This Bittensor Subnet Could Cut Drug Discovery Costs in HALF | E2267

This Week In Startups is made possible by: Luma AI - https://lumalabs.ai/twist Every.io - https://every.io Lemon.io - https://Lemon.io/twistPlaud - https://Plaud.ai/twist Today's show: What do drug discovery, the creator economy, and AI vision models have in common? In the case of Metanova, Bitcast, and Score, the answer is Bittensor. Yes, each of the three companies leverages the Bittensor network to get more work done, more quickly, in a completely decentralized fashion. Metanova uses its subnet to run developer competitions to find exciting molecular candidates, parsing through a mountain of possibilities to pluck out the most promising for further investigation. Bitcast uses its subnet to collect visibility demand from brands, which is served by video creators. The company is focused on the crypto niche to start, but will expand in time to other technology topics. Score uses its subnet to generate highly performant, specialized vision models, which it then sells to customers through a platform (Manako). In each case, the Bittensor's economic engine unlocks global creativity to tackle tasks that were previously time-consuming, fragmented, or expensive to complete. Let's see how quickly each company can scale and whether startups building on Bittensor can grow faster than their non-decentralized peers. Timestamps: 0:00 Intro 2:19 Plaud: If your work depends on conversations — interviews, meetings, calls — you need a Plaud NotePin. You can check it out at https://Plaud.ai/twist and use code TWIST for 10% off! 3:40 What is Bittensor? 7:22 Metanova Labs joins the show 9:28 Lemon.io - Get 15% off your first 4 weeks of developer time at https://Lemon.io/twist 17:16 How Metanova tackles the multi-billion dollar cost of drug discovery 20:48 Every.io - For all of your incorporation, banking, payroll, benefits, accounting, taxes or other back-office administration needs, visit https://every.io 30:20 Bitcast joins the show 31:23 Luma AI - Luma builds accessible, professional-grade AI tools for creatives. Try Luma Agents for free at https://lumalabs.ai/twist 32:30 Mining crypto with YouTube 36:48 Why Bitcast is focused on the crypto space to start 39:11 How healthy is the creator economy? 47:26 When will the AI bubble collapse? 53:44 Score joins the show 54:42 How Score will make vision AI more accessible 57:16 VLMs v. LLMs 1:01:14 Demo of the Manako platform Subscribe to the TWiST500 newsletter: https://ticker.thisweekinstartups.com Check out the TWIST500: https://www.twist500.com Subscribe to This Week in Startups on Apple: https://rb.gy/v19fcp Follow Lon: X: https://x.com/lons Follow Alex: X: https://x.com/alex LinkedIn: ⁠https://www.linkedin.com/in/alexwilhelm Follow Jason: X: https://twitter.com/Jason LinkedIn: https://www.linkedin.com/in/jasoncalacanis Check out all our partner offers: https://partners.launch.co/ Great TWIST interviews: Will Guidara, Eoghan McCabe, Steve Huffman, Brian Chesky, Bob Moesta, Aaron Levie, Sophia Amoruso, Reid Hoffman, Frank Slootman, Billy McFarland Check out Jason's suite of newsletters: https://substack.com/@calacanis Follow TWiST: Twitter: https://twitter.com/TWiStartups YouTube: https://www.youtube.com/thisweekin Instagram: https://www.instagram.com/thisweekinstartups TikTok: https://www.tiktok.com/@thisweekinstartups Substack: https://twistartups.substack.com

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Hello and welcome back to Twist. Today is March 25th, 2026. My name is Alex and I am joined today by my dear friend Lon Harris.

• 0:08Speaker 01

  Hey.

• 0:09Speaker 00

  Lon.

• 0:11Speaker 00

  This Week in Startups is brought to you by Luma AI. Luma builds accessible, professional-grade AI tools for creatives. Try Luma Agents for free at lumalabs.ai/twist. Every for all your incorporation, banking, payroll, benefits, accounting, taxes, or other back office administration needs. Visit every.io. And Lemon. Building a great team is essential to any business. Lemon is a marketplace of vetted, experienced engineers, ready to take your company to the next level. Get 15% off your first four weeks of developer time at lemon.io/twist.

• 0:52Speaker 00

  How you doing?

• 0:52Speaker 01

  I'm doing pretty good. What, what day AO is it? What day after CLAW is it, is it Alex? We didn't?

• 0:58Speaker 00

  Forty, forty-seven, forty-eight.

• 1:01Speaker 01

  It's in the late forties or something, folks. We, we kind of lost track of that.

• 1:05Speaker 00

  Much like Lon himself, in the late forties and has somewhat lost track. Importantly, though, we are not going to be spending all of our time on OpenClaw today. Instead, we are going to be drilling down even further into the world of BitTensor. We have three different subnets on the show today: MetaNova, BitCast, and Score. I'm actually really excited about each one of these companies for different reasons, Lon.

• 1:25Speaker 01

  Yeah.

• 1:26Speaker 00

  But I love getting a diversity of projects, all part of BitTensor, because it shows what the project can do as a whole.

• 1:31Speaker 01

  It's a really cool thing about doing, you know, we were so focused on OpenClaw for a while and it's cool. I still enjoy OpenClaw. I've bonded with my agent. But a lot of the OpenClaw projects are, are kind of similar. It's people doing kind of similar things. Here's how you can use your OpenClaw. Here's how you can make multiple agents. Here's how you could do this workflow and that. And the great thing about making a show about BitTensor is that every one of these subnets, they're doing different things with the core technology. It's one core concept or idea, but then you could use it in a whole lot of creative ways for all these like interesting applications. And that's what, that's what I think is interesting about today's show is that it's three wildly different projects, but all built on the same. Kind of ecosystem.

• 2:11Speaker 00

  Yeah, drug discovery, social media for creators, and then also vision models for commercial applications. So we're going to be getting through quite a lot of things. But Lon, I think before we bring up our first guest, we should do a little PSA about our fun little devices that are listening to us as we speak.

• 2:26Speaker 01

  Yeah, we should talk about how we applaud Plaude. You may notice that Alex is wearing one on his wrist. I have one right here on my collar. These are Plaude pins, and all you do is you hit the button. It doesn't just record you and keep track of notes on everything that you said in the people around you said. It organizes them so that you can go through it easily later, figure out what you said, search through what was said. It identifies the people in the room with you. It gets to know the people in your life. So it really is kind of this magical device that takes interactions that you're doing throughout the day in your regular spoken, allowed life, and then sort of saving them for you and making them searchable and easy to look through later, so that you never sort of miss anything in conversation ever again.

• 3:08Speaker 00

  Absolutely. I use it for less of that, Lon, and more as my personal scribe for when I'm like holding a child and I want to remember a thought, an idea, a task to do. And I just kind of hit the button, drop it in, turn it off, and then I go back and I have kind of a list of things that I need to get done. An absolute lifesaver on my end. I'm a huge fan. And if you want to get a Plaude, you can do so. Go to plaud.ai/twist, P-L-A-U-D dot A-I slash twist. Use the code TWIST to save 10%. Look super fly. And then Lon and I will give you high fives when we see you.

• 3:37Speaker 01

  Or online, we'll give you a virtual high five if you get a Plaude pin.

• 3:41Speaker 00

  If you send us a tweet, we'll follow you back, or whatever the digital equivalent of that is. All right. Let's dive in. We're going to talk to MetaNova, or as I like to call them, Lon, subnet number 68. So please welcome Michaela Bazo and Pedro Pena to the show. Just for folks out there who are less familiar with BitTensor and maybe don't know quite what we're talking about. So I don't know who wants to take this, but the way we think about it, BitTensor is a decentralized network that uses crypto incentives to reward individuals who contribute useful AI models, compute, or results to task-specific subnets.

• 4:14Speaker 00

  Thoughts, guys? How can we improve that? Tighten it up so everyone can follow along?

• 4:17Speaker 02

  I mean, I think, as you were saying, day 47 after Claw, we're seeing that now it's humans and agents, right? So it's a marketplace for intelligence production. And there is a wide range of applications. Today, we're going to be talking about drug discovery. But as you were saying earlier, one of the things that makes it very unique is the fact that you can use this network to train any kind of AI use case or to develop any kind of digital commodity that you can think of, from renting compute to vision to drug discovery. And we're all in it together and benefiting from each other's success, which I think is something really nice that's coded into the way that the protocol works.

• 4:55Speaker 00

  Now, on that protocol point, I want to talk about the individual actors and players inside of BitTensor. So can you explain to me miners and validators as they relate to individual subnets?

• 5:05Speaker 02

  Basically, there's like three main actors, let's call it, in the way that each subnet works. You've got the subnet owner slash operator. In this case, it would be us, who's basically designing the challenges. And then miners can be anyone from around the world or increasingly any form of intelligence from around the world that's solving those problems. Validators are then given a scorecard, let's call it, and basically selecting, reaching a consensus and selecting which ones are the winners for each competition. So these are like competitions that are running 24/7. Think about a hackathon that never sleeps, that you can apply to any problem that you would like to solve.

• 5:45Speaker 00

  Pedro, tell me about staking and how tau and alpha tokens fit into this. Again, for folks out there who are just tuning in and learning this for the first time.

• 5:53Speaker 03

  So I think the best way to talk about staking is to see this as a way to vote on the subnets that you believe, right? Depending on the amount of stake that is flowing. Truly different subnets. The chain is going to define how much emission these subnets are going to be receiving. So it is also a mechanism for you to vote with your power on the different, very different projects that are running on Bittensor or the 128.

• 6:23Speaker 01

  And just to clarify, emissions are, that's how people are getting paid. Like that the the Tau gets or the tokens get emitted to them based on their work. And that's how their their fortunes rise over time.

• 6:33Speaker 03

  Yeah, it is. It works as also an incentive from the chain to everyone that is involved in the project.

• 6:40Speaker 01

  Got it.

• 6:41Speaker 00

  And just like Bitcoin, there's a hard cap of 21 million tokens for Tau, correct?

• 6:46Speaker 03

  Yeah, for Tau and for for the alpha tokens that are linked to every subnet.

• 6:51Speaker 00

  And just to make sure that I'm I'm tracking this correctly, each subnet has their own 21 million token alpha cap.

• 6:57Speaker 03

  Yes.

• 6:58Speaker 00

  Got it. Okay. Explain the goal and economic structure, please, of subnet 68.

• 7:03Speaker 02

  Subnet 68 is part of this grand decentralized platform that we're building for drug discovery. Why? Drug discovery is a very difficult, very expensive problem. Most people are describing it as being in a state of crisis with the average drug taking about $10 billion in 2.2. No, sorry, 10 years and $2.6 billion. Well, we see different estimates. And a lot of people are kind of shooting in the dark. It's a really hard question. There's a lot of points of failure. And so what we're trying to do is improve the virtual screening process so that we can make the best bets, so that they don't even feel like bets, so that we're de-risking what is really like the most asymmetrical bet that you could make.

• 7:48Speaker 02

  And it has an impact, not just like financially, but also on people's lives. So we launched March 1st of last year. And it was, first of all, a proof of concept of can we even do this in a decentralized way? Like, nobody had ever tried to do that before. And since then, we've grown to have two different incentive mechanisms. So right now, our miners are doing two things simultaneously. They're either Submitting molecules of interest based on whatever target we set for the competition. So we're like, hey, find us the most interesting molecules that bind to serotonin. And/or they're competing on their second incentive mechanism that's focused on chemical search algorithms. Why?

• 8:30Speaker 02

  Because chemical search algorithms allow us to basically look within the possibilities of the chemical universe in a very flexible way, and we can plug it to any kind of state of the art model, and also keep certain information private that might be sensitive from our partner's perspective.

• 8:46Speaker 00

  Is the second mechanism a way to automate the first, or are they distinct?

• 8:49Speaker 03

  They talk a lot with each other, and indeed, we can see that miners can get a lot of inspiration from the second mechanism to compete in the first mechanism. Because you can actually choose, do you want an open source or code that you're using to win in the first mechanism? Or do you want to keep it to yourself? So we have these two kinds of incentives, and you can go for the more open source route of winning with the code itself, or you can use this code to structure something that will really allow you to explore very, very vast chemical universes and then submit that.

• 9:28Jason Calacanis?

  When you're launching a brand-new company and designing your first product, you don't want to hire just any developer. You want the best, the best of the best, the creme de la creme. And that's why you got to turn to our friends at lemon.io. We've had many companies in our portfolio work with lemon.io, and they've gotten great results. And lemon.io is not just going to save you all the tedious legwork associated with building a team, they're also going to make sure you're getting matched up with top talent that fits within your budget. Here's how it works. Lemon has an experienced lineup of pre-vetted and curated developer source from Europe, Latin America, and the US, and they only accept the top 1% of applications to this very elite program. So you know you're going to be working with the best.

• 10:10Jason Calacanis?

  Lemon.io will collaborate with you to help integrate your new hires into your team. So go to lemon.io/twist and find your perfect developer or technical team in 48 hours or less. Plus, Twist listeners get 15% off. Off their first four weeks. That's L E M O N dot I O slash twist.

• 10:28Speaker 00

  Okay, when we think about molecules, Michaela, that bind to serotonin, just to stick with that example. As a non-scientist myself, a little economics, a little philosophy, not a lot of chemistry and biology. How many possible molecules are we talking about? Like, what's the library we're flipping through to then select from?

• 10:44Speaker 02

  So the number of possibilities, bigger number than I can even enumerate. And honestly, it's a theoretical combination of atoms. We're trying to focus specifically on what's synthesizable so that it is actionable. So that our miners are not just, you know, submitting something that then we can't get tested in the lab and advance into eventually like a drug candidate. We started off with a data set of a billion molecules. And then we layered on top of it five combinatorial reactions. So our miners are essentially recreating, like they're, they're using these generative approach to recreate what it would be like to synthesize molecules in the lab. And that brought it up to about 65, we estimate 65 billion possibilities.

• 11:24Speaker 00

  That's a lot of possibilities.

• 11:26Speaker 01

  Can I jump in here, Alex? I have a question. So being a non-biologist, after this process is done, when we've identified strong candidate molecules through this system, what's the next step towards actually synthesizing them into like a treatment that we could sort of test out on people?

• 11:44Speaker 03

  It is indeed the synthesis of the molecules. So miners are submitting a lot of potentially interesting molecules. There is a hit picking process, that is, that is the name, that you basically go through the best submissions and also consider a few other parameters. So like, maybe a molecule could work really well, but you have evidence based on the chemical structure that it could be super toxic. So they're not picking that one for synthesis.

• 12:10Speaker 01

  Right, yeah.

• 12:11Speaker 03

  So during the hit picking process, we consider a lot of other parameters to pick the ones that really go for synthesis. And then what lab validation, that it is doing what it should be doing. Because every model will have some level of success, but it is not perfect. But it is so much better than testing billions of molecules.

• 12:32Speaker 01

  So after the candidate molecule, you've discovered it, it passes the test, it's not toxic, it's worth trying to. Do you send it off to a lab? I mean, how, what's the, what's the actual next step in the process towards getting these drugs made?

• 12:47Speaker 03

  Yes. So we operate as what is called a virtual biotech. That means we are running with a very lean team focusing on increasing efficiency and avoiding a lot of overhead for running internal wet labs. So there are companies called contract research organizations that are specialized in synthesizing what you want them to synthesize and then testing those molecules exactly in the assay that you want it tested. That is how, that is how the core from it operates.

• 13:18Speaker 00

  And that's something that you guys are taking on inside of your company. So Bittensor to generate the possible candidates and then your company then takes those out to a CRO and then runs the test to see if they are compelling. And then at that point, Pedro, would you sell it to a different biotech company? Would you manufacture it yourself? What's that final step look like?

• 13:36Speaker 03

  So there are multiple paths that you can follow. Since we're talking about the process that is very long and very expensive, it is a game of derisking the assets and generating IP. So you're creating a IP along the way and derisking the potential drug that you are creating. So there are multiple points where you can be interacting with the, with the industry and they might be interested in, for example, licensing of assets at earlier or later stages, depending on how interesting is the target, the indication, and the validation that was generated in order for, for them to want to acquire or interact with this kind of novel IP.

• 14:19Speaker 02

  So if I can add, basically, the core motivation was to support our own R and D process, but we specifically or very intentionally built something that is completely flexible, target agnostic, so that we could also become a gateway for anybody else's drug development process. So it's very flexible, allows us to enter into any kind of code development or even offer screening as a service. We have one partner.

• 14:43Speaker 02

  Who's based in Shanghai, Yalutane, and thanks to them, we're expanding beyond just small molecules into another therapeutic class called Nanobodies. And together with them, we will be validating 50 of the candidates that are coming out of the, out of the, the subnet. So it's really about maximizing shots on goal.

• 15:04Speaker 00

  Okay, can I ask about that? Because, and this is going to show my ignorance of the actual science at work here, but let's say there's 65 billion possibilities, you guys are looking for serotonin binding, say, and people compete and compete and compete, and they find the best plausible molecule for it. Haven't you solved the problem? Why do you need 65 different candidates? Does this narrow to a single molecule? Or is there still going to be a range at the end for different variations in use?

• 15:29Speaker 03

  So that is the thing. You can go up to a point in terms of prediction if your asset is going to be efficient and safe. Right? In some cases, you are going to need to, to test it to, to really be sure in many kinds of assays. So the idea is that you can, if you can improve versus what would be random, there are a lot of gains that can be made and you can truly accelerate getting to cures. But as of now, I don't think there is any kind of approach that can get to the exact molecule from the, the first moment. It does need refinement along the way.

• 16:14Speaker 01

  Well, you gotta start giving it to people and seeing if they get sick. See if they get it.

• 16:18Speaker 02

  Right, but also it's like, each person may respond differently. So the goal ultimately is finding something that's safe and effective for broad populations. Or at least that's usually how, I mean, there's also a whole future in which we're like delving a little bit more into personalized medicine and we deviate from that. But at the current time, what we're trying to do is find things that are safe and effective. And commonly, they, they describe this process of drug development as like a funnel, because you're essentially losing options along the way as you move from cell lines to animals and eventually humans. And what we're trying to do is improve the predictive power of step one so that you don't have to go three steps in and figure out that you've been wasting a bunch of time and money on the wrong thing.

• 17:02Speaker 00

  And we've all seen biotech companies long go public and then their phase three trials fall apart and then their stock goes to zero and then you kind of cash it in, right?

• 17:09Speaker 01

  Exactly. Yeah, we're always hearing about like, oh, they're testing a new Alzheimer's drug. We'll see. You know, that's always this sort of very hypothetical. Speaking of hypotheticals, I have one. So to me, it feels like we're going to use AI to make new drugs that's going to cure all of these diseases like diabetes and depression and cancer. That's like the cornerstone argument for the AI optimism. Like, Americans are very down on AI. We keep telling them like, no, no, no, you don't understand. We're going to cure these diseases. What is in your mind, because you guys are sort of on the cutting edge of this, what in your mind is the timeline? Like, when are we actually going to start being able to offer people therapies that were developed by AI so we could start to actually make this argument for real?

• 17:52Speaker 03

  So we have a few assets developed using AI in late clinical trials right now. So we could be seeing this like earlier than we imagine. But the thing is, even though for you to prove that something that takes a long time and a lot of resources is working, especially when they're doing this the first times, it will take some time.

• 18:14Speaker 01

  Sure, yeah, of course.

• 18:15Speaker 03

  How much is some time?

• 18:17Speaker 01

  Oh, I don't expect it to be tomorrow. I'm just curious, in your mind, like, is this a, you know, in the future, we'll be dead, but our grandchildren will be fine? Or is this like five years, 10 years from now?

• 18:29Speaker 03

  No, I think we'll be seeing some interesting things in the next three to five years, considering everything that is being developed right now and that is undergoing clinical trials.

• 18:43Speaker 01

  Wow.

• 18:44Speaker 03

  It is becoming kind of the standard to go for these kinds of techniques because, again, it is truly, truly expensive. So if you can improve in any way, this is really relevant. You can bet more.

• 19:00Speaker 00

  That's a perfect segue to my next question. Michaela, you mentioned 10 years and $2.6 billion to get a drug to market today. How much do you think that can be reduced using Pedro's three to five year timeline? Do you think it could be five years and 1.3 billion? Could it be one year and 100 million? Like, how far can we compress the time and cost to get new drugs to market?

• 19:21Speaker 02

  This is where Pedro's gonna be like, do not throw a number because then we're not being accurate. So I'm gonna refrain from falling into the honey pot.

• 19:29Speaker 00

  You're allowed to just rip. No one's gonna, no one's watching. You're fine.

• 19:32Speaker 02

  No, I'll tell you some things. I'll tell you some things. Yeah, but also it's like, we need to be rigorous or else like everything falls apart. Like, do you want to scam? We can scam, but that's not what we're here for.

• 19:42Speaker 00

  No, no.

• 19:43Speaker 02

  What I can tell you that's really interesting is, A, there are many things to accelerate. So we are a decentralized company. We are decentralizing not just virtual screening, but also the whole R and D process. And that means we think that we can do geographic arbitrage, cut through red tape, accelerate timelines and slash costs even further by choosing the right place to get the tests done, by working with CROs. And for example, there was like a very exciting news a few years ago from a treaty between the Brazilian Health Department and Visa and the FDA, because the FDA is still kind of the gold standard for drug approval and like the gateway into the largest like addressable markets.

• 20:25Speaker 02

  I can't tell you exactly how many years, but we know that, I mean, the clinical trials done outside of the US drastically reduces the budget, right? And if we can prove or if we can work in locations and with partners that don't compromise the quality of that testing process, that means maybe we can get the best of both worlds, right?

• 20:49Jason Calacanis?

  So much of what we do here on the pod is about tactical and practical tips for founders. Basically, ways for you to save time, save money, and, you know, get those chores done and reduce stress in your organization so you can focus on what matters most in your business. Your business and here's a crucial one. Delegating your back office administrative needs to a trusted partner like Everi. Everi will take care of all of your incorporation needs, your banking, your credit cards, your payroll, your HR, your bookkeeping and even your taxes. And let's get you started today. When any of your staff spends on their Everi debit card, founders get 3% cash back, redeemable into their own account. Plus, Everi's platform is AI native while keeping all of your data secure.

• 21:31Jason Calacanis?

  So you can get access to all of the analytics you need without having to worry about compliance. So it's time to visit every.io for all your startup needs. That's E V E R Y dot I O. E V E R Y dot I O. every.io.

• 21:47Speaker 00

  Does the FDA's work in this case mean that we can do essentially American responsive trials outside of the country? Just to make sure that I'm getting this right. So the FDA would accept results, say, from Brazil in this case. So you can go faster and more cheaply, but still get that gold standard seal of approval?

• 22:02Speaker 03

  Yeah, yeah, for some specific cases, we're seeing that this is where they're moving to. So I think it is

• 22:11Speaker 00

  That's really encouraging.

• 22:12Speaker 03

  Very reasonable to imagine that we're going to have this substantially more integrated and more common as we are getting to harder and harder diseases to tackle.

• 22:25Speaker 00

  I want to go back to Bed Tenser because I have one more question I want to ask about this. When I think about other subnets like shoots, for example, you know, providing compute, it's pretty easy. It's kind of fungible. People could just bring it. When I think about finding the right molecule from such an enormous set of possibilities, it feels like a different type of challenge to me. And so I'm curious, how many people out there are potential miners for Meta Nova? Is that a large group of people? Is this a problem that a lot of folks can attack? Or is it a relatively small number of people with a niche set of skills and information?

• 22:58Speaker 02

  I think the beauty of this system is that in the way that we have designed this problem is that you don't need a background in this field to participate in it. We've like reduced it to a search problem. And that's actually led to very interesting results. So for example, in our second incentive mechanism that's focused on chemical search algorithms. We saw somebody apply an optimization strategy that's never before been used in drug discovery, outperform a well-established industry technique across a number of targets, across a number of challenges. So we're seeing innovation in being able to change

• 23:37Speaker 02

  the competition format so that it can allow for this like cross-pollination of ideas, right? So going back to this, we need like a hybrid intelligence model to truly automate science. At least that's kind of where we're coming at it. And that means humans, experts and non-experts, agents, and machine learning competitions that are hosted within BitTensor. If we can bring all of these things together, I think we are in the most solid ground to really accelerate the timelines and hopefully get surprises along the way. You know, we want to be cautiously optimistic, but at the same time, it's truly fascinating to see how much the world has changed since we launched the subnet. Like, everything's truly accelerating. And we're seeing like amazing new tools being developed that totally

• 24:22Speaker 02

  change and force us to update our priors on what we believe can be possible.

• 24:27Speaker 00

  So BitTensor then has been the right choice for your company to democratize this work that underpins your future commercial prospects?

• 24:36Speaker 02

  Absolutely, absolutely.

• 24:38Speaker 00

  Has it, has it met, exceeded expectations? I'm curious about like, the enthusiasm you had on day one when you launched the subnet to where we are today, better than expected?

• 24:46Speaker 02

  Better than expected, yeah. No, I would, I would say the day we launched the subnet, it was more like, so it's interesting to have a workforce that, A, you don't know, and that is in an adversarial relationship with you, right? Like, it's not your usual office job. Like, the miners are actually quite unruly. So when you first launch incentive mechanisms, there's this, this warm-up period, or this learning curve, where they're trying to exploit you, they're trying to break your subnet, and-

• 25:17Speaker 00

  Trying to game it, yeah.

• 25:18Speaker 02

  They're trying to game it, but there's actually, we noticed, what we thought, you know, it's actually a feature, not a bug. So one thing that we found is that at the very early stages, what the miners were doing is they were finding the shortest path to a reward. And that meant that they were pointing to us the areas of low confidence in these state of the art models, because they don't have the same bias, right? And everyone else is training us, like in a private company that's trying to protect their valuation, or in a research institution that's trying to publish good results. Our miners do not care. They want to make a token. They want to earn their tokens. And that can, that can be, you know, a little tense. And it's can be a high stress environment, but ultimately, it can

• 25:59Speaker 02

  add resilience to the system. And if we can use that information to build something that has a higher predictive capacity, then I think we have a competitive edge in the technology that we're training in a decentralized way, that goes beyond just resource efficiency and tapping a global network of like, really cracked engineers that are competing for your token and for creating increasingly more valuable commodities. The true challenge becomes, can you program their behavior and align them in a way that generates valuable inputs?

• 26:33Speaker 01

  You're productizing the unruliness, essentially.

• 26:35Speaker 02

  Yeah. And we believe the answer is yes, but it's a very dynamic system. But at the same time, wouldn't it make sense that a dynamic system would be the one that would create the most interesting technology? Like, we're constantly responding to what they're submitting and being like, actually, we need to steer them this way and that way. And, and that means that ultimately, we, we have a living mechanism that can learn from the results, from their behavior, and you can tweak the incentives so that they can yield different results. And that's very interesting in and of itself, outside of the context of drug discovery.

• 27:13Speaker 00

  Do you need like an in-house economist to help manage the stuff that you're describing? I'm not even really kidding you, Michaela. Like, to me, it sounds like you could have, you know, just get yourself a Chicago School PhD in econ and set them loose to tune and tweak and improve your economic incentives on the platform.

• 27:28Speaker 01

  Oh, I thought you were pitching yourself for this.

• 27:31Speaker 00

  No, no, I dropped that back.

• 27:32Speaker 01

  You're a co-host.

• 27:34Speaker 02

  I mean, this is actually more like Pedro is, I think, best suited to really comment on this because, you know, he's constantly tracking submissions and also recently recruited an agent to help look at some of the work that we've been getting from our miners. And Pedro, do you want to talk a little bit more about, you know, how to iterate on the incentives and the submissions?

• 27:59Speaker 03

  Yeah, no, it is actually very, very interesting because miners, they have a power of breaking like really well-published methods in a way that is insanely fast. And that is actually really, really good because in many cases, you just don't know what are the potential problems in your scoring function, right, with what you're using to do the predictions. And so after you know that and you understand how to consider that in the structuring of the challenges, then it really, it really goes well, right? But it is constant iteration. It means we need to be checking what is being submitted to be sure it is aligning with

• 28:43Speaker 03

  the long-term value generation. Like, this is absolutely our responsibility. At the same time, a lot of interesting things comes from that, right? We started with one mechanism, and then the second mechanism was also a way for us to increase the competitiveness and kind of have everyone sharing their super interesting new approaches to look for molecules. And so you need to be creative on how to do that, but when you get it right, you can get really, really good things very fast, right? The way we are going to be plugging all of those things with the agentic economy in

• 29:26Speaker 03

  this wave of agents is also something very interesting because we are considering what is the best way to also make the outputs that we are creating integrated with agents because we do envision a lot of agents doing...

• 29:44Speaker 03

  A bunch of applied science in the next few years. Right. So we need to be able to integrate all of that. We are implementing some very, very interesting agents. There are some that are already in production. And for example, helping me select the molecules that are going for synthesis or helping me check which ones might already be covered by patents. So not ideal for us to explore. This is, this is like really, really interesting.

• 30:14Speaker 00

  Yeah, I was thinking about some combination of auto researcher from Andre Karpety, my local open claw setup. And then somehow doing useful work for Tao subnets or Bitinto subnets. Like I feel like I should be able to put my agents to work somehow to help with something here. So I wonder in time, how much the ratio of humans to agents doing work at the minor level will, will shift, but well, that'll probably take a couple of years, I think.

• 30:43Speaker 01

  Mine's still just writing my tweets for me, but.

• 30:45Speaker 00

  Oh, well, that explains why they're so bad. URL is metanova-labs.ai. You can take a look at it. Michaela and Pedro, thank you so much for coming on. I appreciate it. And I think I fully understand it. So rock and roll, you guys are the best.

• 30:56Speaker 01

  Up next from subnet 93, we've got Tom Bleers. He's the co-founder of BitCast Network. This, Alex, is a subnet where miners compete over who could generate the most social media views for a brand, a product, an individual, whatever, whatever your project is. This is a way to crowdsource user-generated content about whatever you're working on. Tom, thank you so much for being here.

• 31:21Speaker 05

  Thanks, guys. Really good to be on.

• 31:23Jason Calacanis?

  AI models can produce stunning, realistic video, but going from an idea to something polished enough to publish, that still takes a long time. Even the pros spend half their time managing tools and jumping between models rather than actually creating. But now there's Luma and Luma agents. Luma is not just an aggregator of third-party models. That's the future of AI. They know the best tools for that task. If you're doing a website, if you're doing a promotional video, while you stay focused on bringing your vision to life. And Luma has just introduced their powerful new model Uni1, which understands your full context and turns your original idea into a beautiful finished work. Text to image was just a demo, but reasoning to image, that's the real product in the future.

• 32:07Jason Calacanis?

  You're not just typing in a prompt and walking away. Luma puts you in the director's chair. Luma is going to 10 extra creativity, not try and replace it. To try Luma's agents for free, go to lumalabs.ai/twist. That's L U M A L A B S dot A I slash twist.

• 32:24Speaker 00

  The thing that really grabbed me about Bitcast when I was prepping for you coming on was the idea of mining being something so far away from the proof of work, you know, Bitcoin OG setup that I'm familiar with. So before we dive too deep into this, Tom, can you explain just the economics of Bitcast? And how the value and tokens flow?

• 32:45Speaker 05

  You're spot on. We're very unique in a way on BitTensor and our miners are YouTubers. So people essentially mine crypto with YouTube content. And the platform, we can go into a lot more detail on how it all works, but essentially takes care of automating the whole process from creating the briefs to actually measuring attention. Yeah, and the more attention that you can generate, the more rewards that get issued to creators.

• 33:16Speaker 01

  This was so, after years of working on YouTube, working for creators, working on YouTube channels, the brand deal advertising part of it is such a huge part of the job. And it's such a time suck. And it's really kind of like, there's so much uncertainty about it. You get like a little brief from a brand, here's what we want the video to be. But then you're always sending it in and you're kind of waiting on pins and needles. Are they going to like it? Are they going to reject it? Did I miss, did I miss saying the one sentence? And so I love the efficiency here of figuring out a model where it's like, here's exactly what we want in your video. And if you do a good job and you hit these three metrics, you get a little tau out of it.

• 33:54Speaker 00

  Okay, so then how does validation then? Do validators determine if the views that are accrued to content made from briefs on Bitcast are high quality? Tom, talk me through how you ensure that people aren't just putting out slop and trying to, you know, stick 15 views together 100,000 times to make money?

• 34:12Speaker 05

  Yeah, 100%. So when we release a brief, essentially that will say a video needs to talk about point A, B, and C.

• 34:22Speaker 05

  As an example, and then creators create content that matches that. And they're basically scored on how much watch time. It's nothing to do with views, it's to do with how long they can keep people on the videos. So that comes down to, obviously, the competition here is to make the most engaging videos possible, get the information across as accurately as possible. And the more you can keep people watching, the more you will be rewarded.

• 34:47Speaker 00

  Okay, so it's length of time, okay.

• 34:49Speaker 01

  And so it's not like if a video is made by AI versus having people in it, you're not really that concerned about it as long as the watch time is keeping up with the correct metrics.

• 35:00Speaker 05

  Yeah, and interestingly, when we launched the subnet at the beginning, and Michaela touched on it earlier about people trying different things and exploits, et cetera. We did have a lot of people creating AI videos.

• 35:14Speaker 05

  But it just generally, it doesn't seem to convert very well. Like, people like to see people, they like to talk to people. So that has been experimented with. I don't know why the future is going to land us with that. But yeah, it's generally people bringing stories, bringing information to life, and putting their own creative twists on it.

• 35:33Speaker 00

  Does the system result in a little bit less brand control? Because in the scenario that Lon mentioned, you know, no one wants to submit their video or a piece of content and then wait to hear back and then maybe have to do it again. That process is terrible, but it does leave kind of a lot of the power of authority in the brand, Tom, in their hands. Versus in Bitcast, it feels more like the brief is put out, people react against it, and then they earn a share of emissions via view time. But the brand in question has less control over what goes on. Does that worry them?

• 36:04Speaker 05

  The company, it all comes down to the design of the brief, doesn't it? So the briefs, they do go through all the points that you want creators to go into detail on.

• 36:15Speaker 05

  Alongside the brief, there's an information pack. So if the brief says, talk about point A, B and C of how they achieve this. The information pack will give you all of that information. So that's all accurate information provided by the the brand. Um. But then we're also leaning on the fact that the these creators are, you know, th- these are very, very well established creators. They have their own reputations and they are leaning on and they are, they're, they're providing their opinions. So, so like all of that combined, it results in really, really good uh videos. And we've not had any instances where people haven't been happy with the videos.

• 36:49Speaker 00

  Oh.

• 36:50Speaker 05

  Yes, the, the whole, I mean, th- this technology, yeah, th- this, this technology, um. It really what it can achieve is, is not really been done before. So we can now get, you know, hundreds, if not thousands of videos created. All at the push of a button. And if you think about the admin saving on that, it's all validated against the creator, against the brand's messaging. So you can get waves and waves of content at the click of a button, whereas previously that would have taken weeks, months, maybe even years to get that many videos out. So i- it's extremely powerful in that regard.

• 37:27Speaker 00

  So do you think this is going to lead to an overall increase in brands that want to work with creators and creators that can monetize their videos? Or does this more replace the current, the current ecosystem and marketplace of creators and brands that we have?

• 37:39Speaker 05

  Well, it's a tool that can be used by brands and marketing agencies. We're not here to completely replace the way that people do things. Um. But it's a very, very effective way of, of generating a lot of content with reducing all of the admin. Yeah.

• 37:54Speaker 00

  And what industries are the best for that, that fit? Because I presume that, you know, if you're selling two million dollar watches, you don't really want a mass audience with a very target niche audience, but there's probably a lot of stuff that does fit into this. So what's the best or what's the current most popular brand type that want to use Bitcast?

• 38:13Speaker 05

  So currently we're working within crypto. We're working with a lot of Bittensor subnets and we're starting to work with some exchanges as well. Um. We're, we're still developing the software, developing our AI that analyzes all the videos as they come in. But beyond this, we see sort of AI, general AI and general tech as very, very good niches to go down. So if you've got a new product release, let's say you're Microsoft, you've got a new product release, you can get loads of targeted YouTubers to give a breakdown on how that product works. Again, you can sort of apply that to most industries really. But you are right, like a unique watch, really high ticket item might be more aligned to go

• 38:54Speaker 05

  and with your Leonardo DiCaprio's. But if you want to get information out and breakdowns and demos of how products work, and you want to get a lot of people to understand you've got a new feature, you've got a new product that you're releasing. Yeah, we can release that into a targeted pool of creators to bring that story to life.

• 39:13Speaker 01

  I mean, the one that jumps out at me immediately is like fast food. Like every time a new fast food restaurant introduces a new item, you see the wave of people who've obviously been compensated on Instagram to like, go try it out. Like, hey, I've got the new BK chicken, whatever. And like, I feel like that is the sort of thing that this was just like purely designed for. You could just, Burger King could just pay for thousands of people to sort of pretend to enjoy their new chicken sandwich at once.

• 39:40Speaker 00

  Whoa, whoa, whoa. You don't like chicken sandwiches? What the hell?

• 39:43Speaker 01

  Not for Burger King.

• 39:44Speaker 00

  Someone grew up fancy. All right. Sorry, Tom. Would Burger King be a good fit for BitCast if they wanted to promote their new chicken Whopper?

• 39:51Speaker 05

  Probably not right now. But no, we're keeping it to quite sort of technological sort of topics at the moment. If you think about what YouTube's really good at is sort of putting faces behind names and explaining how things work. And so we're leaning, we're leaning into those sort of tech verticals with where we're sort of going to market.

• 40:14Speaker 01

  We're doing it on This Week in AI with demos. That's, that's YouTube loves a demo.

• 40:19Speaker 00

  YouTube does love a demo. What is the health of the creator economy writ large? There was a wave of startups, some number, you go three, four, five, maybe, Tom, when the creator economy is going to do this big thing and everyone thought that there's going to be, you know, 10 times as many creators, but it seemed to kind of end up like the power law, like the top 1% made 80% of the money and things seem to have gotten a little bit quieter. So I'm not as familiar or aware today of just how strong the creator economy is that Bitcast is going to tap into to create the videos, if that makes sense.

• 40:48Speaker 05

  So the creative economy is absolutely booming. And I know what you're sort of touching on there with the top 1% taking a lot of the revenue, which is a problem that we are solving, I'll come on to a minute. But the creator economy is absolutely booming. And the, I mean, I think the projections at the moment is that it's about $250 billion worldwide. It's growing much faster than any other, much faster than any other ad medium. So paper clicks and, you know, sort of traditional ads that you see in newspapers, etc. It's growing much, much faster than any of those and its return on investment is much, much higher. And this all boils down to trust.

• 41:31Speaker 05

  You trust the people that you follow and you lend your attention to, if you will.

• 41:37Speaker 05

  But on your, on your point about the top 1% taking a lot of the earnings. So that's actually a result of the admin behind launching marketing campaigns with creators. So if you were to go and work with, let's say you had the budget and you, you, you want to go and work with 10 creators. The effort to go and get 10 creators to talk about your brand, run through the brief with them, get them to understand what you're doing is very onerous. So you're only going to go for the top creators because your bang for the buck on the admin is way better spent. However, what Bitcast actually unlocks is, is really powerful when you think about it,

• 42:18Speaker 05

  because we now have removed all of the admin and the return of investment and the actual trust and engagement for smaller creators is much higher than big creators. It feels a lot less commercialized. Now you can activate all of these creators at the push of a button. So you're actually managing to tap into that 99% of creators with your same budget. And yeah, it's, it's, it's, it's a theory and it's a sort of point that we're sort of leaning into.

• 42:51Speaker 00

  Yeah, so it democratizes who can participate, who can make money from being a creator. And it probably also evens out the payments a little bit because if it's just done on view time, to your earlier point versus, you know, Mr. Beast probably has a premium on a per impression basis because he's Mr. Beast. So it makes it more fair as well. Lon, this seems very democratic to me. I like it.

• 43:12Speaker 01

  Yeah, I mean, I think you're, you're hitting the long tail of creators. There's a ton of creators out there who have, you know, a few thousand, a few tens of thousands of really dedicated fans who will listen to anything they say and who, if they started to give brand messages, would probably pay a lot of attention. And if you sort of cast a big enough net over that group, you're still talking about millions and millions and millions of people. I think right now, everything is so consolidated on like, you know, the Mr. Beast of the world, the people who have those incredibly high profiles. A system like this is amazing for like scooping up a whole bunch of those people that are in the middle, you know, they're not, they don't have nobody following them. They have a lot of people, it's just not record breaking, competing at the very top of the charts numbers.

• 43:57Speaker 05

  Another thing that we can actually do, which is very, very novel, and I think we are the only people in the world that can actually do it, is that our system can work with obviously any creator in any place, but any language as well. So we could have 100 different languages all on the same campaign because it's AI that validates what they're talking about. It, it's language agnostic. So as we sort of build this up and introduce much more, many more creators. And as we grow into other software verticals, like the, the reach that we can do would possibly take multiple teams, all speak in different languages, all speak into different creators. So the time savings on the admin size are, yeah.

• 44:39Speaker 00

  It's the AI verification though, that I keep getting a little bit. Stuck on Tom, because if I'm thinking about a creative brief, to me, that's a very human document. Because, you know, humans are trying to talk to other humans through the medium of a creator. So I, I, I guess, you know, as Mark said, AI checks the videos that are created. How, how strong is that? Are there problems with it? Have you improved it? Is that an easy problem that I'm overestimating the difficulty of?

• 45:04Speaker 05

  That's essentially what we've been learning and that's what we've been developing. And, you know, that's part of the benefits of being on BitSens. We've got access to some of the best AI tools in, in the world. So, um, yeah, we, we are always iterating our process. Um, and the, the video, like, at the moment, we're sort of checking, you know, 50 to 100 different videos a day against the brief. Um, that system needs to scale up and we're getting it to a very, very stable position now where we could scale up way beyond that. Um, but yeah, it, it's, it's a problem that we're working on and, um, we're getting really, really good results now. Um, and, you know, it, it, it has to, it has to scale up to sort of 100,000 videos a day, really.

• 45:48Speaker 05

  Um, and based on what we're seeing at the moment, we're, we're getting close to that.

• 45:52Speaker 00

  I want to close with growth because when you guys wrapped up your 2025, you said in your Substack post, uh, that in the last two months of the year, I think you said your, uh, hours watch was up 60%. Views were up 56% in the same period. So a lot of growth going into 2026. How has the new year treated you? How's Q1 gone?

• 46:11Speaker 05

  Really, really well. So I think we sort of flipped the switch over New Year. So we were developing a lot last year. We flipped it into sort of scaling up the network. Um, we're now, our creator network's now at 2 million subscribers, uh, and, uh, 50, 50 different YouTube creators. Um, but we are accelerating about 40, 50% a month on creators. And our watch time and views are, yeah, about 60, 50, 60% month on month at the moment. And that's no sign of slowing down. Um, in fact, we are getting creators every, every day trying to join the network.

• 46:49Speaker 00

  Is, is there enough brand demand inside of your current niche? Cause you're Targeting a particular slice of the market to start, is there enough brand demand there to support that many creators? Or do you need to start opening up to other niches inside technology sometime soon?

• 47:01Speaker 05

  The whole of last year, we were just working with Bittensor subnets. The start of this year, we decided to outreach and we've started working with one of the biggest exchanges in the world. We're now in conversations with many others. And yeah, so we're now starting to attract capital from outside of Bittensor. And the demand for this product is, yeah, there's a lot of demand. There's a lot of custom out there just within the crypto industry. And obviously move that out beyond that. We've got creators, if we get creators talking about tech, like I mentioned before, or AI, there's thousands of creators. And there's massive, massive budgets there from centralized labs, the biggest companies in the world. And, you know, they're going to be very interested in what they can achieve with BitCast.

• 47:46Speaker 01

  They're coming for Oliver's AI demos. Oliver, look out. It's a competition. Well, when you are ready for fried chicken demos, that's when I'll be ready to jump in. That's my expertise. Tom, thank you so much for joining us. Bitcast dot network is where to go if you want to check it out and become a creator, become a miner, become a validator. Just learn more about it. Tom, thank you so much for being here with us. We appreciate it.

• 48:09Speaker 00

  Subnet 93, everybody. We love to see it.

• 48:13Speaker 01

  We love a numbered subnet, don't we, Alex? We love a numbered subnet.

• 48:17Speaker 00

  Well, you know, it's okay, it's, it's, I have a lot of thoughts about this. One, there's a lot of numbers used in company names in China. So something that I've become more accustomed to just as time goes along. Also, it just strikes me as slightly science fiction to have like a series of numbered subnets in this way. And so, you know, I'm a big nerd, Lon, and so this actually kind of works for me.

• 48:37Speaker 01

  Like I, it does feel a little sci fi, like a community where everybody is identified by their number. It's like The Prisoner, but for tech.

• 48:44Speaker 00

  Before we jump into our interview with Scorelon, I want to bring up my new favorite polymarket of all time. Because often polymarkets are a little bit binary, you know, like, okay, will Elon Musk tweet five times before noon? Or whatever it is.

• 48:58Speaker 01

  They're so specific now. They get so specific.

• 49:01Speaker 00

  You can literally prediction market like a five-minute Bitcoin price changes, for example, but this one is called AI Bubble Burst Buy. It's a question about when things will turn, if they will. And if you're watching the video version, you can see there's a 24% chance, according to the sharps over at Polymarket, that it'll happen this year. But what's interesting, Lon, is the terms here. Talk us through it.

• 49:26Speaker 01

  We always say you gotta look at the rules. And I think this is maybe, of all the polymarkets we've ever looked at, the most important to look at the rules. Because just will the bubble burst? It's like, well, how, what does that mean? How are we defining it? And here's how they're defining it. So the AI industry will be considered to have experienced a downturn once at least three of the following events have occurred within 90 days of the timeframe. So by December 31st, 2026, 24% of the people on Polymarket are betting that three, not one, but three of the following things will happen, and they are Nvidia's closing stock price is down 50% from its all-time high.

• 50:09Speaker 01

  The iShares PHLX Semiconductor ETF, that's SOXX if you're looking at the stock ticker, that closing price is down 40% from its all-time high. OpenAI or Anthropic declare bankruptcy. OpenAI gets acquired. The rental price for an H100 chip falls to a dollar or lower for five days straight. Or some of these major AI hardware suppliers, their stock price goes down 50% from its all-time high. That includes, you know, Taiwan Semiconductor, TSM, ASML, Broadcom, Super Micro, the big players in the chip world. So I could see one of those things happening by the end of this year, but for three of

• 50:52Speaker 01

  those things happening by the end of the year, it strikes me as pretty remote, more remote than a one in four chance.

• 50:59Speaker 00

  Well, that's well, that's what you and I think, but you know, people can take the other side of that bet if they want. But what I appreciate here about this is if three of these things happen, I think it's actually very fair to say, yes, the AI bubble has burst so much as there was a bubble to begin with. So like, this is a really well framed bet. Now, I don't think that we're going to see OpenAI or Anthropic declare bankruptcy because they have-

• 51:22Speaker 01

  It's impossible, right? Like, they would have to go on like a wild, insane, like Vegas spending spree, like, like all time-

• 51:30Speaker 00

  They'd have to buy Guatemala. Like, I mean, I don't know.

• 51:33Speaker 01

  I don't know how that could happen.

• 51:36Speaker 00

  So that's not gonna happen. Now, Nvidia's share price losing 50%. I did the math before the show, so this is probably a little bit out of date now, but it's off about 15% from its all time highs, which is not that much, given how bubbly that stock has been.

• 51:47Speaker 01

  Yeah.

• 51:47Speaker 00

  But to lose 35% with the 75% gross margins, 2 trillion in spending for this latest GPU, I don't really see it.

• 51:54Speaker 01

  We're talking three fiscal quarters. We're not talking, like, if it was 10 years from now, you know, anything could happen, but that's just not that long a time period.

• 52:01Speaker 00

  Yeah, it's not that long a timeframe.

• 52:03Speaker 01

  Yeah.

• 52:03Speaker 00

  Also, if you're curious, an H100 hourly rental today is about 750, according to the data source they're using for this bet. So to have it collapse to $1 would imply an absolute inference an AI compute collapse law.

• 52:16Speaker 01

  Yeah.

• 52:16Speaker 00

  Not super likely, I don't think, but the OpenAI acquisition thing caught your eye. Tell me why.

• 52:21Speaker 01

  I, well, I just, I, yeah, I mean, that one feels like a weird rule. Like, OpenAI being acquired, it could be a sign of a, like, a collapse. Like, it's worth so much less now, its value has plummeted, everybody stopped using ChatGPT, Yahoo's gonna pick it up. Like, I could, like, I'm kidding about Yahoo, but like-

• 52:39Speaker 00

  I used to work for Yahoo, so like, you know.

• 52:41Speaker 01

  But like, but, you know, I could envision also a scenario where OpenAI gets acquired, and it doesn't necessarily mean AI is dead and over. It just means, like, another mega deal happened and two companies-

• 52:53Speaker 00

  It could be super bullish.

• 52:54Speaker 01

  Right.

• 52:54Speaker 00

  If someone buys it for 3 trillion, then, I mean, then Sam Altman's walking on the moon.

• 52:58Speaker 01

  Exactly. So, so that one struck me as, as, as kind of weird, but I don't know. Obviously, not financial advice, but overall, to me, you know, 76% feels like a pretty-

• 53:08Speaker 01

  Strong wager at this point, you probably make 24 cents on your on your dollar there.

• 53:13Speaker 00

  Yeah, but this is this is like, this is actually what I want people to use prediction markets for. And if this is my old man yells at cloud thing, but like, this is a hedge. If you have a lot of like exposure to like AI stocks, you could buy the other side of this contract and literally just hedge your hedge yourself. And that's super duper cool with me. That's that's the type of thing that I'm most excited about.

• 53:35Speaker 01

  I feel like there's a lot of culture in like, if you're inside the tech industry, the idea that any of this could happen by the end of this year is like, you sound insane. Everybody would be like, what are you, what are you talking about? Like, we're, it's just like we were, we were just talking with Menonova, like, things are accelerating. That's from the perspective here at this, we could startups, things are moving faster than ever. This stuff is being adopted more than ever. Compute in tokens are more valuable than gold, literally. Like, that's how people in the industry feel. So I think this is a lot of like, if you're outside the industry, you're like, I don't even use chat GPT that much. If you're not coding, so you're not using Claude code, or you're not on Codex or whatever, you're not on Open Claw.

• 54:17Speaker 01

  For a lot of people, everyday people, I guess it still seems like, well, this could all just go away at any time. For Elizabeth and tech, that sounds crazy. It's like, it's too embedded in our everyday lives for it to ever go away.

• 54:29Speaker 00

  The, the average person has asked chat GPT to write them a poem four times at this point in time. And that is not enough, I think, to get your feet let enough to learn how fast the water is rising.

• 54:39Speaker 01

  Exactly. I think that's, I think that's a big part of what we're seeing here is people outside of the tech industry wagering on this without really understanding how deep it goes already.

• 54:48Speaker 00

  I want them to fire up Claude code and have it build them an app and then run it. Because I think if you, once you discover that that's possible, then AI does feel like the old bicycle for your mind thing. Like, here's a thing that lets you go and do so much more. Anyways, all right, let's do our third interview because I'm stoked about this one. I did the pre interview for this, and let me tell you, the technology is astounding. So we're going to talk to Score, or as we call it, Subnet 44 over on BitTensor. Please welcome to the show, it's Max Sebti, and critically, Lon, he is in Paris. And we have had people from around the, around the world on the show today, but it's always nice to see France show up, the home of Mistral, one of the world's leading AI labs. Max, hey. Yeah, thanks.

• 55:31Speaker 06

  For having me, guys, definitely.

• 55:32Speaker 01

  Great to have you, yeah.

• 55:33Speaker 06

  Happy to be with you today.

• 55:34Speaker 00

  So let's start with the same question we're asking everybody. What is the goal of the score project? And also, how did the economics work?

• 55:42Speaker 06

  Yeah, definitely. So the goal of our project is to literally give AI sites. So our subnet is building a vision skills, let's say, for agents or human users to build vision AI apps. So the whole thing about, you know, vibe coding and stuff, it's really, you know, you just said it, you know, you can download Codex, for instance, or Claude Code, and you can start building something and it's empowering. It's crazy. It's a beautiful experience. And we think that if we want to go a step further and go beyond text-based intelligence, we need to give agents and people the ability to build with vision as well.

• 56:21Speaker 00

  I love that. Lon?

• 56:22Speaker 01

  I was just gonna say, as somebody who has constantly tried to get his agent to watch YouTube videos and have them rely exclusively on captions or a transcript, I could not agree with this more. Like, I want my agent to be able to see the video, not just read what was said.

• 56:38Speaker 06

  Yeah, and you just mentioned Polymarket a few seconds ago. Imagine someone being able to stream something like Mention Markets or even Sports Markets and then get an agent to act differently based on what's happening on screen. This is the type of thing we want to unlock.

• 56:51Speaker 01

  Yes.

• 56:51Speaker 00

  But this brings up the question of what type of vision model are we talking about here? Because there's a lot of things you could watch. So how broad are the vision models that people are bringing to score via BitTensor? Or how narrow are they, I suppose?

• 57:05Speaker 06

  So it's exactly like your second question about, you know, what's the economy around the subnet? The economy is pretty simple. If you try to use a state-of-the-art VLM, which is a vision language model, it's actually quite accurate, but it's not built to run in production. It's built to be very accurate on very specific things. And most of the time, to access computer vision, you would need to be an expert. You would need to know how to code. You would need to know how to train a model. And you would probably end up with something very accurate, but then way too expensive to run in production anyway. And this is like the biggest problem in vision at the moment.

• 57:44Speaker 06

  So the way we build our subnet is to actually distill big models into very specific and tiny skills. So then people can use them and can buy them the way they want. And instead of running them on very large GPUs, you mentioned H100s, for instance. They could run it locally on a CPU, which unlocks then a lot of vision use cases that were so far not really profitable or not really interesting from a unit economics perspective. So we are incentivizing our miners, because now we all know the different terms around bit sensor.

• 58:22Speaker 06

  We're incentivizing them to take big models and to cut them into chunks that are working, for instance, for person detection, car detections, yada, yada, yada.

• 58:32Speaker 01

  Right.

• 58:32Speaker 00

  That's what I wanted to get into, which is the...

• 58:34Speaker 01

  Can I ask a sort of a doofus question, jumping off point, just before we go further? How does a VLM differ from an LLM, just in terms of like the training process and putting it? I think we all have a pretty good conception of what an LLM is and how you make one. How do you train up a good quality VLM?

• 58:53Speaker 06

  So most of the time, and that's a bit of my background as well. I was working for a data annotation company like a few years ago. You use human annotators to tell you, captures, for instance. Captures were the best way to start building VLMs, because you're using human beings to tell you if there's a bicycle within the image, right?

• 59:13Speaker 01

  I could spot a crosswalk like that.

• 59:14Speaker 06

  Yeah, like crosswalk now. Okay, so that's pretty much how you train the first... You just build like a very, you know, a good dataset of human annotated pictures. So that was the first step. And that was a big hurdle for us as well, because we wanted this to kind of really scale. We had to find a way to create very precise datasets that could bring also what we call ground truth. So I don't want to go too much into the details, but basically, if you want to validate something in vision... You actually need to know if miners are going to produce quality data as well. So you need to find a way to automate this process of annotating content.

• 59:59Speaker 06

  And that's a bit technical, so I can probably answer to more questions there. But just to your first point, you need to collect a lot of data. So the computer would know exactly how a crosswalk looks like, for instance.

• 1:00:11Speaker 00

  And this is why, even before you launched the Subnet formally, you were already getting partnerships with other companies to collect that data you needed, Max, right?

• 1:00:19Speaker 06

  Yeah, it comes from, I mean, there's two reasons around that. First one is obviously, we needed to test our approach. But now, I would say, you know, the day and age of just needing people to solve captures to get a good model, I mean, are over. What you need to know now is you need to, you need to get, basically, you need to train AI models to understand how human beings are solving problems in the real world. So you're not just extracting an annotation like, this is a bicycle, this is a crosswalk, this is a horse. You're also codifying the way someone is solving a problem in their day-to-day operations. So it goes a bit further. So you want reasoning, you want, you know, you want models to kind of say,

• 1:01:03Speaker 06

  okay, so when that happens, well, Bob did this, so it's the right way to solve it. And then I can kind of remind that.

• 1:01:10Speaker 00

  So you mentioned distillation earlier, and then also, I think you said chunking a model. So it sounds like what you're doing is taking a general VLM and then letting your miners essentially cut it down and heavily tune it. So that way, it does one thing, well, I think you mentioned people detection. But the models that the miners create are problem-specific. So if I made a model as a miner for Subnet 44 to help identify, let's just say, chickens crossing the road, to keep it nice and generic and neutral, then would I win all the emissions from that if my model stayed the best for that specific task?

• 1:01:49Speaker 06

  So for each task, each skills, we've got a winner-tex-all mechanism. So we always want to have the best model and to pay All the rewards that we're allocating to a task to one individual model. So we can also run this model through our front end, but I'll show you later on.

• 1:02:10Speaker 00

  Yeah, well, actually, that's where I wanted to get to next because we're talking about the bit tensor backend here. But the front end of your company, which is called Manico, is a pretty, and I say this with love, standard looking software service.

• 1:02:24Speaker 06

  So the non-standard part of things is that our front end is actually collecting all the skills from the subnet, so from the infrastructure, and putting it together automatically. So it's a full agentic platform that knows exactly what you're trying to achieve just from a chat with you. So you just come with a prompt. And from the prompt, this platform is going to build a full computer vision pipeline from fine tuning your model. So this is the chat from fine tuning your model to creating your computer vision pipeline and also your deployment. And you don't have to know anything about computer vision. So as you can see, we can create, we can generate the code for you. We can generate an SDK that you can plug into your own app. And also, yeah, and maybe you have a question.

• 1:03:09Speaker 06

  Sorry, I saw you're raising your hand.

• 1:03:10Speaker 00

  Yeah, I want to, I want to clarify something for folks out there who may be a little bit less familiar with this. But when you say fine tuning, that is taking the Alex Incorporated's information and giving it to the bit tensor selected best model for my task. So that way it has the base intelligence for the task that I have and it knows my company's context, right?

• 1:03:30Speaker 06

  Yeah, exactly. And in this case, for instance, I was mentioning car detection. I was mentioning, you know, person detection. This is something one of our partners created, you know, using the the alpha version of our platform. They wanted to know across all their stations, they're running, they're running gas stations. They wanted to know every time something like a car or a truck would kind of crash into a pump. And they, and they, and they started building their own custom model. And, and, and within, within a few minutes, they realized that they actually found. Something that happened, you know, across one station, one of their stations, which is a truck completely smashing the roof of a pump.

• 1:04:14Speaker 06

  Um, and I'm, and I'm, and I'm laughing, but in reality, when that happens, um, without things like Monaco and agents plugged into subnet forty four. Um, they would have to wait until someone would realize that something happened. You know, called someone at the station because most of the time in Europe that those stations are completely, you know, automated. And then the time to action would be in hours, sometimes twenty four hours. With this system, you can literally get a message on Slack, WhatsApp, whatever, in a few seconds.

• 1:04:48Speaker 00

  Yeah, Lon, this brings back our prior points about agents eventually playing a large role here, but I can absolutely see like, you could have an agent doing these for you and then just passing the information to you and kind of being the go-between.

• 1:04:58Speaker 01

  Yeah, you just get a text from your agent, like, hey, a truck just hit our roof.

• 1:05:02Speaker 06

  Yeah.

• 1:05:03Speaker 00

  Yeah, that's, that's super cool.

• 1:05:05Speaker 06

  Yeah, and I, and I mean, when, so like all businesses, they have kind of like a, you know, a timeframe to kind of file a complaint, you know, uh, to their insurance, uh, company.

• 1:05:15Speaker 01

  Sure.

• 1:05:15Speaker 06

  And if you miss that, it's like a hundred thousand dollars every time a, a truck, you know, crashes into a roof. So, uh, for them, that's, that's really cool to have access to that type of thing. Uh, but in general, just for you to know, uh, and maybe you guys have other questions. Our subnet is built for other agents to also access to those skills. And we built it in a way where we have a twin competition. We have the public track, so it's fully open source. So any of your, um, you know, uh, open claw agents would be able to use what miners would produce in open source. And then you have a private track that is going to be launched this week where, you know, Monaco is going to be, trained on the actual real customers we have, so they can, um, have access to their own, uh, skills.

• 1:05:56Speaker 00

  Does that mean that some of the winning vision models over on the Bittensor competition, will actually be different by the time they reach production on the customer scale?

• 1:06:06Speaker 06

  Yeah, but depending on, on which track they are processed through. So if they're processed on the public track, they're going to be generalized approach to computer vision problems and they're going to be open source. So you would be able to grab them and make them, you know, tweak them the way you want. On the private track, they would be tweaked based on the prompts people would, you know, write on Manico.

• 1:06:29Speaker 00

  This actually brings up a question that I had, which is who's paying for the inference? Because on the use of BitTorrent competitions and token emissions to help find the best vision models for a specific task, I'm totally with you. But when Manico serves them to a customer, you guys are handling the inference costs thereof?

• 1:06:47Speaker 06

  Yeah, so we kind of fixed that problem. We created an app that people can download on their computer and the inference is then running on their CPU.

• 1:06:56Speaker 00

  But I'm shocked that it's that efficient. What am I missing?

• 1:06:59Speaker 01

  If you had your OpenClaw agent running in like a Mac Mini, it could like run this on its own. It wouldn't need to. Wow, that's amazing.

• 1:07:06Speaker 06

  Yeah. And the reason for that is because what this kind of, let's say, you know, decompose approach allows us to move from a model like SAM3, which is like 3.4 gigabytes to a model for the for the gas station that is like 50 megabytes. Because it's because it's an expert model. So you can run it on CPU.

• 1:07:27Speaker 00

  Versus a mixture of experts model, you're just grabbing the one slice of it. Mmm.

• 1:07:31Speaker 01

  Wow, that's neat. I mean,

• 1:07:33Speaker 00

  This is what I'm glad I actually read the MOE papers.

• 1:07:35Speaker 01

  Yeah, that feels to me like a very futuristic fit. Like that's what we all need to do all the time. Like I rarely need Opus 4.6 for my problems. I need a very specialized model that my agent could use just to help me write tweets.

• 1:07:51Speaker 06

  Yeah, 100%. And I think we should build AI that is smart enough to just use the exact amount of resources you need. You need to use, you're right. Yeah.

• 1:07:59Speaker 01

  That makes so much more sense.

• 1:08:01Speaker 00

  Yeah, 50 megabytes is nothing. 50 megabytes is like, I sneeze 50 megabytes. I mean, often I'll have a Chrome tab that uses a gigabyte.

• 1:08:08Speaker 01

  Yeah, that's like a casual game is more than that. That's why.

• 1:08:11Speaker 06

  That was one of the biggest blockers in computer vision as well, because you could you could come to a client like the gas, you know, the fuel distribution company, but the minute you you tell them that they have to buy a machine, like a specific machine that they would just install on site and then they have to like buy like a HED 100 or something like that. You know, the conversation is over. You know, you can't talk to, yeah.

• 1:08:33Speaker 00

  Well, that's one thing I like about a lot of these BitTensor projects, Max, that we're talking to is that they take all this really complicated economics and tokenomics and then it's kind of abstracted behind the scenes. It almost feels like an API hook that takes away, you know, the difficulty of telephony in the case of Twilio. But in this case, it's just like, do you need a custom tuned VLM? Well, cool. Yeah, we'll do it. You don't need to know that BitTensor is behind it. And so to me, that's just so powerful. I fricking love it. But it implies demand on both sides. Clearly, you've shown that you can help make better vision models for commercial use. But how do you go about finding the customers who want to tap into it? Traditional business problem, but still applies in this case.

• 1:09:17Speaker 06

  Yeah, I mean, and also traditional, I would say traditional, I mean, not so traditional, but we do believe that. And we're building a community of enthusiasts, I would say, at the moment. We think that we would get more clients by letting people vibe code with the product. So we believe in vision vibe coding. This is one of our kind of strongest opinion on how our go to market strategy should look like. The second thing is we do have, and one of my co-founders is an expert in business sales. We also have, within our team, people that are already connected to a lot of large companies. And also, I can't talk about this right now, but we also managed to sign a very big agreement with a large corporation that is going to help us fix our distribution when it comes to large enterprises.

• 1:10:02Speaker 00

  And is that a technology company that you're partnering with? Give me, give me one little hint. Sprinkle some hints on me.

• 1:10:10Speaker 06

  Sort of.

• 1:10:13Speaker 00

  Oh, it's IBM. Okay, got it.

• 1:10:14Speaker 06

  No, no, it's not. No, it's not. No, no. I don't want to do like an announcement of an announcement. But basically, they, they help a lot of businesses, you know, implementing tech, you know, in their day to day operations.

• 1:10:25Speaker 00

  I know who it is. Lon. Max is in Paris. The company is based out of London, European company. Who's it gonna be? It's gonna be SAP. But I bet you, I bet you.

• 1:10:37Speaker 01

  No comment, no comment.

• 1:10:38Speaker 06

  I'm not going to say, I'm not going to say anything.

• 1:10:39Speaker 01

  Stop putting our guest on the spot.

• 1:10:41Speaker 00

  It's, it's fun. We're into the show.

• 1:10:44Speaker 01

  Yeah.

• 1:10:45Speaker 00

  Being a little loose is good times.

• 1:10:46Speaker 01

  Cause the show's wrapping up. That's what, that's what's happening.

• 1:10:49Speaker 00

  So, so as you bring on more partners, is this kind of the year of commercial growth for Manaco and Score and the subnet?

• 1:10:56Speaker 06

  Yeah, definitely. This is how we see our, our kind of 2026, you know, year plan rollout. We, we need to go to market quickly. We need the app to be used by a lot of people. And also we need to show that it's actually bringing more value to the whole ecosystem. So yeah, this is the, the 2026 is definitely our kind of commercial year for us. Yeah.

• 1:11:16Speaker 00

  Well, between small models, great technology and economics that I understand, I fricking love it. If you want to learn more about what Max and his team are working on, go to manaco.ai, M A N A K O dot AI. Or you can go through the various BitTensor world and look up subnet 44. Max, I think you've taught me more than anyone that I've interviewed in the last two months. So thank you very, very much.

• 1:11:37Speaker 01

  I learned a lot.

• 1:11:38Speaker 00

  And we'll have you back on when you announce that major partner.

• 1:11:42Speaker 06

  Amazing. Thanks for having me, guys.

• 1:11:43Speaker 01

  Thanks, Max.

• 1:11:44Speaker 00

  Thanks, man. Lon, this has been a real treat. I really hope people like the BitTensor focus. It's a cool new ecosystem. It's an interesting project. Jason's made a couple of bets and we're doing our standard learning as we go.

• 1:11:56Speaker 01

  I'll tell you who does love it, the BitTensor community. They have been very supportive throughout our exploration of Tau so far. And, you know, we, we love lots of enthusiasm, lots of passion from the Tau.

• 1:12:08Speaker 00

  It's gotten to the point now where I'm thinking about asking the spouse if I can take some chunk of cash, buy some Tau and stake it as a learning experiment. Very similar to how back in like.

• 1:12:18Speaker 01

  I, I said it on this very podcast. I'm, I'm, I'm putting, I'm putting one stack into Tau, I think. I think I'm, I'm dipping my, dipping my toe in. I'm wetting my beak a little bit here.

• 1:12:28Speaker 00

  A stack of hundreds, a stack of tens or a stack of ones?

• 1:12:31Speaker 01

  A stack is 10K, Alex. That's, that's, that's how, that's how we talk on the street. That's, that's our, that's our.

• 1:12:38Speaker 00

  I'm from the mean streets of rural Oregon where we didn't have stacks.

• 1:12:40Speaker 01

  Oh, okay.

• 1:12:41Speaker 00

  Anyways, Lon, an absolute treat as always, guys. Twist is back on Friday. My name is Alex, at Alex on Twitter. He's Lon Harris, at Lonz on Twitter. We think you're fantastic. Thanks for hanging out and we'll see you next time.

• 1:12:52Speaker 01

  Bye bye.