Bittensor's Bull Case with James Ross & Mog

Transcript

so what detail does is it turns that power over to everybody pure democracy but in the context of these incentives you know synth is live does that mean synth is also going to have its own token so yeah 250 you can have 256 minus all contributing different models I mean the industry is evolving at such a fast pace I think even AI open AI can't answer the question welcome back to the rollup your home for technical talks and open conversations that shape the modular future nothing on today's show is to be construed as Financial advice you can see our full disclosures list at the roll up.co Ventures sit back relax and enjoy today's episode after a quick word from some of our awesome Partners imagine if the world's GDP moved on chain guys say GM to Safe net

the newest protocol by safe isn't a blockchain and it's not a new layer 2 it's a transaction processing Network built to send money anywhere and swap anything on thousands of layer twos from a single Universal balance in sub 500 m seconds it's fast it's smooth and it's simple and it's built with safe multisig level security in mind safet unlocks smart accounts across multiple chains allowing users to supercharge security scalability and speed so get started today at net. safe. Global and remember to follow at safe onx GM GM GM GM we are the rollup today MOG we're going to get absolutely mogged

by MOG I already know it and James is here to rep representing mode um this one's going to be super fun James you know you're a friend of the show welcome back um MOG it's your first time so uh very very cool to have you on and uh very excited to hear um a little bit about you and then ultimately get into our subject today which is going to be about bit tenser and the transitioning that's happening there talking about decentralized and distributed compute and then we're going to be talking about synth the new subnet on Bit tenser Lots to cover so let's let's Dive Right In MOG welcome first time thank you very much glad to be here absolutely James thank you for

coming back how you been yeah uh things are good you know we keep building that's uh that's what we have to do oh yeah that's what it's all about um and so today distributed compute I think the whole world of compute got thrown for a whirl a little maybe about a week ago with deep seek we kind of saw Nvidia react to oh maybe we don't need billions of of of dollars of compute to train these models and so Nvidia kind of reacted um ultimately we're still going to need plenty of compute uh it's just there's a parallel there to like data availability these things tend to just fill up and that we need more and more

and more as use cases get more expressive um and so when we're thinking about distributed compute MOG we haven't had a whole lot of deep coverage on bit tensor and and the way that it works how it works uh what people use it for so maybe you could kind of give us a Prelude to bit tenser before we talk about some of the new and exciting things that are happening there yeah uh bit has been around for a couple more just over a couple of years now um and originally as a network designed to distribute um AI training um and then pivoting or moving over to inference as the uh the kind of industry need and

narrative focus more more towards inference um and then going through a an evolution into or revolution but an evolution into uh subnets which essentially allowed what many people equate to l2s to be built on bit tenser allowing individual uh entities uh companies organizations individuals to build their own incentive mechanisms and uh value commodity uh marketplaces um the thing that bit tenser really does which is what really piqued my interest is it really uh sums up uh an in a true incentivized usage of blockchain distributed Ledger um and

distributed compute um so obviously this is often equated primarily to AI um but you can use distributed compute for anything the most efficient thing to do in terms of cost savings is to use GPU and the most uh in demand use for GPU of course at the moment is focused mainly towards AI awesome and before we get into subnets and and how these things are are are becoming prolific and ubiquitous in the ecosystem like what when we're thinking about compute in general what is the value ad of of doing distributed compute for inference or anything else versus using a a Microsoft Azure or or you know AWS etc for more of a

centralized computer what's the what's the value in doing it in a distributed way it's not just compute it's expertise it's both it's Distributing your workflow and incentivizing it so whereas you can of and there are many times when we've been building on bit tensor in the last few years where we would question would this be easier not to do on bit tensor uh and if the answer is yes then potentially you're not using the network to its Optimum uh design because you can uh employ inhouse expertise and uh fund in-house compute um but what you're doing by Distributing the compute is you're also Distributing The Operators um and those that are producing those commodity values and that's really where the uh power comes from awesome okay James you know feel

free to jump in here because mode has a lot of exciting uh AI agentic uh and defi Primitives that are being launched on mode this DFA AI uh space is is starting to take hold um so how are you guys starting to leverage this this new sector that's popping up and then ultimately how are you guys using uh bit tensor to to progress this initiative we buil mode and the L2 ecosystem to focus around like decentralized finance and AI use cases uh within decentralized finance and as we were building this out we kind of built out this stack we built out the application layer we built out the execution layer L2 and we we were

missing this data layer and the type of data we needed was uh training data price data um and loads of it for agents to essentially train on to get better at forecasting and trying to model volatility and as Mar said like it's not just the the compute which is important this isn't like a massively intensive compute task but it's more like the decentralized expertise which are very important um and approving yeah already the subnet has been live just over I think around 10 days and it's already improving already proven extremely useful to have this access to expertise like we could have just taken five data

scientist put them in a room and said model this out right but instead we've uh decided to launch this subnet Design This incentive mechanism so miners from all around the world experts from all around the world are comp competing um to produce uh the top models um as the output of the subnet and yes so far we've seen that to be like extremely powerful um and yeah it's been it's been really interesting we we registered the subnet yeah roughly 10 days ago um and the kind of process for that was fairly straightforward uh you speak to lots of different validators they give you weights your subnet gets emissions um

miners come um and essentially yeah work for those emissions um and yeah validat validate and also receive Admissions and yeah uh this is all about to change well not all about to change but some of this is about to change um I think it's actually a really exciting uh upgrade to the network and something that yeah I think a lot of people who are listening are probably also very interested in learning a bit more about I'm sure can tell every one yeah more about detail what that means yep um as you uh very rightly said um spinning up a subnet is uh not was

not so difficult but there was a slight barrier to entry uh in that you need some towel to do it uh it's a dynamic Marketplace so the the cost of registering a subnet was defined by the uh price uh by by the demand um and there is a limited number of subnets at the moment currently 64 um and so the cost or barrier to entry somewhere around anywhere between 1 and 3,000 to uh with toed around $350 $400 at the moment is a fair old barrier to entry which means obviously you need to get some funding or be well funded yourselves um it also means that you are uh dependent on emissions emissions are the percentage of the network uh output that flows to you

7,000 uh mirror tokenomics to bitcoin 7,200 uh pre-ir harving towel are emitted into the network and a percentage flows to each subnet based on in the current mechanism the voting uh a stake weighted vote by the top validators now for shepherding the network and getting the network in the right direction um this has been a really really uh powerful Catalyst uh for the last year or so however as the financial barrier to entry increases um and validator start to become funders essentially take on a VC kind of role within the ecosystem there is inherent bias that comes through uh that coupled

with the need to really understand value you have to meet teams uh analyze code bases stay on top of vulnerabilities because I we can discuss the uh the beauties of building an incentive mechanism and all of the uh idiosyncrasies that go with it later but there's a it's a lot of work to understand that value and it was understood that that system was flawed if we were to expand Beyond 64 subnets and there's no reason why the ecosystem cannot expand be Beyond 64 subnets the data exchange on bit tenser goes on off offchain each individual subnet and its operators uh operate between themselves offchain and produce value offchain the

chain is used for the incentive mechanism records uh recording the value of the par ipation shall we say so we're not limited by TPS and uh and metrics that you would normally measure a blockchain so we can expand to many many subnets but in order to do that we needed to have a natural uh value Dynamic value Marketplace so what deta does is it turns that power over to everybody pure democracy if you hold tow you are able to govern the flow of emissions um and that is done done through this this might get complicated for an evening podcast through something called Dynamic tow uh currently everyone

Stakes to a validator and you essentially give that validator the weight to act on your behalf uh the more weight a validator has the more power their vote has as to the subnets that they deem to be of value whereas whereas you can still do that that is still an option you can now take that Tower and say I'm not going to allow the validator to vote I am going to vote I am going to say that this subnet and this subnet and this subnet have value however it is not a a free ride it's a risk reward scenario you can equate it to purchasing tokens or acquiring shares in that subnet if you are right and that subnet increases in value then the value of your shares your tokens they're called

Alpha tokens in that subnet increase uh against the base currency Tower um in in a very crude way it's like investing in l2s back and forward between eth uh you can hold eth and eth will go up but you're going to make more from investing in the right l2s I'm not sure I like that analogy too much but it is the one that is probably the easiest for for people to understand um as a result the uh price market cap value of each subnet defines the flow of tow um so if you think of the overall market cap of tow being 100% the percentage that your subnet holds is where the emissions flow there are some slightly more intricate dynamic to it but I think that's a good starting place I I I do want to get into

synth as a concrete example that that we can explore as it relates to some of these incentive mechanisms um so you guys you know are launching syn as a subnet and and you know maybe in the course of this question we can also talk about what exactly synth will do um but in the context of these incentives you know synth is live does that mean synth is also going to have its own token or maybe better described it will have these a tokens that will represent its subnet relative market share in the context of all the subnets on bit tensor is that the right way to think about this do you want it or should I take a chance I can I can yeah I can take it so I think yes so yeah essentially uh yes

um every one of the 64 plus subnets will have its own Alpha token and synth will have so synth is subnet 50 um I actually don't know if it's going to be called synth token or 50 Alpha token called synth token it has a uh either a Greek Hebrew or Aramaic letter you don't know your toen I have to fire up Discord in the middle of this calling uh but it is there um you will be a subnet 50 are you ready you want to know what you're gonna be you oh that's a good point because you are there and it doesn't say what the name of it is I can tell you I'll find out you car find out find out live yeah

essentially yeah um the the best place as well for lot for anyone listening for all this information is definitely to Stats which uh mgan Team uh run um but essentially yeah what what this does and why I think it's like so interesting especially to everyone from yeah the overall crypto Community is like bit tens has previously being fairly a fairly small ecosystem right there's I'm sure there's hundreds of thousands if not millions of tow holders amongst exchanges and everything but uh the previous Dynamic has been like validators Miners and subnet owners right and as MOG said like

you have to as a subnet owner you have to go to a um to the validators you have to like tell them how great your subnet is why it's improving um the ecosystem why it's really like net positive for the space for yeah for all these all these things that we're trying to build right whereas now um it's going to be like the open market so as a subnet founder or a subnet owner you have to now tell everyone right every towel holder um not just validators um why your subnet has so much potential um as there will be this Alpha token there will be all these Alpha tokens which will which will launch um so yeah it's

going to be it's it's a huge upgrade and it's definitely going to increase speculation uh and give like a market value to subnets which previously uh the value was just emissions um obviously previously as James said you were very much marketing to the bit tenser Community uh because they were the ones awarding your value which was very good to keep things a sort of a quality control metric within what we were building and that the Community approved of it and it was the right direction what we do now is to appeal to the larger tow holder but again that alone will be shortsighted because we're not appearing appealing to current tow holders what we want to do is attract investors into the subnet uh into the

subnets but in order to do that the true metric of value will be the value of your product which um essentially should hold value to users outside of bit tensor successful subnets will be running companies that produce valuable Commodities and data that they monetize outside of bit tensor where their end user may not ever need to know that the product they using was generated or part of a product offering that comes from bit tensor and it will be the revenue and business models of the subnet businesses um that and how they they they organize flow back to the acrel of value within their Alpha token that will see those that succeed and don't and this is where things really change for a

while we pushed for Revenue as a metric uh alongside pushing for Value sometimes it's hard with scientific research based subnets but there will be an onus towards uh generating um quantifiable value and being able to translate that value back into the alpha token and if you are looking to maximize that you should definitely be looking beyond the scope of just bit tensor users which is of course where things going right and that ultimately will build these these inflow channels into the subnet ecosystem that that are kind of like tentacles in all of these other Industries outside that ultimately bring value and then we'd like to think that there's Network effects within the

subnet ecosystem on on bit tensor where people can kind of uh kind of spread we've had similar conversations to these similar to like these uh rollup Stacks not to not to belabor that analogy but you know how does uh one net new user into an L2 on the op stack Ben fit another chain on the op stack and and I imagine that similarly subnets that attract users from maybe the you know exterior AI or just general Community um ultimately impact and and beneficially provide some some net new user or net new unit of value to the rest of the subnets as they work together um I mean if you were to come in and realize that

you were renting compute not from runp pod but from a company that was built on Outsourcing compute through uh decentralized Network and understood that you might look around and think hold on a minute I could get my storage from there I could get my inference from there so yes absolutely there's a br effect to all the participants and subnets and and so specifically about synth like I you know this is particularly in the DFI space so you know I imagine that those who are looking to engage in crypto and potentially a user on boarding experience very easy way can leverage some of the tools that are going to be built on the synth subnet for or using for instance like an AI agent to go and buy tokens on on an our behalf or you

know pulling some data and and setting some sort of algorithmic strategy that will programmatically bu you know the the tokens that are newest launch or have the most volatility you know James maybe you could speak to the use cases that synth subnet is specifically tailored for absolutely so synth is uh concerned with like output of scent is um forecast like Monte Carlo simulation simulated paths um of price data for BTC right now and it will expand to many different assets um and actually many different Industries but it's all about um essentially trying to forecast potential paths into the future for

prices and yeah creating synthetic data for this and this is important to mainly to train agents on because in the same way that Tesla when they're training their autopilot they don't just train it on the roads they also train it in these simulated environments right so we can find all these different edge cases it's essentially What synthetic data and having loads of it enables you to do um so yeah s's main focus right now is talking to teams who are building Financial applications agents that need to uh forecast into the future and understand like volatility of markets so yeah we've seen I think 12 13 teams now

uh use like adopt API and start to build on that and yeah the the interesting thing for us is like the subnet has grown fairly quickly I think our our weight is now like 1.6% which is quite good um and the subnet is receiving emissions so the the data scientist machine learning Engineers everyone who's submitting this data is earning a mission but yeah soon this is going to change with DET and we need to think of like what the kind of holistic overview looks like um within yeah within growing since Alpha token and also within yeah growing like agents across crypto um and making sure that

data stays very useful for them okay that that makes a ton of sense I want to I want to Circle back and just reinfor force one of the questions that had early on which is you know the purpose of distributed compute and and why you know in syn's case let's just to continue to use that as an example here why is it particularly relevant for synth to utilize a decentralized intelligence Network versus using one of these centralized Alternatives Redstone is def's modular Oracle supporting over 100 clients on over 50 chains offering 1,000 assets Redstone is the leader for Bitcoin

staking and yield bearing assets they offer the most robust customizable price feeds available in any model join the Redstone exhibition [Music] today absolutely so I mean I've done like a few data science competitions in the past and you know some of them they give you like 5K 10K if you're lucky um and they usually start and they usually end and that's it right um and we try and so there's that as like a way like how things are

currently run right like how you currently get people together to compete on something um then you have uh yeah the other alternative which is um you hire five people 10 people 20 people 50 people 100 people you put them in a room and you try and get them to all solve like this hard problem maybe they do it individually maybe they do it by working with each other like that's what an investment Bank does right um or high frequency trading for that so the reason why uh Bens is like very powerful is you can you have if you you have a subnet that has purpose and has a good incentive mechanism um and a good scoring

mechanism um you can essentially bring people from all around the world to contribute their expertise that compute to try and solve really really hard problems um and what we're trying to do is currently a pretty hard problem like modeling volatility for so Bitcoin is very hard um and you need a massive variety of approaches and models so right now I'm poens there's slots for it's 256 minus mug 250 yep 256 on subnets with the exception of one of them so yeah 250 you can have 256 minus all contributing different models um and yeah your validator score those models the Top Model is that like yeah what we use or

what people can call um API to get the data from to power their applications so I think yeah it's very much the the kind of like if you need something that needs a variety of approaches and expertise then poens can also be a very good fit for that okay okay so I I think I'm starting starting to wrap my head around this so we've got and so to draw a comparison here like for instance say you know open AI or something in one of these centralized models that is entirely built and and in open AI Case Closed source and so you know they have their Engineers contributing to this model and then they put out you know when they

feel like they have they have a new version that is ready to release now open AI does that Claude does that deep seek does that deep seek is open source but essentially same thing they put out their own model they're really I haven't seen any production level tools that allow these models to to coordinate with one another and to collaborate with one another James what I'm hearing you say is that you know there's 256 miners or models is it the case that you know once we have these models on a subnet or on the bit tensor Network they're able to coordinate they're able to learn from one another they're able to collaborate

and share their level of intelligence with one another is that a more general idea of the way to think about these models on and subnets on bit tensor uh yes and no what you are actually alluding to was one of the original uh ways that the one of the first sub was looked to be designed this kind of mixture of experts model where any any minor could run any model and uh would be rewarded for diversity of responses um if you're talking about inference you have to bear in mind that a subnet running inference is incentivizing speed latency accuracy of response now humor consensus is very data agnostic uh you can uh assign value and tiered value to to any kind of uh

data not just a binary yes or no but actually awaited scoring however you still need to know what you consider to be the right response if I was to ask uh uh deep seek GPT and Claude to write me a haiku uh on on cats uh who would be the objective judge as to whose was better who should be rewarded correctly which one I mean you if you were if you were making a haiku subnet then your incentive would be on the formation of the highq etc etc so you would have other incentive mechanisms but in terms of uh getting together to better each other yes this is possible but in order to better each other you have to be producing a data set uh to be including some uh potentially human feedback in

the loop um and all of this is possible with multiple subnets so you have a subnets such as uh shoots which can deploy any AI model or it's basically serverless AI compute it's currently uh serving uh deep seek um you also have subnet 19 is serving deep seek uh to open routa so you could use both of those also subnet four targon so there's two inference and one serverless AI compute that are all serving uh deep seek so that that gets your responses uh then you have fine-tuning subnets there are many that are doing specific tasks uh based on what the the subnet owners are trying to achieve so dippy are training uh

character models on very specific uh data set outputed from their commercial app they're also doing a text of speech uh then you have some that are just trying to achieve uh state-of-the-art uh model uh benchmarks with one specific data sets and then you have subnets like gradients which allow you to arrive with any base model and any data set and say right train those for six hours 12 hours and give me something back so you can already start you start to understand what you're doing there you then have subnets like Dojo I'm not giving you numbers here I'm only giving you names but uh um they provide human feedback so you could then take any of those results and say hey we're going to get these over to Dojo to include this in our workflow so what you start to see is the

more you understand about especially to do with the AI training pre pre-training fine tuning inference uh human feedback those are available and there are some of the best quality subnets we we have at the moment in terms of their their infrastructure their build the teams that are behind them some of the the ones with the most experience in the ecosystem then if you want to add in elements like Financial predictions uh let's say let's take the outputs from synth let's mix them with another Financial prediction subnet or let's take those outputs from synth and put them into one of the other prediction subnets uh as as a data set and then let's use that output to train another model on you see where we're going the answer is yes uh but it's going to take

some imagination and understanding of all the different powers of the the different subjects essentially yeah and and more more understanding of the specialized nature of each of these right because ultimately you know at a high level all of this looks like a decentralized intelligence Network that's how I understand bit tensor at a at a holistic level but underneath the hood when we start looking at individual subnets these each perform a very specific specialized task and and those tasks kind of line up in the intelligence supply chain if you will that you know you went from pre-training to fine-tuning to inference to reinforcement such as this the intelligence supply chain each of these

subnets uh produce and uh contribute a particular task in that supply chain to ultimately uh to ultimately create the most uh accurate and sophisticated intelligence uh that is outputed to the end user so that that really does help to to add color and kind of fill in the gaps to these models work together it's not just that we have one endtoend model that is learning from another endtoend model these these subnets actually contribute one piece of the model and then they can collaborate with the others to to each make the next piece better so you may not want all of the pieces of the puzzle you may come in as

an established provider of intelligence and say well we just want that or we want to use these two bricks with bricks within our data flow yeah yeah and and I think another parallel to to to draw here is um the idea of of a monolithic uh intelligence model such would be an uh an open AI or a Claude that is all developed end to endend by the same produce production company versus a more modular architecture such as bit tensor that has each of these subnets contributing modular components to the intelligence workflow I mean the industry is evolving at such a fast pace I think even AI open AI can't answer the question as to whether we'll have a monolithic model or specialized models there are contrasting views within the

industry on this um and what we thought six months ago may not be what we think today and will certainly not be what we think in six months time and what bit tenser does have is the adaptability to move at a far space and provide the one thing that we were missing before which was Computing compute power being able to access that compute you know you spoke at the beginning of deep seek uh and what they've achieved of course there's some ambiguity as to whether they what what training uh resources they actually had and they were allowed to have um and whether without getting into the political side of this whether this should be a barrier to being able to produce models surely the barrier to producing models should be uh

imagination quality of data sets um and pushing forward the frontiers of this ever fast growing industry yeah and and so you mentioned that there W just to to pause on deep seek just for a moment you mentioned that that there was a couple of subnets here that are helping or or deep seek is using these subnets is is that is that the case how and you kind of mentioned that deep seek is is on open routa and the uh I think uh subnet 19 has surpassed azur uh in providing inference for deep seek I think they've provided over a billion tokens in the last week or so in the last couple of weeks uh I'd have to pull out the stats but I 100% sure they surpassed a zero as a number one provider uh of of deep seek

conference through open routa wow so so this does highlight the idea that an a a distributed system with its its more modular parts has been able to supersede you know some of these other inference contributors in in the context of of deep seek and the the providers that they're taking into account absolutely you can you can consider mining on subnet 19 as a full-time job and it's not one person sitting there spinning up 4090 or whatever they need and providing inference uh to that that people are running load balanc highly technically uh uh very efficient setups and it's incredibly competitive and

there will be times when there'll be dominant by a minor or group of miners who will start to eat up more spaces uh and the others will have to catch up but by having this competition uh through limited space uh and and and essentially a fight for resources it cre creates this darwinian effect of a constant Evolution um and that's allowed 19 and other inference subnets that to to Really provide um in terms of speed latency and reliability uh the up as good as if not better than the centralized providers who are relying on their own choices in infrastructure whereas you've got competing infrastructure providers saying I can do this quicker better faster more reliable I I love it I love it because

ear in this podcast as we were talking through the incentive mechanisms I was thinking to myself okay well you know if if there is a relative market share component and those who earn the most end up dictating the incentives for everyone else it felt like a rich get richer type of thing as you gain more market share you dictate more incentives you can allocate those to yourself and the flywheel continues until one subnet dominates the entire ecosystem but I think you point and through this course of conversation I think I I think we hit on very important point which is the biodiversity or client D better subnet diversity of the ecosystem and it's not such that one subnet if all subnets were the same one could Dominate and one

could get richer if they had a better go to market strategy etc etc but because of the diversity here that one subnet that ends up getting really popular is actually potentially dependent on other subnets in the ecosystem and wouldn't so much call it altruistically but but it would it be in its best interest to reallocate incentives to other subnets other than itself in order to keep the the diversity of this ecosystem alive and flourishing yeah I mean inference has always been the popular narrative for quite a long time and the inference subnets have taken up the largest uh majority share of emissions for a while but this has just been I would say because it was easier to bring those to uh Market uh parallel performance with

mainstream providers whereas of the other things that are being built on bit tens have really taken a slightly more creative approach of doing something new and it's a new breed of subnets such as shoots uh such as celium compute uh that are really challenging existing models of businesses and doing something that can potentially only be done on bit tensor um and these are the ones that that are set to rise up um I feel like detto will be this Catalyst for this um because people have I wouldn't say become stagnant in their appraisal of emissions in the last month or so but with so much focus on this new mechanism we're looking at it we're studying it we're preparing for we're living in the

this future moment that will be live next week that a lot of us have just sort of said well you know let's leave the the the past way of deciding emissions behind so I feel like there will be definitely a rising up of new uh champions of emission um very very very soon with novel business business models we've got one right here go ahead J yeah I think the other thing that like some people some people discount with bit tensor as like how how powerful the incentives are um and yeah I think I don't know in dollar terms how much in the last 12 months has gone to powering subnets um and miners holidays but is it

around like close to a billion maybe like 500 million between 500 million and a billion so there's been just under a billion going into essentially yeah this system and it's yeah been incredible how like that I mean the output is obviously incredible because the incentives are so strong and I think yeah daow is going to accelerate that further um and yeah po level the playing field uh for new people coming into the ecosystem which I think is super exciting yeah

um and MOG you mentioned that the incentive mechanism here loose follows Bitcoin in the sense that we have a diminishing inflation rate I'm not as familiar with the the tokenomics of to the tokenomics of tow uh are a test swansong a testament to uh uh Bitcoin so they directly mirror with a first Haring at 10.5 million emission which at this point will be uh late this year I think uh it's a slightly elongating date because uh delayed date because towel can be recycled through registration um and then various other aspects but we're talking I think in 2022 it was in August and now we're

sometime in November so it will be either mid to late November or early December this year will be the first Haring got it well that's certainly a date to look out for um cool and I you know most of our audience is is it's crypto native for the most part right Builders investors so we have recently had quite a few AI conversations but I think still people are kind of opening their mind to the idea of AI and you know coming out of the trenches if you will and and and kind of expanding their knowledge base and so you mentioned that inference kind of kind of dominates and overpowers like a lot of this ecosystem could you just could you just break down inference for

the lay person and and maybe speak to why inference is so important for AI development uh yeah well inference being prompt response uh G you know I type in some text and model gives me a response I request an image image and I get an image uh or fact within bid tensor a 3D model uh but it's essentially uh ask question get response uh it's the most understandable and us and it's a GPT effect you know GPT is what brought this uh to the public eye it's the easiest thing to build applications on it's the thing that we use most of all in our day-to-day lives especially as developers content creators uh any of us sitting in front of a computer our lives have been made infinitely Easier by by

being able to have a friend who'll write the long form version so we only have to write the short form or correct our code or answer a question that we used to Google before we don't Google it anymore we we look it up online and so there's huge value in this uh this is where the initial gold rush towards building things were what has been in terms of applications um but this is a not a shortsighted but it's a a view on the now it's building on the now of AI and of course those within the industry know where this is going uh with the drive towards uh AGI uh ever more powerful models ways of uh getting agents to uh perform tasks on our behalf and then and and you know and as I say said to you in the beginning in private I'm not an AIML

engineer uh everything I've learned about AI I've learned through my involvement with in bit tenser my specialism has always been products and the products that can come out of this and I can still imagine products that uh we will see in the next couple of years that I always dreamed about as a child and these will be powered by by by ever increasing powerful models and compute um and so it's it's obvious why inference was the easy one out there but the more you understand the more you understand necessity Behind these other tools um you know it's like it's like comparing uh email as the antithesis of the internet so much more is done with the internet but our parents understand email as what it was that brought the

internet internet to their to their prence it was post replaced by email the internet does a lot more powerful things it's now we don't say it's only for going shopping do we it again does a lot more powerful things but we understand the the need for uh public um absolutely cool guys I mean this has been an incredible podcast I really appreciate it and uh I certainly learned a lot not just about bit tensor and the ecosystem there but also the way in which these these components interact with one another to create an intelligence supply chain and the incentives that play in that supply chain so uh guys congrats on the partnership with synth the 50th subnet deploying on to bit tenser and uh the

entire bit tenser Community transitioning from maybe it's early stage bootstrapping mode into a more mature Market uh a more mature mature market and diverse uh ecosystem thanks so much appreciate it Epic