of them.

[00:47:06]In this instance, Evvel said two. No, the last one kind of got solved, but uh we had to catch like uh we talked a couple months ago const and you were saying like you have to be careful about miners overfitting their prompts to really like nail the evals and so there was like a sneaky edge case that kind of avoided our first set and now they're getting stopped out um on some of the longer context evals.

[00:47:38]So we're we're pivoting to a a bring your own key to also expand the number of because one one interesting thing is it's not just a prompt right they get to define their temperature the model >> all the different parameters >> that you could possibly we lost you there Max got some feedback on your mic you might have to mute that >> but yeah just to finish up what he's Um the evaluation set is evolving over time. We've just launched V2. We've only put like six or seven of our evals in there. Uh we have hit like kind of uh a scaling limit with shoots. We're getting like a lot of 429s.

[00:48:23]Um and so we are going to open up open router with our next our next release just to alleviate some of that traffic on shoots and to allow you know at least the miners to compete on a fair you know open evaluation. And then again, these eval right now are, you know, five samples. We've only got seven loaded.

[00:48:40]So, we're going to really kind of load these up, 50 or 60 eval going with way more samples. And hopefully we can get this prompt up to a place where it's it's it's state-of-the-art. We'll see.

[00:48:51]>> What does the prompt look like right now? Do you guys want to try?

[00:48:53]>> Um, yeah. Can I Well, you can pull it up if you're on the dashboard. Let's just you go to the the top the top winner right now. I think his name is draft.

[00:49:05]>> Okay, let's try.

[00:49:07]>> So, click on draft. If you scroll all the way down the bottom, you'll see user prompt template and system prompt template.

[00:49:13]>> Okay. Okay. So, I I I'm on the dashboard. Okay. And I'm going to go all the way to the bottom and I'm going to find sorry. Uh did >> you find the leader? So, first find the leader which is his name is Giraffe.

[00:49:25]It's UID 81.

[00:49:28]>> Click on that and scroll down the bottom.

[00:49:32]And you'll see user prompt template.

[00:49:36]>> Yes.

[00:49:37]>> And then you should it's like collapsed.

[00:49:40]You can like expand that.

[00:49:41]>> Yeah. I found I found it. I found it.

[00:49:42]>> And so that's the current prompt right now. It's pretty good. It's not bad.

[00:49:46]It's better than the one I wrote.

[00:49:49]>> Guiding principles for appeal variant cataloges. Treat apparel items with shared SKU roots and different and skin and size color prefixes as color size color variant grid not as an outfitting building catalog. How do you think the miners are iterating against this? Like is there a just a loop where they run the evals over and over again?

[00:50:10]>> Yeah, that's interesting you say that because I actually have an auto research loop running right now on our own that I'm running this against and I am seeing improvement. So, I would assume that the most savvy miners are doing this. Um, it's hard to really tell though to without actually seeing their evolution, right? We just see the final prompt that they submit. We don't see how many private evals they did themselves, >> right?

[00:50:33]>> Um, but I will say the prompt is getting better from from what we're seeing from the evals. Uh, there has been some overfitting and some hard coding. So, it's this constant battle of, you know, okay, let's let's shrink the allowed tokens they're allowed to put in the prompt down. Let's increase the reax to make it more strict. And I think we're going to land on a place where eventually we get a good kind of balance. The question for us though really is is a general recommendation prompt the way to go or should we actually um slice it up by taxonomy?

[00:51:04]Meaning in e-commerce there's all kinds of categories, right? There's clothing, there's home and houseware, there's TVs, electronics, etc., etc. You know, maybe it's it's more sensible to have a prompt competition for each category. versus trying to do some super general recommener master prompt, right? Um, that's something we're going to answer this year. We'll figure that out.

[00:51:27]>> Great. Good luck. Thanks for coming up on stage. Like one of the things I really like about your guys' subnet is that you know in in the advertisement world like um you're very very close to where you can string a line between like optimizing the imperformance of this means dollars for my customers and like it's a it's a very clear linear argument to be made. Um so it would be really awesome if you guys could get that working um and and and showcase that to your customers and really sell that.

[00:51:58]>> Agreed.

[00:51:59]Cool. Um, okay. Well, I guess there's 15 minutes left on the call. Um, I think a lot of people here are are here to also hear about um like some of the the the upgrades we're making um going to next week. Um, one moment.

[00:52:20]>> Hey, Jacob. How are you?

[00:52:22]>> Very well.

[00:52:24]Um I was I noticed I know you were going to change topic which is fine but I I thought maybe I could talk a little bit while we're doing Cathedral if if that was there's time for that.

[00:52:33]>> Five minutes. Five minutes and then we'll end it because it's good to have some some teams on stage because we can do an AMA about conviction. This is the last call about conviction that's going out next week. Um there was some calls that I got uh this week for people that are they're curious about this upgrade.

[00:52:47]like they have some questions and like there's a couple like headliners that I want to get out there just before maybe you take the last five or seven 10 minutes on on the call. Uh the first one um is around how conviction is going to be loaded on subnets and um how it's turned on. So um specifically subnets are not going to have conviction on until there is 10% conviction on the subnet which means that 10% of the outstanding supply needs to be locked for 2 months roughly before uh that is triggered on your subnet. And this is this comes down to the hyperparameter specifics but this will be something that all teams can monitor. And if people on the subnet do not choose to lock their stake, it's not something that a team actually needs to think about at all. Um, so it's sort of triggered and created by there already being a contest for the ownership of the subnet, which is something that will be uniquely seen and visible by teams uh before it happens. So, so if you are in a position as a subnet owner that you don't want to participate in that and you don't think that your community is going to want to pressure you to participate in conviction ownership of a subnet, it's not something that's on the agenda right now. It's not going to affect you whatsoever. That that's I think you know maybe a sigh of relief for some people. Um still allows for contested subnets for that to happen. Um so where there are subnets where there is there is con there are token holders that are contesting the subnet um there that vote can be triggered by people locking and actually putting conviction into the ownership of the subnet by locking their stake and that will be something that's visible and doesn't just happen on instantaneously but but shows up it's not 10% of all staking all keys um Keith it's 10% of the stake um terms of the conviction convergence to and asintopes the value conviction asintotes the actual amount that's locked. So if you lock 10% of the subnet alpha, your conviction would go from zero all the way to 10%. Um so those things are similar and and we can measure conviction in terms of outstanding supply. So like for instance, if everybody locked all of their stake perpetually um you could get conviction that was equivalent to the outstanding simple supply. There's also some some um handy um tooling to make it easier for subet owners in the event of um let's say uh there's concern about like let's say a hostile takeover. Um the subnet owner has the ability to immediately get the maximum amount of conviction for the subnet. Somebody locks 10% of the supply. Um it takes them about uh two months for that conviction to mature in that time. If the subet owner convicts the same amount that they have locked, if they lock the same amount that they've locked, they can immediately get the the maximum amount of conviction.

[00:55:30]So, it's always a game of can the subnet owner just lock the amount of stake that is being um put at play. So, in basically every instance, it it it's b it's it's possible for the team to quickly and easily defend any let's say like nefarious or spurious attack. Um um in addition um you know autolocking for the owner key is possible um uh and you know unlocking and relocking is possible um sort of like standard operations that you would expect. It's it's actually a very beautiful design. We're very happy with it. It's quite elegant technically.

[00:56:05]Um so we're excited to see um we'll have more details on that coming out next week and people can play with the website showcasing how it works.

[00:56:14]The other upgrade that is going to be coming out next week is a tweak um to tflow. Um the specifics of that tweak are that we will be subtracting the net subsidy from the net flow. So that means that for every towo injected into the subnet as subsidy from the chain as in you know liquidity injection andor chain buys um counts against tflow and you know that might seem negative to all sub bit tensor but it's negative in in equal part to also sub bit tensor. So the renormal after normalization it it merely changes the distribution of the tow flow across the subnets. Um and that will be coming out next week. So this is this basically says if you think about intuitively it means that tflow was a measure of the net flow into a subnet.

[00:57:07]This is a measure of the net flow minus the subsidy from the chain. And what this does is it subtracts the bias created by the subsidy. So if people are b if people are buying just because of the subsidy that signal is removed from the Tflau signal and and what this does is actually adds crytosis increases the steepness of the emissions. When you do get emission you'll be getting a bit more emission um uh and uh but but it specifically says like if you're just a subnet that is that is getting just just liquidity injection just subsidy from the chain um then then you don't have any emission over time. So th this is um this is like a a slight a slight tweak to tflow that's going to happen next week which is which is interesting and I think that um will affect and improve tflow uh that in in in in conjunction with conviction um there are also there's also a change to allow subnets to turn off emission for their subnets. So it's not forced uh and it's not um something that is set by validators for instance or other subnet owners. Specifically the subnet owner can turn off the emissions for their own subnet. Um uh and this is this is uh there to enable teams that um have been sitting with a mission not doing much and they feel like they don't really want to have a mission. They'd rather give it to the rest to to other submits on bit tensor. um they have that ability to actually stop injection from uh injecting into their into their subnet. So this is interesting and I think exciting for for for like moving towards even more control for sub owners to control their emissions.

[00:58:50]Another change um chain buys historically have been a purchase from the chain and then a burn of alpha. We will now be switching that to a purchase and then the accumulation of the alpha that has been bought through the pool and that uh amount of alpha will be attached to attached to that um subnet so that on dregistration that those alpha tokens are paid back to the chain. So the chain is not just burning uh its own ownership of the of the tokens that it's purchasing um through admission. It's actually attaching those to the subnet and then if that subnet is dregistered uh those go back to fund and then burn the towel uh which um actually specifically recycled the the tow that that it had previously bought into that subnet. Um so this this is an important just sort of like balance of of um like let's call it like fairness between the chain and and the liquidity injection. So if a subnet that has got a bunch of chain buys through let's say manipulation and then goes to a dre registration in order to pull out the max funds or like to extract from bitensor um bitensor itself pulls back a good portion of those funds um that is was accumulated through those chain buys. So that's an important aspect. Um, some upcoming things and like this is a bit of a conversation that we're going to we're going to kind of uh bring out into the community is uh the potential removal of root dividends uh and go towards a pure burn mechanism on on root dividends. Now this is something that I'm actually extremely excited about for Bit Tensor. Um the reason for it is that um we think that this will move more um more investment into the subnets and less uh passive staking on route.

[01:00:48]In addition, it also truly sells and explains the mechanism in general. I think many people don't fully understand that the bit tensor blockchain actually makes revenue from its incentive um which comes from the root dividends. Uh and by burning those root dividends we can we can uh let's say analogize um the way in which bit tensor actually makes revenue by burning its own um tokens through increasing the prices of the subnets. This is this is in conversation. So I'm I'm actually stimulating this this conversation here on the on the call. I know people are going to talk about this um and there's no better place than than giving it out on the search. I'm sure there will be lots of conversation. This is not coming out next week. This will be something that is that is discussed um at infinitum by the economically minded people in bitensor. But I think it will be extremely positive for for people understanding um uh how the economics of bit tensor the bitensor system works like and note that as the the sum of subnets the sum of subnets prices increases on bit tensor it actually means that bit tensor can go purely negative uh in terms of purely deflationary uh in in in terms of tow emission which is I think an exciting pro um pro prospect for for the chain itself. So those were the major upgrades. Um there is also governance coming coming uh in in let's say the next the next few months. Um and and also changes centralization but nothing more there. Uh these are just the changes these are just the changes that are coming out next week. Very excited to to show those next Thursday uh when we do the actual upgrade. Yeah. Okay. Um well then, hey, let's talk Cathedral.

[01:02:37]>> Oh yeah, thank you. Um I hope you're all well. I'll just be really quick because of time. Really three things. Um I want to explain what Cathedral is doing or at least what I'm trying to do and then sort of talk about why it matters real briefly. So what how it works today? Um like people submit agent profiles. So what do I mean? So miners submit like a Hermes agent where we are we are locked on Hermes um as an agent stack for now.

[01:03:00]Uh so that bundle involves like soul skills memory model and when we get that bundle they also submit like the the the node or the the computer is running on and so we we log into we we ingest the the the module they've submitted the profile they've submitted we log into the machine and we we execute a series of tasks on that on that agent evaluate it sign it uh score it and then give um give uh the score to validators to to to to reward. Okay. So, what I've described is not necessarily a subnet.

[01:03:34]It's actually just a mechanism uh for evaluating agents that live on boxes.

[01:03:39]Okay. Um and so the the subnet then becomes like these jobs. So that was the first thing I wanted to build was the the the the mechanism. And why is that mechanism interesting? Because it means that we can actually see exactly how an agent has been configured relative to the job that we assign. Okay. and and so the job is kind of important and I'll talk about that briefly but like just think of all the data you can get from logging into the box. You can look at um tool calls um you can look at um um how how they how they configured their their their their prompts what sources they used. So we're getting like a really rich source of data that we can use for training for pairs for distillation for for for many things uh the context the outputs um and so much more and so like that's the that's the core structure of the mechanism and so the the interesting part the design then becomes jobs which is kind of what I've been working on.

[01:04:38]The first job we had uh was just hey go look at this EU regulation and write summaries of what's changing or who it's impacting and send that back. But that was just like a basic job to help us test the mechanism. Our next job is we're going to we're going to work on like bug reproduction and test suites.

[01:04:54]I'm designing those now for us to to launch it. Um and you can if you go on cathedral.computer you'll see like a little a little wall that we're building a different kind of wall. Um, and you can click there's a bit of a bug there, but you can click on the first the first the first block and you can see an example of a minor submission from the job we have now using the mechanism. So really once uh we design like hey we want to do jobs we want to do jobs for bog production we can ship that um miners uh work to produce um um you know work to produce um um assets relative to that specific job. Um, and then we verify and we reward. But the real key thing is just like previously when I built Polaris, we we were having you people give us compute. Um, and we would rent that compute. But now we want people to give us agents in live boxes and we have these live entities and we can verify, hey, this agent did this work and we have this trail of the capability of X agent and we have the agent bundle so we can actually rerun that agent and we can actually have that as an asset um that we build over time.

[01:05:58]And so you imagine the kind of train data we can get there. um for for other parts uh of of the network whether it's training or or and so on and so forth.

[01:06:06]Um this is like super briefly what we're trying to do. Um so the the miners are um upgrading Hermes agents.

[01:06:18]Yeah, the miners are are creating custom Hermes agents per job we define >> and they do that through actually writing Hermes Hermes code or by >> yeah they they go finish your questions >> they they they they depending on the job they write skills they have config they configure specific tools they configure their particular instance of permes like the soul they like they craft an agent that does something really good um and submit that to us um why did I end up here. I end up here because I wanted to build a subnet that is minable for basically anybody. And so somebody can be lazy and use Opus 4.6 to power their Hermes agent and submit that to us. And somebody may not have that kind of access. And so they can do really good prompting, really careful thinking, really good user tools, um, and craft an agent that uses a small model but works really, really well. And they only need a CPU to run to run Hermes. Um and so it's like it's like challenging that idea that you know more compute equals better outcomes. Um yeah, >> thanks Paul Scaler. Okay. Um are there any questions for the chat? AMA um for for any of the teams on on the stage any of the updates that that I I made on the call here? Um I can answer those now.

[01:07:41]But conviction will be implemented in some way powered by lock. Yes, conviction is based on lock. Yes, Jake.

[01:07:46]Why is conviction not proportional to the pool depth?

[01:07:50]I actually don't know what you mean by that. Uh Sam like why is conviction going to be related to the pool depth?

[01:07:59]Oh, as in like the the conviction is is like the simulated swap of the amount that you're locking.

[01:08:06]uh I don't think that really matters because in conviction is only measured with respect to other people that are also locking with inside the same subnet. So that's that simulated swap is going to be the same for everyone. You know, the slippage is the same on that subnet for everyone. I suppose actually some people have more slippage because they're locking more, but I'm not sure if that scaling really matters. Is that was that the Was that the question?

[01:08:30]Um any plans for improving liquidity management V3 V4? We're actually we actually have a push for balancer pool sebi um but it's not right now it's not the like the tip of the spear in terms of what we need to be implementing on potensor and so it's not it's not our largest issue right now. Um as in a whale can just rock up with more money than the owner and convict more than the owner. Um yeah so there's a couple things. Um, a subnet uh can only be in conviction after one year. So, they should have pretty deep pools if they've not been dregistered in a year.

[01:09:08]They're also um a fairly liquid team um and they have a fairly high price. So, it it's it's quite difficult to do that.

[01:09:17]In addition, the the subnet owner can respond um by locking. And in fact, I think we will tend to see that that that in some cases it's actually the the the holders of the subnet that that ask the teams to lock through this mechanism. Um um Nikolai, is it possible to elaborate on the motivation between um of zeroing root AP API? Um the the the motivation uh is is twofold. Um, one right now is I I would say like primarily is to shore up and prove the economic um, argument um, for why and how Dynamic Tao provably improves the economics the token economics of Bit Tensor by exchanging for its subsidy optimizing for teams that can pay the protocol more than the subsidy is. And so it becomes a very clear argument for Bit Tensor itself how improving improving the subnets and increasing the subnet prices in increases root yield which is actually just purely burned. Um I I think that paying a dividend um in crypto is a little bit less understandable than just doing um a standard burn. And I think that it can be net deflationary for for Bit Tensor and that's why that is being argued um for and I I look forward to your argument for why not. I'm I'm assuming that there are reasons um Nikolai that you are not interested in that. Uh I think that it will it will also drive root stakers towards investing in subnets which is where we want to see capital move rather than sitting uh passively on root.

[01:11:02]Um T slice do subnets under one year old have it? Uh no. So the yeah initially root was especially valuable because we needed the root validators to stabilize the we stabilized the subnets on the launch of DTA that dynamic has substantially changed um since in one year into DTA and it I think that it's not um as justifies justified as it was before um and moving towards um uh higher yield and uh more demand for subnets is probably net positive for potensor.

[01:11:46]If root was validating the chain, it would make sense to reward, but that's not the case.

[01:11:51]Exactly.

[01:12:15]It also should be the case that a lot of root validators have a large number of alpha tokens at this point. Um I in in in in most cases active validators especially um validators like yourself um Nikolai you do have a large amount of um of alpha on your validator.

[01:12:36]So the validation is not is not is not done over. In fact, it just transitions to primarily being in terms of alpha tokens.

[01:12:46]But let's leave that conversation. I know that was quite a that was quite a boom and perhaps I should not have um um yeah brought that up on this call. It's not relevant for the next week and and like I said, the purpose was to actually stimulate the conversation, which looks like it has, which is fantastic. Um uh and and we can uh we will be discussing this more with with the community going forward.

[01:13:11]Cool. Okay.

[01:13:15]Thanks, Nikolai. And thanks everyone for coming to the call. Uh it was great to to see like a plethora of teams, you know, Xavier, Miguel, and uh and to learn about Cathedral.

[01:13:26]Thanks everyone. Take care. Have a good night. Bye everyone.

[01:13:35]All right, my friends. That was the novelty search. Um, if you were curious on what I was doing with my pastime, uh, right now I am making a bit sensor emissions website. The reason for it is that Toflow V2 is coming out and I think a lot of people are kind of confused on what it is. So I am trying to build something in plain English for that. My thinking is that I'll probably like code a new website every time I have one of these bit tensor streams. Of course, my pleasure. Absolutely my pleasure. I think a lot of people are looking for novelty search. They're curious on where it is, why nobody's streaming it. So, I'm just filling in that hole and doing it. Um, also at the same time, I have the notes that I got, which I think will be a little bit more useful than what you just saw there. Um, this is what I'm going to be posting onto Twitter, so you're getting it a little bit earlier.

[01:14:26]So, the main points for everything. See, does this look better? It's kind of the same thing. So, main points, um, hyper burning. So, any subnets that are doing hyper burning, which is burning 100% of everything, um, they're currently getting protested by validators, which are saying, "If you're not doing anything on the chain, we're going to be selling our coins essentially." Uh, validators typically speaking, hold their coins with productive subnets because the upside is higher and they also want to support price stability.

[01:14:59]So, that's kind of it. Uh the upgrades got delayed. They were supposed to be done today. So Conviction was supposed to happen today. It ended up going into next week. So we're uh going to see when it comes out, but it's supposed to come out next week. So in terms of conviction, it's not going to work on subnets that are less than a year old.

[01:15:19]And also they need to have 10% of the total supply staked for two months to build enough conviction to replace a subnet owner before it turns on. So interesting. If people in the subnet decide not to lock their stake, then you don't need to worry about it, there only needs to be an active competition for the subnet. So, someone needs to try to take this thing over and all of that will be very public. You will know in advance. So, um if the subnet owner locks the same amount in those two months, um they will retain ownership.

[01:15:50]The idea is that if if a subnet owner has been around for a year, they're going to have plenty of towel. They're going to have plenty of alpha and they're going to be able to easily retain ownership if they are acting in good faith. Another thing, subnet owners will be able to turn off emissions for their subnet. Um, this is there to enable teams um that have plenty of emissions and would like to give it to other people. Maybe they're not using it for anything right now and they want to have the good faith of the tow ecosystem. So, it's optional. Um, TWFlow V2 I thought was really interesting. Um, pretty much in essence they're going to make the staking rewards for root which is just like staking your towel a lot worse. The idea is that by making it worse it's going to encourage people to stake into subnets instead. Um, initially root was very useful which is like normal tow staking as these like subnets came alive last year. Uh, but now they don't think it's very useful.

[01:16:47]the subnets are validating the chain perfectly fine and it's not as justified as it was before. So all these validators are going to be moving from root over to um alpha which is all of the subnets that are here on Bit Tensor.

[01:17:00]He thinks it's going to be a net positive for Bit Tensor and whatever APY you would have gotten from TAL um will actually just come from the price appreciation of Bit Tensor as a result of the subnets having more capital to go ahead and compete for. So that's cool.

[01:17:18]Um other things governance is coming in the next few months. Also stuff about chain decentralization. He did he just didn't want to say more about it. Um, and also I thought this was an interesting thing that someone said and I copy pasted it. It says 100% support.

[01:17:32]Why pay tow holders extra rewards if they're not contributing risk or market information um into subnet economies?

[01:17:38]Remember the only reason to exist is because of these subnets. Without the subnets, this chain would be pretty much nothing. Also, naturally, as subnet prices go up, root APY will go up. Um, you know, as someone that is mostly in root right now, this is actually really good news because if they increase the amount of rewards on subnets, then, you know, I'll be more encouraged to be on subnets because more people will naturally gravitate staking into subnets versus just having it on root. Uh, so root APY effectively hampers incentive for alpha holders in DTA. So that's cool. The incentive for them to hold TAL is already high since the demand for alpha is demand for TA. So any buys on these alpha tokens is uh demand for TA.

[01:18:16]So, those are my findings. I hope that you found value from the stream. If you did, leave a like so it gets out in the algorithm. And I will see you all in the next one. See you.