Episode 32 - Macrocosmos.ai with Steffen Cruz and Will Squires

What's up, Dcentralistus and welcome to the BitTensor Guru. This is the auditory intersection of blockchain incentive and artificial intelligence. I'm your host, Keith. It is Wednesday, May 15, 2024 and this is the 32nd installment of this podcasting series focusing on the decentralized open artificial intelligence project taking over the world one subnet at a time called BitTensor. In this one, we interview Stefan Cruz and Will Squires from macrocosmos.ai. The type of talent that is picking up subnet slots at this point in the BitTensor ecosystem is ridiculous. The chops on these guys is off the charts. If you go take a look at their bio, you'll see that these are the type of fellas that have the heads on their shoulders that allow them to pretty much do anything and they kind of have at this point and now they're stretching their intellectual legs into the BitTensor ecosystem and are about to cause a ruckus with what they're creating and delivering. They're not only responsible for managing subnet 1, 9 and 13. So that's text prompting foundational models and 13 is the dataverse. They're also launching a protein folding subnet which is basically a working proof of concept for like any type of actual academic institution or applications using BitTensor's decentralized network to solve real world problems and they're showcasing what this network can do with that new subnet which isn't even registered yet but is up and running on TestNet for anyone that's interested. Check the show notes. I'll put a link to the GitHub in there. This network's growing up and these type of guys are the proof. So I think you'll enjoy this hour or so with Stefan and Will and make sure you check out macrocosmos.ai, one of the most exciting teams to be building on the BitTensor network at this time. Thanks. Bye. Welcome to the BitTensor guru. I'm here with Will and Stefan from macrocosmos. We were talking before the call and these guys have apparently been jumping out of planes together for a number of years. They are now parachuting into the ecosystem jumping from the open tensor plane and landing in the subnet zone. They're launching, well they have multiple subnets that they're launching including some very interesting ones that I just read about this morning. Welcome to the show guys. Thank you. I'd like to congratulate you on the excellent metaphor as well. Yeah. Yes. Thank you. I was working on that before the call. Human generated also. No chat GPT, no core cell. That's straight from the source. All right. macrocosmos. Now, there's been a couple of successful forays from former open tensor members into the subnet ownership role. We've got subnet 21 file tau. We've got subnet four which is run by Kero and both of those subnets are doing very well performing in the ecosystem. You guys are now also doing something similar except you've teamed up here and can you talk a little bit about how and why you formed macrocosmos and then we can kind of get into the subnets that you guys are running. Yeah. So I guess the start of this story is definitely, as mentioned, only each other for 15 years terrifyingly. We've always had this rolling six month block of like, is now the right time sort of business together? And you know, Stefan's been involved in bit tensor for over a year now and I'll let him speak to his own point. But I got involved through doing some work with the foundation, with const and share of on sort of, I guess, operational strategy and like helping to work a bit on the structure of the foundation, some other things back in September and October. And then with the sort of outset of revolution which I sort of did a very small piece to help on and I'll give most of the credit to the rest of the team. I think that success sort of bloomed and we saw the sort of proliferation of subnets and we decided now is a really good time too. I guess formed something together that we thought could meaningfully help drive bit tensor forward in the broader sense. And I've spent the last 10 years building AI companies and tech companies in I guess don't work too space, go through them FTSE 100s and within sort of series A series B startups. I actually started my career building nuclear power plants as a civil engineer which is a bit of a site topic. But Stefan and I really thought it was time that we could do something exciting here and we saw this sort of quite amazing community coming together around AI, around compute, around incentives and thought we could add a certain flavor in the space that around research around doing some of the deep work on the network and that's what we're really excited to be doing. Stefan, I'll pass over to you to give I guess your sort of brief introduction on that. Yeah, sure. Well, as we already said, I joined the foundation early last year. I was one of the developers on 7.0.1 and it's a original form. Little while into my tenure at the Open TensorFlow Foundation, I became a CTR. I was appointed there for a little while. But what became apparent towards the end of the year is that the foundation we really needed to help our game in how we communicated what we're doing. There's a very high sort of technological barrier to entry we found. So a big part of what I was working on with the foundation towards the end of last year is just to develop a dark, really improving the experience for the developers, the miners, the validators, the builders in the ecosystem. Of course, things really started to lift off with the revolution launch. So we really needed to catch up with all the interest and the eyes that were on the screen. In addition to that, the foundation under my guidance also we developed a lot of stuff around the 7.0 template, which has hopefully helped a lot of builders great our own subnets. But what became apparent at the end of the year as well is that the foundation doesn't need to sort of distance itself a little bit from the game playing of subnecoration and maintenance. And so that's when I sat down with the founders of the foundation and also my old friend Will. And we talked about how we can participate in bringing bit tensor to the next level as subnets creators. And also our part of this was to really hopefully raise the bar to make subnets creation more of a craft, something that we sort of flesh out the rules, to do the don'ts, the philosophies around how to build good incentive mechanisms. And that's really what's proud of us here today. We're here to explore lots of different subnets, subnets ideas. We've got a backlog as long as you're on of different subnets that we want to build. We're currently trying to find the people to work with us to build them, but we are short of nothing if not ideas. So yeah. What's it like coming from the open tensor foundation and can you can you talk a little bit about like the working environment there because I think for most people that are involved in this project or maybe invested in this project, they just kind of know it as this sort of entity that's maybe responsible for the bit tensor code base. But what's it like actually working for that organization? Yeah. Really, really smart people working really hard for too much. It's probably how I would describe it. It was really humbling and really fun to work with them. Their teams are absolutely at us. Really, really good. I learned a bunch being there. But really what they're doing behind the scenes constant is of course like there's the next tenth version coming down the pipeline, which is hopefully going to bring new features to the community. But there's also just so many, the scope of what the foundation takes under its wing in terms of its responsibility as sort of the de facto custodian of the protocol. They're really making sure the lights stay on at this point. And so their work behind the scenes is often not that glamorous and certainly not that celebrated, but it's really, really important. I've got massive respect for everyone working there. Like I said, they're really, they're really good at what they can do. I would actually feel that the lights staying on is the wrong analogy. They're more of trying to make sure that you can run like electric powered railways after you've just invented the light bulb. Like I think the main thing I found is they are like three years ahead of most of the people in thinking about how does this thing break? How does it evolve? You know, I had conversations with const about like, you know, there's obviously a lot of conversation about weight setting and things. This was something you was worried about before it was really. So I think none of the concerns or fears people have about the network. You know, none of these are a surprise to those guys, which to me is comforting. And they're very much dedicated to doing a lot of the really deep technical work to make this thing successful long term. And you know, that's sort of evidenced by, you know, almost where we are here today. Like it pops in a less altruistic system. You know, a bunch of the subnets that were sort of seeded within the foundation, but still be within the foundation, you know, padding their own books. What is actually the direct opposite? Like you see the foundation creating and then shoving it very almost aggressively out into the wild to fem for itself. So I admire that about that, their behaviour. And you know, they're focusing on the things that come next, not the today. Yeah. And I think the last thing I'd add to that as well, if you're jumping back to a point you make Keith is the fact that some of the best subnets right now, not about ourselves, but about Faltao, about subnet forest hargon. Like that shows kind of the pedigree of people that are working in the foundation that they come out and they put on a clean. and they show us all how to build a down-to-net well. I think that's the experience. - Yeah, Will, you mentioned the subnet like aggressively pushing things out into the wild. So you guys are kind of the thesis of that here by what you guys are doing. So, and Stefan, you mentioned too that there was this, you wanted to kind of avoid, maybe you didn't say conflict of interest, but you said that open-tensor kind of wanted to absolve itself from the subnet management role. Can you talk about why that would be that that open-tensor wouldn't want to have subnets is that maybe a kind, like I said, a conflict of interest or why would they kind of want to distance themselves from that role? - Well, I'll be blunt here. I mean, they run a very substantial validator and there's always a little bit of a, you know, once you have a very large influential validator, which is a really important part of the allocation of emissions. It becomes very tricky to justify and do so in a transparent and fair way how you support your early initiatives, let's say. So that was the concern. And like Will said, we were worried about this within the foundation long before revolution even released. We had a subnet, we always planned to continue building subnets, but it became very apparent very quickly. It's like, we as the foundation are trying to put ourselves out of business long term. That's the goal. We want to become completely obsolete eventually. And so actually sort of self-financing through our disproportionate influence on the network is not the wrong deal. - And I guess you can see the next stage of that right now, like with the announcements about validator take rate being reduced. Now the foundation is funded, first and foremost, you know, through its validator take rate. And, you know, again, a less altruistic group, would perhaps seek to cling onto that as opposed to pushing the next wave of the network. And you know there, there's certainly not a group that's afraid to create pain for themselves if it's for the good of the community. I think it's one thing that really stands out about bit-tensor. Above other projects in that, you know, it was a firm, there was no pre-mind. They continuously almost stabbed themselves in the leg if it's for the good of the community to push the project forward. And I think that altruism is why certainly one of the big reasons I'm here today, because I see it as a good place that by people who care about the project in a broader sense, not necessarily personal image. - Yeah, that's something I would echo too in my two years in the ecosystem. It's the reason why I'm in the position that I am today is because the, you know, the founders and the team that has built bit-tensor has aggressively tried to essentially give it all away. And it seems like the more you try to give it all away, the more it comes back. And I've tried to emulate that in the way that I run my validator and it seems to be that, you know, that's just the best way to do it. The universe sort of rewards you for, you know, the equanimity and the charity that you distribute to those that are kind of supporting your initiative and bit-tensor is a great example of that. Can you guys talk about now that you're sort of exiting the open-tensor role moving into macrocosmos? Like, how does that free your mind creatively? And you mentioned you got a backlog of subnets. Like, what's the difference in what you're focusing on day-to-day from what you were doing, let's say three months ago to what you're moving into now with macrocosmos? - Yeah, sure. I can take a stab at that. So, I mean, our first pass is Semnet 1. It's no secret that we try some techniques that in retrospect were much more difficult to, let's say, sustain in a decentralized, entrustless network, allowing LLAMs to be the judges that were the basis of the reward mechanism was a really neat idea, but in practice, it was just an invitation for people to find creative ways to jailbreak into the reward stack. We sort of self-defined our own rubric internally where we wanted to lean very heavily into sort of the research possibilities of like, how do we have a completely autonomous self-sustaining AI system that rewards itself and sustains itself? A very lofty dream. And I'm very proud of what we built. I think it was a really interesting piece of work, but nonetheless, I think that mandate made it really hard. Again, sort of a self-prescribed mandate, not necessarily something that came from the founders, but that commitment to trying to do everything in the most pure, it in a way possible meant that it was quite hard for us to kind of cut through the weeds and really get a subnet that clicked and to get an incentive mechanism that was really driving all of that effort in the direction that a good product of subnet needs to. So I'd say comparing them to now, part of it is just that we're equipped with more experience. We have lived that experience now. We're seeing what happens when you try that way. We've learned from it, but also sort of the difference being a completely stepper entity. Look, frankly, we can push the boundaries in different ways now that we couldn't do before because the foundation is something that we want to be really careful with. We don't necessarily want to do things that are, for better or worse, they become synonymous with the intent and the actions of the foundation, whereas we're operating more autonomously and we can try those things out on our own back. And I think that's better for both sides. - Yeah, you mentioned subnet one. So subnet one, I'm sure has had more development hours behind it probably in multiples compared to any other subnet on BitTensor. It's the one that's been running the longest. It essentially was the network for the majority of still the majority of BitTensor's life, maybe a year and a half or two years almost. The subnet one was kind of BitTensor. What is the state of subnet one today? Because I'll tell you, my last sort of big download on subnet one, Stefan was that presentation that you did towards the end of last year in one of the Thursday sessions, the TGIF, TES sessions that open Tensor runs. And you guys at that point were pushing subnet one to sort of be the operating system for the Tau landscape and that you could use subnet one to access the other subnets. How has that sort of utility function developed over the last six months or so? And what is the state of subnet one today? Yeah. So the massive ambitious build, what we ultimately want to do is the well-position for when we have so many subnets that no one knows how to access them or do what to do. In that eventuality, exposing your subnet to the outside world becomes the bottleneck for the subnet owner and bit tensor. Its ability to showcase its utilities also bottleneck by how effectively you can sort of surface all of that value. So the design of subnet one, its LLMs, are, as you said, absolutely correctly, they're effectively the operating system. They're the reasoning engine that takes human queries in natural language, which we believe is the natural language that we've got to want to use. You know, it's much more convenient to be able to communicate in that way. And they're going to translate that and understand what the underlying task is. And then they're going to dispatch that workload just to the respective experts subnet, let's say. I sometimes think of this as sort of down-economes, thinking fast and slow. You've got system one, system two. Think of subnet one as sort of the system one of bit tensor. It's the intuitive fast brain that can handle decisions on the fly. 80% of pedestrian use cases, it's going to be a smashing job. And the other 20% that are more detailed, more involved, they require a slower, more thoughtful measure response. That's when we would basically send those requests to output a subnet. So what does that require from subnet one? Well, the miners need to be native speakers of natural language, and they also need to be able to know how to use external APIs. So what we've been building since January is really, we've been slowly refining and improving an agent-based network, where the miners have to prove that they are hosting sufficiently capable models. While at the same time, they're constantly being measured on how well they can use APIs and tools. And of course, this is a precursor to fully connecting subnet one for the rest of the answer, which we're already conversations with various teams on how we're going to integrate that. So where we're at right now is just a lot of deep foundational work thinking about how can we measure agent performance in the most general way possible. You don't necessarily want into subnet communication to take part during validation. That becomes a little bit involved. So what we have is sort of an internal validation process. So something that is a suitable dry run or production, where you're actually reaching out into the subnet. So we hands down and have the most complicated validation mechanism, which is like multitask. It has all of these things going on. And that's not really necessary for us to go into. The main thing is we're constantly adding additional APIs that the miners have to be able to use in order to measure their usefulness. And eventually, or actually, soon, when we actually release our product on top of this, people are going to learn that the miners are now very capable, they're state of the art. They're very useful. And this is only going to get more and more capable as we integrate with other subnet. Yeah, for me, this is-- Go ahead, well. No, you go ahead. I was just going to say this is some fascinating stuff because what I'm going to say is think about this, I'm not saying we're trying to build the self-aware artificial intelligence here or AGI or whatever, but what you are trying to do here is incentivize miners to be able to understand that they sort of have an appendage and that appendage is the other subnet. You know, and so like for anyone listening to this, what we're talking about is using the pool of miners using the pool of decentralized compute on subnet one so that if you go into, you know, a prompt and you try to communicate with the bit tensor network through one of these miners, that these miners will, you know, intuitively understand whether you'll be able to answer that question using kind of their own resources or whether that question exceeds the ability of whatever they've got under the hood and that answer would best be served by being sourced from one of the other today, 31 subnets outside of subnet one, but eventually, you know, who knows how many thousands of subnets at some point. And I think there's two other points to that, I pick up on key, maybe even three. One of those is like, subnet one needs to be something that celebrates the fact it's bit tensor intelligence. So like the fact that, you know, a lot of the other brilliant subnets out there are doing things that are like brilliant within their subnet, but don't necessarily like champion the decentralization or the fact that it's part of this broader intelligence. So a big part of that is as Stefan mentioned, having a subnet that has the consciousness that other subnets exist and can, you know, suggest I've got an answer but go somewhere else. But one of the other things you want to do and you'll see this in our product is celebrate the fact there's a thousand miners here. So we're looking deeply at some of the cutting interest search around response ensemble, but also looking at, you know, infinite canvases and can you explore what happens when diverse miners respond to a response. Maybe I want some of that wisdom of the crowd that was sort of some of the original thesis behind bit tensor, but, you know, the other thing we're looking at there is how we measure the intelligence of that network in an open source way and in a sort of interregible and digestible way. We're in the process of looking at pushing the validates stack to alarm a 70 billion. We believe that like certainly for subnet one there needs to be a big open source component of it, which is why we're celebrating that. We're doing as much as we can to reduce the alarm a 70 billion models through quantization to make them as efficient as possible for validators because, you know, there have been questions about how does bit tensor manage itself to be economically efficient. We're doing some deep work on that. We have two AI PhDs working on that right now, so credit to those guys doing some seriously deep work we think is cutting edge. But those are the sort of things that are quite interesting to us on one. Like how does it be like a truly bit tensor form of intelligence? How does it be state of the art in the open source? And I guess the final bit would be how does it have some way of measuring the intelligence that's coming off it in? In I guess a way that others can look at and respect. And so those are the sort of three pushes for us on one today. And in fact there's one final element because I couldn't remember this one when I was mid-flight. Subnet one I think as you articulate it has had probably the most development time of anything on it. One of the things we do want to look at is what's the next like genus of incentive functions that you need to succeed in bit tensor. So there's been a lot of work around working out how do we get quality intelligence from miners? We're actually interested in as bit tensor drives to becoming more performant applications. How do you turn up the heat? So how do you push not just for quality but quantity? How do you incentivize miners who can push 5 million tokens? 10 million tokens a second? Because if you want to run seriously large scale applications on this sort of thing, you need quantity as well as quantity. And you need to start looking at serious like performance computing problems which are the things that have make bit tensor scale to the next level. So that's a lot of our interest on one in the next few months and you're going to see some exciting developments that we do. There's so many little alleys I want to take us down here. But the one that I kind of want to latch onto is I was listening to a panel with a bunch of miners. Well, I guess they weren't really they were mining on bit tensor but they're actually just like well-known personas in the AI space at ETH Denver maybe a few months ago. And one of their criticisms of bit tensor was that we were using stuff and like you mentioned models to essentially do the grading of output from miners and then miners were fine-tuning based on those reward models and overfitting and you would see this sort of like reduction in utility or intelligence in the contribution from the mining pool. Now the subnet one has developed far enough to the point where you're talking about using the and I forgot how you described it. I guess I'll call it the aggregate but you're using responses from across the mining pool to then figure out essentially how to take that response coming from dozens or however many miners and you know instead of picking just the best one that is overfit to the reward model you're trying to combine that response in order to create like a consensus intelligence or like an aggregate intelligence from the mining pool. How can you like what is the process of doing that and could you talk a little bit about how you've gotten to that point today? Yeah sure I guess just one point before I jump into that. I think there's always a intentional decision that a subnet owner or builder of a subnet has to decide upon which is the notion of sort of are they a big government or a small government? How much do they want to actively police the participants in their network? I think that the LLM as a judge is still really valuable in the ether. It's just not that adversarily resistant. Now can it work? Yes if you're prepared to like slap everyone on the wrists that tries to jailbreak your rewards diet you can just blacklist everyone into submission. It's not our jam but it's a way of doing it and it's totally blamed by the rules still as far as I can right. You can just say miners like oh we don't accept adversarial people here. You end up in a bit of a cat or mouse race with everyone on the subnet because of course miners will subvert the rules as they can and will and you have to constantly follow upon that. So we prefer to just be kind of puritan about this and we're small government. So we just try and let our incentive mechanism speak for itself. We write it and then we let it go into the wild and then we're all we find or at least what we design for is fine tuning is aligned with what we want the network to get better at. They're not just getting more efficient at subverting the rules. So that's just a sort of a historical note on the previous version of subnet one. Would I do it again? Yes but I'd have to be a hard asset out of it next time. Everyone's being blacklisted next time. Regarding your point about Ensembling, yeah it's a really interesting piece. It's considered to be pretty state of the art in sort of the general research around AI inference right now. I'd say there's two broad scenarios that you can look at. There's there's a scenario where you kind of want all of your answers to be divergent. So actually maybe you're exploring the concept. What you actually want is you want 10 totally different opinions that you can explore and that's actually a very fertile ground for your activity of creative design, ideation, whatever that might be. And then I sort of call that sort of you want a divergent ensemble and then the flip side of that of course is you want something convergent. You want to be able to somehow aggregate and bring everything into one and have them sort of gravitate towards some central ground truth. In that case we've already done a fair amount of work on the convergence scenario and what that really involves is let's say I ask a question which is mathematical or ask a question which is based on a specific fact. Well what Ensembling would do in essence in their case is it would look at all of those responses, analyze the semantics or it would extract various facts from them. They would compare the facts across all of them and then what it would have is two things that are both valuable. It would have again of course the votes of all of your responses so you'd have some majority vote mechanism that tells you oh look seven out of ten and seven happened on this day. Great good to know but also you have some measure of uncertainty because you also have some sort of spread of opinion and of course you can use that as a proxy for like how sure are we about this answer. So what you get for free is some kind of like soft, basely improbableistic guarantees around how certain you are about your answer. So it's really really useful. In fact a lot of state of the art models are doing these days is to give certain responses they are generating many outputs and then they're aggregating them under the scenes and this is well known to give you substantially improved results because typically when language models hallucinate they hallucinate in somewhat stochastic of random ways rather than predictable reproducible ways. So if you just get it to inference enough times it will sort of cluster around its ground truth and then you'll have a diffuse cloud of sort of wrongness. So now you can imagine just taking that distribution just filling it down to a fine that should approximate the right answer. So we're trying to do that research on some that one in the background right now. We're also trying to weave it into the incentive mechanism itself but at least it's going to be present in our product once we go live and again the main takeaways there are you get some enhanced certainty on things like factual objective results when it's convergent. And some measure of uncertainty about that. - Yeah, this is something that, you know, for me as a fan of just open source AI in general, it's exciting to hear you guys working on this because when I think about, you know, the centralized competitors, your Chatchee PT's or whatever, you know, they're spending so much of their resources on essentially filtering out what they originally generate as a response. And if you go back, if anyone who's had experience using these tools when they were initially launched, they operate very differently than they did in the beginning. In my opinion, due to some very heavy, you know, RLHF response, you know, type training where they've essentially tried to make the answer most palatable to the most number of people. And this has led to essentially a dumbing down and a whitewashing of whatever awesome intelligence exists underneath the hood. You can't hardly even get to it anymore. And I love the fact that, you know, on BitTensor and on Subnet One, that you guys are spending so much time on creating this ensemble intelligence and this convergent intelligence, utilizing all of the responses from these diverse sources rather than spending resources on essentially obfuscating the, you know, the, like the real, the sauce of the intelligence, the ingredients that people want when they're interacting with these things. Like nobody wants that clean cut version. Or maybe if they do, there'll always be the centralized source to get it. But on BitTensor, it seems like, and especially what you guys are talking about, you know, this is the exact kind of like development that I would want to see in open sources, pushing the boundaries on creating these, like, amalgamations of intelligence from diverse sources. So awesome. That's, yeah, we could do a whole podcast on that probably. But I think one of the things that was really important when, when staff and I chose to find microcosm was like, we're both researchers at art. Like staff and spend 10 years doing the particle physics. He's got a PhD. He's the world export on a very obscure ice adover strontium. I've read the book in boring, don't, but like, we're both researchers. I want to research fellowship at UCL. And we, one of the commitments we made was we wanted to do some of the deep work that would make bit tensor competitive and beat centralized solutions in the right spaces. So whether that's like pulling together a group of people that can actually crack the ZKML challenge, whether it's taking on stuff like on sampling, whether it's publishing and writing papers, like a lot of our mandate was to be in many ways an AI company first or decentralized AI, and a sort of bit tensor subnet builder second. And you know, that means we, in contrast to some of the teams that have been very quick to push product, push other work there. We've been doing a lot of deep work on incentive mechanisms, on creating really high quality commodities like that others can use to build. But also looking at what are those big technical hurting, hurt, technical, hurting planks have lost the word. Hurtles? Hurtles, that'll do. (laughing) They're like, we need to be small, for the tends to be successful, but also to win in its field. And you know, we're committed to doing a lot of that work and putting some of our brilliant team of resources against that, and it's exciting for us to sort of get started on some of those initiatives. And I guess the purist in me also wants to do this with a smaller government as possible, the level of the subnet owner. And I can we write the incentive mechanisms that just get what you want instead of intervening and interrupting it? - And why do you guys choose to spend your time on pushing the boundaries here on open source artificial intelligence? Because Stefan, whether it's you with your experience in exotic isotopes or will, you've got, you know, you could go build a nuclear power plant. You know, you mean like, these are big, like you guys can basically, you're smart enough to pick what you want to work on and make a decent enough living at it that you essentially get to decide. So why are, you know, we sitting here talking and you guys talking about launching a couple of subnets, you're forming macrocosmos. Why did you choose to work on this problem? - I think you sort of covered it right there, actually Keith, like I built the reactor. Stefan makes the explosions go off inside. That's kind of the way we work. Look, I think artificial intelligence is the defining technology, probably of our light-times. Like I certainly believe that. And I personally believe, like I actually look at this as an entrepreneur, first and foremost, like, you look at the way the businesses built on chat, GPT and OpenAI, they just collapsed when they released their AI studio type thing. Now, as an entrepreneur that pisses me off, the fact that a centralized entity can just steal everyone's cheese. So I think open-source decentralized systems are a way for creators, for people who care about doing deep work to be incentivized or rewarded in a way that's genuinely fair for their contribution. And I think when I look at BitTensor, it's a group of people that look like a BC backed unicorn with the collective of capabilities and skills we have here. And you look at the resources, some of them we've discussed, you have an entire storage subnet solving that problem in a decentralized way. You have data scraping subnets, you have inference subnets, you have pre-training subnets, one of which you know, subnet line we run. And you've got all of the pieces of like a decentralized business that can compete with none of the like necessary behaviors of having to go, you know, beg Sequoia for $300 million. And that's really quite special. And I think that's certainly one of the reasons I'm here, but staff, I don't know if you want to give you a brief. - To me, it's pretty simple. If BitTensor's good at its promise, it's going to change the course of history. I think it has a lot of the machinery necessary to be a multimodal AI super computer. The first of it's kind, perhaps the only one. It's a ability to absorb resources at scale and enable coordination through this strange, competitive mechanism. It's just really fascinating. It pickles my brain. It's really hard to do. But I think it's really, really important. And I think we've got a good shot at getting it here. So that's enough for me. (clears throat) - I'll have one more point. Constory speaks about Bitcoin being the world's biggest super computer. The really interesting economist that I'll call the other day about data center build out. And if you actually look at like the addressable stack planned by 2026 for AI data centers, it's smaller than the currently addressed stack against crypto currencies. So you can see that decentralized market incentives can actually organize groups around a goal better than Big Tech. If you create the incentive mechanisms for people to point that at intelligence, there's already more compute directed to cryptocurrencies than there is artificial intelligence. All you have to do is point the needle in the right direction. And you could do something really, really interesting. - Amazing. So you guys mentioned the subnets you're working on. So maybe you could give a quick breakdown of the ones kind of falling under the micro or the macrocosmos umbrella. And then I also want you to touch on the, I saw a post, maybe it was yesterday on Twitter you guys mentioned, you have an academic research subnet coming out that does protein folding. It sounds like as its first objective. So let's, can we talk about what you guys are working on in macrocosmos? - You want to talk about privacy folding? - Let's do protein folding. - Okay, cool. Yeah, so why protein folding? Primarily, it's because it's, it's suitably hard and acceptively, like an accepted difficult research problem that we hope will get the attention that it deserves to bit tensor as a viable, decentralized computer, for researchers to use. So we could have chosen one of a thousand different interesting research problems until a few years ago, protein folding was considered to be impossible, frankly, you just had to do the hard work and it would take a really long time and you needed to allocate super computer time to make any real research effort. And it's, I think it's also very tangible to people. Well, perhaps protein folding in the abstractism, but it's a particularly useful technology is quite a direct line. Once you've folded a protein, whatever that means, you have now a biological replica of a system that you can use to develop pharmaceutical treatments, you can understand disease, you can do all kinds of things. You have a digital twin of a part of your body and that's something that you can examine and introspect an experiment with and it's just endlessly useful in a water. In fact, I believe a large part of the treatment for protein folding was actually a result of people doing really extensive protein folding simulations. So while the way that a protein folds is perhaps very opaque and complex, the finished product is something that is ready for analysis. It's distributed around the world. Anyone can analyze that protein now and then they can develop, again, treatments based on it. So that was really the excuse for existence for protein folding as a subnet. At a technical level, it actually maps very gracefully on Twitter. incentive mechanism. I believe that a good incentive mechanism design is very asymmetric. So it requires a lot of work on the valedade side that cannot be gained, cannot be circumvented. On the valedade side, the verification process is relatively cheap, straightforward and very objective. Physical systems are exactly that. Basically any kind of computational physics, usually there's real extensive calculations or equations that you have to carry out to evolve the system or to get it into some state, but to measure the quality or otherwise sort of assess whether your simulation was a success is usually as simple as extracting a physical property like the temperature or the energy and these distilled down to a single number where less is better or is worse, which is kind of a language that valedade should be speaking, just like loss functions to nature. So that's really a high level overview of why we chose folding. It's a really nice example of an incentive mechanism that actually solves a physical problem. And again, we hope that the benefits of being able to solve proteins at scale is tangible to the community into the wider world. And we want to invite researchers into use this, on demand free of charge to do their research and just treat the subnet like a super computer, which is optimized with state of the hardware for folded proteins. And this was part of a bigger initiative for us that we've codenamed the blue sky. And this is about, look, I think we've said it already on the school, bit tensor is a really big super computer. It's not just a gen i hot hax. So for us proving that you could do distributed super computer work on bit tensor that was relevant to research, it was good for humanity, that was equitable, which are like some of our founding goals was a massive win. And so, you know, Stephens backgrounds in particle physics, we've had some conversations without bridge national lab about could we do a particle physics subnet, could we do computational biology subnet? You know, we both hold research positions that we sort of probably don't do enough on. You know, can you provide a way for research entities to access essentially limitless compute in a way that the community can decide if it's important is a really, really exciting prospect for us. So that was a big push behind, you know, ethos and the mission behind protein folding was can we show what their tensor can do? You know, what other cryptocurrency is helping to solve like genuine research problems for human humanity? And could it be a sort of launch pad for us to develop the incentive mechanisms we need to do that? And there's been a lot of really brilliant research that our team have put into about new incentive mechanisms that didn't exist before. Like, how do you handle like reward problems that last over like 100 hours or like really long periods? Like, how do you deal with like tracking, in boxing and carrying those things? We don't see any other subnet that has done some of the things we've done on protein folding today. So it was a great bit tensor research problem and also something that sort of like tickles a bit of heart muscle. So we're excited for that to go into the wild next week. And is this this subnet's more than a proof of concept because it's actually delivering real world utility? When you do get the protein folding up and running and you have these, you know, academics that can come in and use this subnet and they want to do something like, you know, computational biology in a different direction than protein folding. Is that something that would require them to launch their own subnet or is it something that could be done on the same subnet? Or how does the architecture of solving those different problems work as bit tensor becomes a place where you can do something like this? So protein folding is a massive, massive deal. In some ways, every sort of family of proteins is its own optimization problem. What we have by design is during validation, we have what we would call stable hyperparameters that enable proteins to fold successfully but not to their most optimal configuration because there's some inputs that you have to provide. What this would unlock for researchers is we've designed the subnet so that through the sign apps or in more sort of layman terms. On the validator side or on the API client side, they can specify the conditions under which the protein must be folded and this is a really consequential decision because it gives the research our ability to inject their expertise. They're not just going to say, oh, this is the, you know, the four letter string that characterizes this protein, folded, give me a generic result. No, they're going to ask, give me 100 iterations with really specific parameters that even those researchers sometimes they spend weeks or months developing parameters themselves that they can pass as inputs to the network. So the network is actually as generic as possible because we learned very quickly on this path that trying to find the optimal parameters for all proteins is in fact an open research problem that no one in the world has solved. It just makes sense. Those proteins exist in different biological scenarios, different biological environments and they all have some of them fall in acid, some of them fall in water, high pressure, low pressure, high temperature, low temperature. The parameters are endless. So really this is just a vehicle. It's kind of like a compute optimized subnet. So it's like a subclass of a generalized compute subnet which is optimized to running state of the R protein-fall in software, but it's flexible enough to allow the user to specify how the specific folding simulation works. Perhaps zooming out one level further from protein-falling itself, the longer term roadmap as we'll point it out, we refer to it affectionately as deep blue sky. What this is is an initiative where we're allocating a large amount of resources internally to building what I would call incentive mechanism primitives that are going to be required to map traditional research problems onto subnet problems. So something that we already have in the pipeline right now is something that can just diagonalize absolutely massive matrices of scale. And this is actually something that can also be reframed quite gracefully as an incentive mechanism problem. So we're going to be sort of the in-house partners that researchers come to and maybe they bring us their legacy 1980s C++ code or 4TRI code Godfreyd. And this is something they've been using and we know all too well that academia is a very conservative community. They don't absorb new technological tools until they're absolutely proven. So we need to be able to work with the tools that are familiar and accepted in those industries and those fields. And we need to basically reproduce the working elements of those in a way that can be leveraging the massive parallelization opportunity of a subnet while at the same time allowing competition to take place and innovation to take place within the context of a subnet so that you actually can outperform your baseline software. That's actually kind of something we want to do here. What happens if every minor in our subnet decides to not run our base model but they run alpha fault? Well that's a good thing. Now we've got 250 alpha faults. That is a zero regret upgrade for the whole subnet. So we try not to overspecify the solution to our problem. We're just really careful to articulate exactly what we're looking for. And again in physical systems it's quite often quite simple. You just give me the lowest energy I don't care how. And that's enough to even convince my rather stoic PhD advisor that it's an acceptable technique. Excuse me. Very cool. It sounds to me like like a biological emulator. That's what the subnet. Do I have that understanding right? Something I'm going to mention for that. Yes, yes. I know all about the cellular autometer and the game of life. It's actually been a passion project for me as well for years. That stuff is really really interesting. Have you ever seen Wolf Friends or recent work on the sort of universal physics theorem based on that? No. I do know he's working on it but I'm not familiar with it. Oh, it's a great read if you've got three weeks. It's super interesting. His idea is just take Conway's Game of Life and just apply it to a limited dimensional grid basically. And you get all of this really, really trippy emergent complexity that just completely dwarfs what you see in the two dimensional space. And I look at that and the cogs in my head start turning and I'm like, a sensitive mechanism, sensitive mechanism. I think that that use case is perhaps a little blue sky even for deep blue sky. But maybe down the line we could jump onto that one. Next year. I think after protein folding, yeah, what we really like to see is maybe what we actually do is we rotate the subnet slot on let's say a 90 day cadence and for 90 days it's dedicated towards protein folding but we don't really want to just ride that to the wheels fall off. The point here is to launch to launch hard to make it clear that it works. The incentive mechanism has cleared. Then we can just rotate the entire subnet will switch out the incentive mechanism and we'll just allow researchers to submit proposals and they will just get free grants basically. They come to us they'll you know they'll make that case. Hey, I'm a computational biologist. I'm a chemist, whatever it is. This is the software that I want to run. I would usually have to wait nine months to get access to a tier two supercomputer to run my calculations but it's political. It's a little bit prickly. So instead what I'm going to do is I'm going to come to microcosm. I'm going to sit down with them. They're going to help me understand whether this can be reframed effectively as an incentive mechanism. First of all, if it can, we'll basically we can look at potentially rotating our dedicated research subnet onto that. If we think that it merits its own you know, if you're perhaps much longer of lifetime subnet, we could just build out a bespoke subnet for them. so that they can cover their research. So, she can get publications in high impact journals. We can support that. BitTensor as a whole, basically, as Will said, various, do you believe? BitTensor as a whole gets to celebrate the fact that it's creating actual utility to researchers, which is a lesson. Did you order that? (clears throat) I suppose there's a broader point here that developing out this suite of incentive mechanisms. So, weird side, but I actually sit on the mayor of London's infrastructure advisory panel for steady systems like Al Quiritt, and I ask. But a lot of real world problems, researchers are very interesting ground come into. So, Stephens said, "Optimized for energy, optimized for temperature, it's easy." There's a lot of similar analogies in certainly the built environment in the physical world, where, if you can address optimization problems people care about in an incentive mechanism, you can start to really solve problems. So, whether that's like energy good optimization, whether that's like, how do you optimize transport time tables? Like, I'm purposely throwing a couple more abstract concepts in here, but I'm really interested in-- - Contrary to the agilization. - Make it the organization. - Always an agilization. - But as our language of incentives evolves, and as like some of this deep worker, and how can you express complex physical world concepts in code and in incentive functions emerges, I'm really excited to see whether BitTensor can solve some of those tricky problems, because my third degree is in smart cities. What's interesting about a lot of city problems and distributed infrastructure problems is, they're really hard to help government invest in, but a lot of citizens of community people really care about solving these problems. So, finding a way for community and decentralized mechanisms to organize around these things that may be, like, personally, and community-wise very investable, but may be hard for a profit business to organize around, is really interesting to me. But that's like a 2025 problem. Yeah, we haven't put that in the world. That's a Q3 problem. - That's a Q3 problem. - That's Q3. (laughing) - Cool. - You guys talked about a couple interesting things that I want to touch on. So, one is the incentive mechanism primitives. And what you're talking about there are essentially the ways of implementing solutions to certain types of problems that you want to solve. You talked about the matrix diagonalization. So, a primitive would be like a, you know, a, the manifestation of the solution of the incentive mechanism in code, right? Do I have that right? This is like you solving it through computer code. - Yeah, it's like the primitives are almost the building blocks that you build more elegant incentives with. Like, you know, one of the functions of software is it's super composable. Like, you can pull bits and pieces together and build them into this wonderful edifice you want to solve a problem for. And, you know, in the same way that like, bit tensor and BTK help you sort of access bit tensor, we need to build out this whole language and package software, brackets and incentive structures for people to ask ever more complex questions and help people solve those problems in an exploit-resistant manner. So, one of the things protein-foilings really, really great for is it helps us solve this whole new wave of incentive problems that none of the other subnets have had to solve yet, which we hope will breed another wave of innovation by somebody sees our work and is like, shit, you solve matrix diagonalization and incentive problem. I'm gonna go use this with these other bits or this sort of bit of genius I have to build something else that you guys didn't dream of. And that's the joy of open source and decentralized communities. And, you know, generally for this year, we're planning to open source everything, to be honest. Like, we're gonna launch a chat app on top of one, we're just gonna give it away. Because we wanna help raise the bit TensorFlow's and help the whole community do better. So, you know, if we can build something that helps everyone steal it, please build on top of it. That's kind of our mandate, certainly for this year. - Do you think these primitives eventually become like modular? Because we talked about Wolfram, our team's done a couple calls with them, trying to convince them that they should be taking the research that they're doing and especially with the cellular automata stuff because Steven Wolfram posts stuff about it all the time and I always throw a post in there, like, when are you gonna put this on bit, tensor, man? And we've offered to like front their subnet registration. But in the end, I think the problem is they don't quite understand what the capabilities are around the incentive mechanism that's already been developed on the 32 subnets. And so do you think that will eventually like create, like I said, some type of modular system are almost like a plug and play deal where you can build a subnet without having to like go through line by line of the stuff that you guys have built or the other subnets and try to put it all together yourself. Like, do you think eventually there's a way to use artificial intelligence and some of the coding abilities to put these subnets together and create like, you know, your own Frankenstein incentive mechanism taking the best and the ideas from all the subnets that you want. Do you think that's a future that we are gonna be in at some point? - I think it makes a lot of sense to me. I think that what is an incentive mechanism, if not a strange love child, of an objective function that's adversarial and is resistant with some economical component. It's just more all sort of collectively trying to come up with a new lexicon around this object, which is clearly very powerful and when you sort of let an incentive mechanism lose on a decentralized system, you get your harness, this massive, massive compute reservoir. They're very powerful, but they're still very young. They're still very new. I think that we're gonna start thinking about incentive mechanisms, a little down the line, kind of like we think about building a model in PyTorch. Okay, so, circa 10 years ago, if you wanted to build a deep learning model, it was hundreds of lines of code. I was there, I was doing it. I was trying to figure out what the hell wasn't working 'cause it wasn't working. And you needed so much expertise and so it goes. This is how all technologies mature. At appropriate levels of abstraction, that release your concentration, allow you to think at a more executive level, allow you to be better at what humans are good at, which is sort of more creative abstraction, instead of thinking about implementation details. That's why co-pilot says coding assistants are really useful 'cause I don't have to drill into seats. I use co-pilot every day. It's not a great code, but it's much better than me to not forget where my eyes and jays go. And that's kind of the point. It abstracts away things so that your attention is released to think about things that you're uniquely good at doing. So just like building a deep neural network is as simple as just stacking the bunch of abstract layers together in PyTorch. It's a beautiful design. Why can't we do that in the time of time? Why can't we say, okay, well, my system is gonna operate on a time scale in the order of, you know, a synapse for the events, gonna be on the order of 10 minute round trips. So I probably should use tooling appropriate to that time scale. Also, I'm sending this much data over the wire because my requests are so big because I'm sending models or videos or pictures or whatever it is. Once you can sort of clearly define the, I say, sort of the hard constraints of the problem that you're solving, there should be a lot of off the shelf tools that you can just drop into place. So that, and again, there'll still be a lot of space for the innovation process. That's gonna be assembling these pieces and also adding your own bespoke twists to it. Again, just as you can get fantastic results out of making moderately small changes to a transformer architecture, you can do huge amounts with it. But the next research lab that comes out, they don't start from scratch necessarily. Occasionally, there's a disruptor and it's good for the ecosystem to bring a member or a jamber or whatever the next one is. They're a net good for the whole community, but for the most part, we have sort of as a community in research in general, we have the stepwise function moment, where everything was suddenly exposed to a new idea or a new tool and it's a massive productivity boost. And then there's several years of sort of consolidation, careful thinking about the different ways that we can apply it, making it more optimal. And then I think where we are right now is, we've all just sort of downloaded this massive stepwise change in thinking about how to harness compute. And we need to sort of mature, integrate that knowledge a little bit and then we can start thinking at the next meta-let, which is maybe like, you know, this is like 40 chess level. What do we want? My 12 subnet mixture of experts, of mixers of experts, of Mr. Vexperts to do. And like, you need to unlock that conceptual framework, but first of all, you need suitable abstractions, otherwise you just, if it takes 16 syllables to say, you're not gonna get that right guys. - What's the long-term plan for you guys being able to tackle all this stuff? So you've got, you know, two, three, four, subnets under management, you've got it, like you said, you've got a list as long as your arm for other subnets that you want to launch. How, like, are you guys hiring? What does the macrocosmos team look like? Are you looking for help? If someone is, you know, engage and excited about the stuff that you guys are working on, are you looking to bring additional people onto your team? What's the status of this? - Yes, so we are hiring like crazy to be honest. Like, we've hired four people this month. We're looking to hire eight people next month. And we're on ping our team really, really quickly. So, you know, we're a couple of months old, we're growing, we've got some great people. I think we already have more PhDs than the rest of BitTensor combined. So that's like a bit of a calling card. But, yeah, look, for us, we're looking for brilliant people enthused by AI deep learning, machine learning, backend engineers, conceptual biologists, I guess. And if you're excited, flick your CV to careers at macrocosmos.ai and our team can have a look at it. So, So we're very much recruiting, we're very much growing. And we want to push the boundaries of what a subnet owner team looks like. And I think some teams have shown us exactly what a team building product on top of subnets can do. We want to very much raise the game for what people building subnet incentive mechanisms can do and show what, how powerful this computer can be. Staff, I don't know if it's anything new or not. That's great. Yeah. Fantastic. Thanks for coming on the show today, guys. Anything else you want to share before we close it out today? No, thanks so much for having us, Keith. This was really exciting. I feel we could have gone for another few hours. Yeah, for sure. We could. We will at some point. When you guys get protein folding launched, it sounds like it might be coming as early as next week. Yeah, for sure. Yeah, for sure. Okay, cool. Yeah, thank you very much, Keith. That was a blast. Yeah, thanks, guys. We'll talk soon.