qBitTensor Labs Live — September 4, 2025

Speeding Up Peaked Circuit Generation

We kicked off the first-ever qBitTensor Labs Live with a deep dive into the performance work happening on subnet 63. Our quantum team collaborated with Yu-Shen Zhang, one of the original authors of the peaked circuits paper, on a stitching approach that would break large circuits into smaller sub-circuits and reassemble them. While early results showed a conservatively estimated 10x speed-up, we discovered a vulnerability the day before launch: the localized entanglement made the stitched circuits too easy to simulate. That approach remains under investigation. In the meantime, Ryan and the engineering team delivered meaningful gains through classical optimizations -- reducing floating-point precision from complex128 to complex64, optimizing CPU-to-GPU memory transfer, and filtering out unlucky seeds during generation. These changes brought reliable 39-qubit peaked circuit generation into reach.

Subnet Stability and Lessons from Bittensor

Operating a Bittensor subnet has been a crash course in decentralized systems. Varying validator hardware, defensive programming against novel exploit vectors, and the sheer pace of the ecosystem have all presented challenges unlike anything in traditional enterprise development. We discussed how the lessons learned on subnet 63 are being directly applied to the architecture of subnet 48 and Open Quantum, with the goal of launching a more robust and resilient system from day one.

The Hidden Stabilizer Solution: Pauli Propagation

The hidden stabilizer challenge was solved, and Will walked through the winning approach. A single miner submitted a solution based on Pauli propagation -- a technique that preserves a specific invariant by "massaging" Pauli strings through the circuit using identity insertions. It proved to be sub-second for 50-qubit circuits, significantly faster than the more common de-obfuscation approach. We evaluated solutions on effectiveness, depth of mathematical understanding, elegance, and uniqueness of the idea. The Pauli propagation solution scored highly on all four dimensions, and a technical paper was published in collaboration with the winning miner.

Open Quantum UI Preview

We gave the community its first look at the Open Quantum user interface. The demo showed the job submission flow: selecting a QPU provider and backend, uploading a QASM file, configuring parameters, and tracking jobs through to result download. Free access will be available for jobs submitted through the public Bittensor network, with private plans for users who need IP protection. We also previewed the wallet-linking feature, which ties into a credit system where stakers on subnets 48 and 63 receive monthly credits for priority access, increased shot counts, and private execution.

The Next Challenge: Shor's Algorithm

The upcoming challenge centers on executing one iteration of Shor's algorithm. Validators will generate Shor circuits and miners must faithfully execute them, returning sample sets that demonstrate correct computation of the period-finding value r. Importantly, miners do not need to be experts in Shor's algorithm itself -- the challenge is about reliable quantum circuit execution. Before handing the code to the Bittensor integration team, we ran extensive internal red-teaming, discovering and patching multiple non-quantum attack vectors.

Innovation, Compute, and the Road Ahead

We closed with a strategic discussion about the distinct but complementary roles of subnet 63 (innovation) and subnet 48 (compute). The vision is for Open Quantum to attract the global quantum computing community with free access to quantum resources, then funnel those users toward innovation challenges funded in US dollars. Innovations from subnet 63 -- better simulation techniques, quantum error correction advances, more efficient algorithms -- get deployed as backends on subnet 48. The long-term roadmap moves from provable accuracy (today) through synthesized real-world problems to open quantum innovation, with a focus on open-source submissions, attracting academic talent, and making Bittensor invisible to the end user.

But yeah, it's cool to know that the tech is working all right. So we'll be quiet for a minute and then we'll roll.

but yeah, welcome to the first, the first ever Cubit tensor labs live. We're super excited to be here. super excited to have you all here. we're not going to be checking all of the chats, as we're going with this, but you know, would welcome you guys, you know, on Riverside chat or anywhere else to post things up. And if we don't see them while we're going, we can always hit them up later. All right. So I came from like, the corporate world and like my entire career, I always wanted to do stuff like this, where you're just like,

have random meetings where the developers can say whatever they're working on and whatever points of view they have, with like users and investors and stuff like that. like the lawyers always say, you can't possibly do that because you're to make a bunch of forward looking statements and you're going to get sued and it's going to be terrible. So in an abundance of caution, I told, grok to cover my ass.

And that's what Grok told me. I did review it to make sure that there was no references to Mecha Hitler or, you know, anything like that. But, but I'll give you my, my read on this, which is basically, these are all ideas. We're going to be talking about a lot of stuff. That's sort of like the way we see it right now. And a lot of that stuff is going to turn out to be right. And a lot of that stuff's going to turn out to be wrong. And so our aim, if everything works right, is to do these things.

pretty often every couple of weeks maybe. And, you know, our ideas might change over time. It's definitely not intended to be investment or legal advice. So, you know, you guys can listen in, do whatever you guys want to do with the information, but we're not, if we make any statements about any kind of like alpha price or anything like that, that's, we're not qualified to say those types of things. So don't listen to us. And we would appreciate it if you didn't weaponize this material or use it out of context.

beginnings with 63 being like dipping our toes in the water and we wanted to dive in head first. And I sort of had this like grand vision of like, we're going to be perfectly diving swiftly into, into the whole qubit ecosystem. But it kind of felt a little bit more like this because, like, you know, we, on one hand we got sort of like all of this praise for, you know, a pretty solid novelty search, but then on the other hand, like, you know, a lot of people.

we're sort of surprised by big things. The charts didn't necessarily, you know, show to the moon green bars and stuff like that. And so we sort of learned some lessons and the biggest lesson I learned is avoid big surprises for people. And so the aim of this is really to like, for those that want to know stuff, we want to give little incremental updates as we go, as we continue to evolve our thinking, we want to bring you with us. That way we can have kind of communication on this as we go.

the other thing that, know, I am worried about honestly, is just to understand like the mindset of the detail investor and to understand if these things sort of become one of these weird, like sell the news kind of things. We're hoping that this actually doesn't cause that type of behavior, but instead by putting information out consistently, we're avoiding that type of behavior. But yeah, I don't remember who, who posted this up or where, but.

I totally agree. It seems almost unintuitive. It seems almost unintuitive when you share good news and markets crash, So, okay, so the agenda for today is we're gonna cover, we're gonna start with like some really tactical stuff, just more like tech support type topics. We're gonna talk about generation times, stability of 63. We recently...

Also the H-stab solution got solved, so that's big news. So we're going to have Bongo Cat Feynman will take us through kind of what that solution looked like and just a little bit about like the broader market of miners and how they all reacted to that. And a quick update on the next challenge, which is the Shorahs oriented challenge. And then we're to focus on some things I probably should have called that strategy, but it's more like community. So all throughout the last two weeks since novelty search,

Like my whole team has been sending me like screenshots of what y'all are saying in different parts of the community, which is like spread out across lots of different tools. And so, you know, the big, the big sort of messages that we've heard about in the community recently have been sort of like, what is the value of innovation in a world where 48 exists for compute and what is the interplay between 48 and 63? So we'll spend some time talking about that too. Sounds good. Let's dig in. So,

Tech support. First, we're talk about generation times and you guys all know, you all know Bongo Cat Feynman, Will, so Will, I'll let you talk about kind of our first quantum approach at speeding up peak circuits.

a lot of optimism that we had that this would work out and it was showing some great numbers. for quantum reasons, I don't know if people have questions, we can get into exactly the quantumness. for quantum reasons, it was turning out to be too easy, right? If you can advance, please. Yeah, so like this sort of stitching approach to making smaller peak circuits and

It does create a problem that looks really similar to the original problems, because the entanglement is sort of localized, it ends up being really simulatable, right?

And then also, you know, one thing that we noticed is that there's, you know, certain, there's a, there's a percentage of, of what you call, I guess, lucky seeds. There's a percentage of what you call unlucky seeds. And there's like in the middle where things are sort of average. And what we try to do is avoid generating circuits based upon those unlucky seeds. And that way we could improve the speed of the generation of that circuit.

And then there's also things that we haven't exactly pulled the trigger on yet to even optimize more, know, seeing if we could get miners to use multi GPU. You know, one of the big challenges that, you know, talk a little bit about later is that, you know, varying hardware between different validators could produce different results and then different speeds. And then, you know, how do we kind of balance the weights between all of the

validators for the miners out there. So I think what we have has done does great up to 39 qubits and we're trying to get past that as soon as possible in order to keep the challenge going as opposed to kind of topping out at a certain value.

what the validators want to share with us and what they're willing to run. you know, also on top of that is defensive programming. We're being attacked and exploited in different ways than, you know, kind of the standard suite of exploits and things that come out for the cloud. you know, learning to be better at that and to provide more stability from that perspective. And then also, you know, using what we've learned on 63 and apply it to 48.

breakneck fast, it's insane. Like, I mean, there's so many people doing so much moving so fast, like every day, new ideas are coming out. People like miners are weirdly working together. You know, like I didn't expect that. I thought they'd be like competing against each other. Like, but you know, you had to move fast in bit tensor, which has been a lot to, a lot to like, you know, get used to.

Yeah. It's good and bad. It's like, you you get a lot of good feedback from the community, you know, about new ideas and kind of new paths to go down. But at the same time, it's, you there's a lot of expectation, I guess. And, you know, being able to keep things stable and perform and, you know, keep continually producing challenges and, you know, it's a lot to manage. So.

just more natively built in. Grafana also wanted to hit on and, know, Ryan, feel free to jump in if you have other stuff to add, but we're super excited about it. took some fresh screenshots this morning. Things are, you know, looking pretty solid. You know, whether you're looking at validator health or circuit generation or sort of like minor weight alignment, like in my opinion, things are looking pretty healthy. Also just as like an interesting thing.

You can see people are definitely reliably requesting 39 qubit peak circuits now, which is pretty insane. we haven't released this to the public yet. And the key reason for that is, we sort of have seen how people go deep on the data right away. And we want to make sure that all of the data that we're reporting is as accurate as possible before we get the scrutiny of the public world on it. also we do have like a little issue that we're trying to figure out where.

the if a validator restarts, there's like a couple hour gap in data. And yeah, that feels sort of like too much to be be gapping out at.

Cool, yeah, so I'm gonna share something with you. Right now we have like multiple different streams going on Open Quantum. There's like the subnet development, there's the traditional centralized full stack development APIs and database and the connectivity of those things. There's front end development to spec out what the user interfaces and implement it. And all of those pieces like need to be connected. So.

What I'm showing you guys right now is not a fully working system. This is the front end, which needs to be connected into those other ends. But I thought we'd show early because seeing something really codifies what it is. So Open Quantum is going to be the website that users go to, that non-bit tensor users go to. You can see that on the website, you can just click Create a Job. You can go in and pick one of the

QPU providers that is available. Pick one of the backends that is available. You can go in and just grab a chasm file, which is like the quantum circuit to be executed. In this case, this is a VQE.

simulation, collect a little bit of metadata from people, and each quantum computer will have different configuration parameters that you can set for it. You can submit the job. You can see that we're going to actually provide free access as long as people are comfortable submitting their jobs through this public network, meaning that it goes completely through the BitTensor ecosystem. We also will include, you know,

That's cool. It's like neat eye candy and stuff like that. And it tells you kind of what we do, but it is probably not the way that anybody's ever really going to use it. Everybody's just going to pip install an SDK and integrate it into the frameworks that they're working with. So like today, people develop software. They run it on simulators. They'll just install load in their license keys, which they'll have through the system and just run the code.

The other thing that I do want to call out, cause it's going to be super relevant later is that for the insiders who are wanting to support innovation, you can actually add a wallet and we'll go through and validate your wallet and add it to your list of linked wallets. And whether you're staking on 63 or 48, you'll actually get this well, and this is, again, what our intention is right now. I think this one's pretty solid, but again, again, everything's kind of fluid.

But the intention is to essentially have a dividend in the form of credits. So if people stake on 63 and support innovation, or to a different extent, if they stake on 48 to support compute, they end up getting credits on a monthly basis that they can then apply to either switch from public to private or to jump in line to reduce wait times or to increase the number of shots, et cetera. So in a way you're sort of giving the community to either

buy credits with a credit card on the fly or to stake and depending on the duration that people have staked for, they'll get those credits. Cool, so maybe that's enough of that for now. But yeah, we're super excited about it. Which parts of those features end up in the initial launch, which parts of the features end up in subsequent launches is a little bit up in the air.

finishing up a lot of database design and open quantum specific design tasks and then going to be hitting hard into development over the next couple weeks here.

You know, we need validators running really thin web servers and running rest APIs and we need to have a full stack open quantum, which is like, you know, front end API is backend and we need to have the Python SDKs to connect into the quantum frameworks. So it's really like, you know, it's not like a like twice. It's not like even like twice as hard. It's like, you know, four times as hard, but I've been super impressed with what you and the team have been doing around like documenting artifacts, unit testing.

you know, bringing in a lot of like stuff to ensure stability because that's going to be really important as we go with this.

All right, moving on then to HDAB. Will, I was super excited about the publication that you collaborated with a minor on here. I'll, you know, again, tell me when you want me to progress slides, but I'll hand it over to you to chat about this.

We are dealing with quantum stuff, but there's really not any sort built-in assurance that there are a lot of proper quantum people who would be able to go through and sort of validate whether a solution is worth sort of, you know, breaking the pool out to or not. So I don't mean to call out tensorized. I don't know if you're here or not specifically, but just sort of to highlight the general sentiment about how we

we go and choose which the best solution is. Currently, the solution, which I think is really the only feasible solution right now, is that miners would go and sort of DM their solutions to us with a full technical explanation. This seems to be the most reliable solution, at least right now. We may revisit this in the future. Not a super hard promise or anything, but maybe once everything, once people have gotten more up to speed with all the quantumness that

that goes on in the background with all these problems. But yeah, for now, I think the only real workable solution is to have this sort of DM structure.

That's kind of the standard that we want to think about here. And so there's sort of like a dual version of the entire problem of operating on the subnet itself, which is like, have to be able to see this and verify whether it is publishable. at some point, everything has to pass through us because we're kind of the ones that would, in principle, make this wider publishing happen. If that makes sense.

of this viewpoint, right? So it looked from sort of a broad survey of what was going on in the subnet that there were sort of two solutions. There was this one, polypropagation, which is only one minor came up with and ended up being the winning one. But it seemed like most other people were doing de-affiscation, which is not a real technical term, but that's what people are calling it, which is where you take the H-type circuit that they get and you convert it back to sort of the original Clifford form.

And that allows you to do fast simulation to get the stabilizers out quickly. Yeah, so anyway, just to give a little bit about polypropagation, the idea is to take a poly string, which I'm calling p here, and

go through and preserve this particular invariant. And it's like the term that people usually talk about is massaging equations, right? So you massage the equation enough until you get to a form where things are workable. So the idea here is to massage this leftmost equation on the top left of this math here. If we have a circuit C and a polystring P,

then what we can do is insert identities, C dagger C and C dagger C on both sides. And that ends up giving you sort of another Pauley string, the CPC dagger here that represents the stabilizer of the output circuit or of the output state of the circuit. And the thing that allows Pauley propagation to work is this deeper understanding of what goes on when you try to multiply Pauley strings with non-Clypher gates.

variance and what solutions can look like. But broadly, I'm going to say that solutions, in order to win, should get points for effectiveness. There was some note about de-effuscation being pretty fast. When we were testing the winning solution, we found that it was actually faster than at least what's being claimed here. There were not a lot of people who submitted thorough solutions. That is to say, source code plus

plus a full technical write-up like the winning solution did. But we found that polypropagation was actually a lot faster than several seconds. It was sub seconds for 50 qubits. So we found that it was more effective. Let's see. The other point. Yeah. Yeah, so it was pretty amazing. So the other thing that solutions should get

and basically look at all the gates, maybe do some merging in a circuit, and then basically just test all the available Clifford gates in the space to see which one fits in. And so to me, that's not very elegant, because what happens if we go up to three Clifford gates or three qubit gates in the obfuscated circuit? And this is an important point, because the

Number of Clifford gates grows double exponentially, approximately, in the number of qubits, actually. So the space gets very combinatorially large. And so that's sort of the I guess, obstacle to the elegance of de-affiscation that I want to note. And then finally, I think it's worth giving solutions more points for the uniqueness of the idea.

fixed that, came back again. And so we have something now that our white hat hackers are comfortable with. And so we just handed the code from the quantum team to the BitTensor team to start integrating it. And Will, if you wanna give us just two or three minutes on it, that'd be great.

such that for some a, you can raise a to the power of r and have it be equivalent to one modulo n. And you can use this r value to generate these two, to find the two pseudo primes. I guess in terms of the actual circuit that's actually run, looks something sort of the naive picture or implementation of Shores looks like this. So you have a couple of registers,

You do some operations that look like applying this modular multiplication in a quantum way. And then you have to apply some broader transformation of the inverse quantum Fourier transform to produce what ends up looking like peaks in the output of the circuit. And the reason this algorithm works, that got messed up, is that

It'll be out in advance of when we actually release the challenge itself. Yeah, anyway, that's the, I guess, quick version of this.

which is sort of around the value of innovation and revenue interplay between 48 and 63. And so, you know, what I will say is, and I just grabbed a couple random things that came in from different people on the team, screen grabbing this stuff and shooting it my way. You know, this was kind of exciting people, like what was it, if you had to only invest in one DTOW subnet for a long time, which one would it be? I was pretty blown away that quantum

stood out above these other ones. So thank you all for helping us with that. You know, I see people talk about like, Hey, the activity of miners and thinking about that or what we post and getting really excited about that stuff. And like, you know, I guess what I would say about all this stuff is like, you know, all of, all of that stuff to me feels very like instantaneous and like that sentiment is going to go up and that sentiment is going to go down. Right. like,

we'll probably say something stupid. don't know, maybe we already did on this podcast. We'll say things that are smart. We're gonna learn, we're gonna iterate. But what I'm really excited about is like the real value, not like sort of the speculative value. And so like I grabbed this post from somebody named Kai, which was also in this database we were collecting. And I saw a couple of things that I was really excited that were noticed. One is that we're cautious about over-promising.

So that definitely comes from the fact that like we come from large business where, we want to make sure that we can really deliver on something before we promise. I also was really excited about seeing that people see that quantum is sort of stuck in this intellectual elitist space and that really there's tremendous value, tremendous value in taking that innovation and making it accessible to people. And that people were seeing that subnet 63 is the innovator.

Giga brains from the labs to be competing for money. Also, I like that cause it's like a US dollars and not tile, which I think is really right on. I bet Scott Aronson wants to play. was thinking probably like probably Scott Aronson doesn't, but probably like his PhD students do. So that's maybe more of the, more of the play for it is to say, can we actually get those types of people on? And you know, we could even build a front end that lets them sign up and try things and then run on compute, but.

you know, 48 can also be that compute. And then this, think was like a shout out to ridges being awesome for removing crypto from mining, which I also thought was a really insightful thing that's in alignment with what we are also thinking about. And so if you think about like, I think a lot of people have made this comparison that quantum computes is essentially the quantum shoots. I do want people to start seeing quantum innovate as working towards becoming the quantum ridges.

And what you can, like what you sort of, might take away from something like that is. Ridges is, has very much focused on allowing outsiders in. And we've been really impressed with what the bit tensor community has been able to do, but like, we do want to get quantum people into the space also. And like, it is, it is a lot to take in for somebody from the outside. so we want to abstract away a lot of the complexities of bit tensor.

We want to talk about things in U S dollars when we're talking about these sort of innovation challenges and awards. And we need to get miners to providing their answers as open source, as opposed to proof that they solved the problem without, you know, requiring them to get to how they solve the problem. We also want to get to problems that are quantum meaningful. So right now we started with a problem set that was focused on getting the BitTensor community more in the weeds of quantum.

a lot of things, like especially some of the optimization problems that we have. People aren't even sure if those are going to have like an exponential advantage. There are some where people are highly confident it will, but if we can get to actually developing more efficient quantum algorithms and rewarding those on the platform, that's going to be of tremendous value. If we can start doing things like challenging people to push the envelope with quantum cryptography so that we can understand, you know, what is the current state of the art that people can do with quantum algorithms, that's going to be highly valuable. So

starting to expand into more meaningful problems, you know, is definitely the roadmap there. And then building from ridges, you know, one of our missions. Isn't just going to be getting, you know, highly qualified people in, but it's also starting to attract pedigree. Like one of the things I hate about quantum is you need pedigree, but like, it's actually just a truth about the market. And so yeah, that, that question about, you know, can you get the Seth Lloyd's, the Scott Aronson's it's like, I don't know if you can.

But like, if you get their PhD students or their labs in here working on this stuff, like you've crushed it. Also, can we incentivize quantum experts to eventually moderate those problems? So don't just put it on the small set of people that we have in our company, but you know, actually expand out to other people and challenge them to bring problems into that network. And so, yeah, like I hope that really articulates like,

I mean, you know, a novelty search, we talked about like this world that I want my grandkids to grow up in where like cancer has been solved. Cars just drive forever and are hardly ever have to be charged with low impact on the earth. The goods that are needed to get to the people like all in order for all of that to be true. We can't just provide access to compute. We need to be providing innovation to make the compute actually solve those problems and to improve the trajectory that the compute is on. And so like there should be no question about the value of innovation.

Okay, so then talking a little bit about the interplay. So, you we kind of talked about it. well actually, you know, the side sort of weirdly out of place now that think about it, but we already sort of talked about some of the interplay, but we'll talk about a little bit more, but just to codify that roadmap, you know, provable accuracy is where we are today. synthesized problems is where we're going, which you can imagine becoming sort of one of the challenge areas that we ask for continuous innovation on is being able to execute.

broadly generated synthesized problems. But where we definitely want to go is to quantum innovation period. that's beyond just execution. That's other very important problems that are going to accelerate quantum. It's open source, not just getting the answer. It's rewarding innovation with us dollars and not just tau or not just alpha. And it's about getting outsiders in. And so when we look at that ecosystem together and we say, well, how do they interplay?

Well, imagine a world where open quantum built on quantum compute, subnet 48 attracts essentially every user that does quantum computing in the world because they immediately have access to free computers. We generate a little bit of revenue. That's cool. A lot of people like that. And we do that by letting them execute privately, jump in line or removing some of the limitations to the jobs that we submit.

through the sale of credits, which again, you can subsidize and we incentivize that staking, we incentivize staking on 63 and 48 by generating credits based on how much you stake. I also saw some questions about like, is there a market for reselling credits? We honestly hadn't talked about that, but what we had talked about is making these credits like specifically useful in certain ways, which might mean that you have to use them.

If you want to use that free quantum computer time to solve these meaningful quantum problems, there's cash available for you and we pay you in USDs and we publish what we've paid out on open quantum to make sure that it's all transparent and above board and real time. And while bit tensor is the engine on this, these guys just, you know, look at the car, right. And they see, you know, a beautiful shiny car that they get to drive for free.

And when an innovation award is on one of these simulation tasks, right? How do we execute quantum circuits generically better? Those innovations actually get to be deployed as backends, right? You saw in the demo that you could pick which quantum computer or simulation backend you wanted to run on. You know, we can kind of continuously update the image for that, you know, that backend based on who's winning on 63.

And all those other innovations, there's the opportunity to launch those innovations as hosted services directly on open quantum. And so, well, the near term value proposition is free quantum computing. As soon as we have created undescribable value to a very needy market that is rather large, and also everybody knows everybody, so we should be able to spread like wildfire.

once we capture a share of that market, we can use that market to pull everybody into these innovation challenges and make those thrive. And also it gives us a platform to broadcast the innovations as they happen too. And so you can see like in my mind, this has always looked like, like, you know, two different parts of the same solution. And, and, know, with our original naive thinking of single subnet, we essentially had to stop one of these threads in order to do the other.

So, okay, so with that, yeah, we sort of opened with this. You know, I kind of felt like this after novelty search. I haven't looked at Discord and I haven't looked at the charts. I have no idea, but I'm hoping that we feel a little bit more like this. We didn't quite belly flop, but, you know, we'll hopefully, you know, continue to learn and evolve and keep this series going. So with that, I think we will sign off. Okay.

Thank you everybody and thank you ShorShot, AKA Ryan and Bongo Cat Fineman, AKA Will.