Watch

Transcript

Disclaimer: The transcript that follows has been generated using artificial intelligence. We strive to be as accurate as possible, but minor errors and slightly off timestamps may be present.

Jacob Steeves (00:11):

And thank God for your watch. Many of you guys are still here. So I wanted to go back and do a bit of nostalgia about the last year, showing where we were at that time. So this is what the website looked like. I had to use the Wayback Machine. We used to have the run a validator that if you had run that script back then, it wouldn’t have taken anything. This is actually from October. We were mining a new element. And none of you really understood what that meant. A couple of you had been on the previous network that had closed down in March.

(00:49):

So it had taken six months. We’d raised some cash and hired two of our devs that are on the call, Isabella and Eugene, to take us to network two. Now, there was many things that we’d implemented at that time. The raw consensus mechanism had not even been in place in the previous network. So truly, BitTensor in its current form didn’t exist until November 2nd. I pulled up this Discord shot because there was three names at this moment in time that are on this call right now that arrived all at the same time.

(01:27):

I’m not sure from where. And of course, their names were not WizzlesWithTwoTs and PinkLlama. I don’t think it was TauTentar6 at that time either. But there’s Scrim coming in on the 9th, sorry, the 15th of September, 16th of September, all in quick succession. So it was a really interesting time back then. We were getting not that many people coming into the Discord. That’s somewhat changed going forward.

(01:57):

So that was where we were. And then on November 2nd, we launched the new chain. And it’s been a hell of a journey since then. I mean, we were talking a little about that first month. Many people were up for nights on end. There was no documentation really about how to do it. It was just the Discord. And the tenacity of the early miners to figure out how to run these miners.

(02:34):

There was a difficulty of 5 million, which we ended up having to cap at 10 million because the issue was too low of a difficulty rather than too high. Good times, really good times. And as I said before, I think many of us wish we could go back in time and register twice as many miners and get those UADs, especially in the first 100. It’s someone taking the zero UAD quite sneakily from us and then losing it later, I might add.

(03:04):

So it’s been a hell of a year. And I think one of the things that we haven’t really focused on this very much, but it’s something that’s happened just naturally. We’ve had a very natural growth in our Twitter presence. And thanks to Todd Quillen, who’s on the call, we’re beginning to build an audience in this regard. And it’s something that we want to push next year because we feel we’ve built this very natural growth to our system in a community that’s not in any way forced through marketing. But it’s something that we think we’re going to push once we get the Bearchain launched next year.

(03:48):

Also, the growth of the Discord, at the time of the launch, it was only 200 people. And now we’re over 3K. So we’ve had more than a 10x growth in the size of the community this year, going from completely unknown to being talked about in Twitter spaces, as Dr. M mentioned at the beginning of this call. We’ve been honing our brand. Now, this is not the final product. We have a, I’ll talk about this a bit later in the call, but we’ve gone through branding, a process with a firm called Saffron.

(04:23):

And we have some really exciting growth in the way the website is looking and some fine-tuning of that design that we’re trying to pull out of the ethos. So this website could be gone pretty soon. So love it while it’s still here. Another thing, we launched the Playground. This was back in July. Thanks to the very hard work of Caro, it was in July that we started actually exporting knowledge from the graph, from the network. And this was just a sneak peek, really, because it’s alpha. But what we’re doing right now is building more and more technology so that we can peer into the tensor and understand it.

(05:09):

This was done, really, by one developer Caro had mentioned. And we think that as more sophisticated developers come on board, as many of the miners here purchase, sell Tau to build applications, we’ll see many more of these coming here. We made it easier for you guys. We launched our documentation thanks to Mac, who’s on the call.

(05:43):

Obviously, Mac was hired directly from the community. We just look at how well people mine and think, oh, they’re good people. So we’ll get them on the right side. That’s pretty cool. But behind the scenes, I think some of the stuff that you guys have not seen so far has been the research side of the foundation. We are a research foundation. Most of our engineers are hard at work building and testing what we can build on top of the tensor, because AI is a research technology. If you look at a lot of other companies in the space, they expend an incredible amount of money hiring researchers to work on this. So these are just some photos of what that research has looked like that none of you have seen.

(06:32):

And we worked at length to craft and fine-tune the way in which we were evaluating the miners, because at the beginning, the validators were doing a fairly bad job of detecting the size of the models that were running and how significant they were to the loss objectives. They were moving very slowly, because the network at that time, a lot of people were running contable machines. People were running on digital ocean droplets. There was hardly any GPUs on the system. And that made it difficult for us to do research, because the response times on the network were just too slow. That’s changed now. And I’ll demo that at a certain point in this talk to show you guys where we are today in terms of building the supercomputer and making it accessible to everyone who’s working now. So these are some images that were created by Taco.

(07:29):

And then there’s somewhat more that I won’t go into the details of this. But we’re looking at the way in which the validators are using the logits from the miners and producing valuable shapely scoring in an adversarial-resistant way. So these are some blueprints for what we implemented.

(07:58):

We did a lot more research investigating new types of consensus and how we could fine-tune the consensus mechanism to be more adversarially-resistant as time, as we begin to fine-tune that in the future. This is something that we’re going to be working on continually as a foundation, making sure that the consensus mechanism is truly concrete and very strong. And that requires a mathematical investigation of the technology. All of that culminated in a paper. So we pushed our first research paper to NeuralIPS. This all culminated. We have a whole bunch of plots and descriptions of what we’ve done, how we’ve changed the shapely value calculation of incentives. And that push, that technological push that we’ve done in the background has been what drives the growth in the system that we’re seeing today, the forcing computers to run on GPUs with decreased sequence lengths, with better approximations of the value that the computers are adding to the system. That’s made many of the miners here jobs harder. But the reward is that the system has become much, much more performant and valuable to any applications that are built on top of it.

(09:22):

We’ve grown substantially. At the beginning, it was a lot of contable machines in Europe. And there was strong clustering around Germany because of it, as well as the DigitalOcean data centers on the east and west coast. Now, we’ve expanded to six continents. We have Europe with America. We have South America. We have Asia and Australia.

(09:50):

So the growth in the geographical spread of the technology has been something that’s been really great for us as a team and a community to reach to people that can’t potentially enter the world of AI because of where they’re situated. Because of the way that we’ve grown the system, because of the way that we’ve been fine-tuning the mechanism, we’ve seen an increase in difficulty.

(10:29):

Our approximation is that we have 576 RTX 390s just on registration alone, continually. And that’s based on the current difficulty. So we’re pushing up into the thousands of GPUs just on the registration side. That’s the pushing into the system. And that’s driven our network up into serious activity. So we have almost reached the total UID count that we could potentially get that is active. And we think that this last little buffer here, it’s just some miners that are not setting weights and can’t detect whether or not they’re on. But this is, I would say, a really great culmination for the year, that as we have reached the end of the year, we’re reaching the limitations of this network that we created. And we now are ready to push into a higher UID count, more subnetworks, et cetera, as we push the new chain.

(11:30):

So I wanted to show you guys what we’ve built in a demo so that the power of this token that we’re mining can be seen directly. So I’m just going to close this slide and go over here. So this is just a really simple script. And what we’re doing here is just querying the network. And I want to show you guys the response times and the speed at which we’re able to talk to BitTensor and what this actually means from an engineering perspective and a research perspective. I have a 12-second timeout. That’s our limit for the network. And so we’re querying here on this example at 21 megabits per second.

(12:15):

Now, today and with this script, this is about 10 people we’ve been able to clock on some of our better machines and better experiments, hitting queries per second at 93. So we’re talking to 93 inputs per second, something akin to a 100-billion-parameter model that we can talk to at any moment in time because we’re holding the keys with Tau. If I were to take this wallet out, if I were to put in a wallet with just a few Tau on, see that the rate of success and the timeout would increase substantially. And the queries per second that we could get would drop substantially also. This is something that we could not have done one year ago. And it showcases how we’ve improved as miners, how we’ve improved as a technology company to make the neural internet incredibly performant. When I saw this, when I started to run these experiments myself, it got me very excited about the power at the tips of my fingertips holding one of the larger keys in BitTensor.

(13:24):

And I think the next year, the behind the scenes that we will let you guys get some interfacing with at some point will be working on us extracting this value and turning it into applications and turning it into tooling for people that are holding Tau. That’s really what we’re creating here. Can we talk to the network very, very profoundly? Can we get these response times? Are they producing knowledge?

(13:53):

And this is really the first example of that, I think, today. It took us a year to reach this. And so I think we should be all very proud as a community that we built a technology now that runs very strongly and produces a lot of value. So that’s where we were. That’s where we got to in a year. And there’s much more to come. I want to talk about where we’re going.

(14:19):

There’s a couple releases that are coming up in the next couple of weeks. I think this one has been called out by the community. A lot of something we want. We have an OpenTensor wallet. It’s built by us. Trust us. It integrates right into the website. It’s going to be the wallet that people use when they’re using the applications and the tools that we build as a front end for BitTensor. So this is coming in November 17, the next TGIFT. So really excited about that release. We’re also launching on December 1 a BT. Klee bot into the Discord.

(14:58):

We’re going to be able to do many of the things that you guys do with BT. Klee directly from the Discord channel. So we’ll be able to interface with the network, query peers, look at the incentive graphs, get information from miners, stake values. All that’s coming out from the Klee bot that’s launching on December 1 in the next couple of weeks. This is all sort of short-term stuff. And one of the big pushes that we have coming out in mid December is new branding for BitTensor. We worked with Saffron for the last four months, five months, to really hone the branding for BitTensor. And we want to release that on the 15th of December. Around December, I can’t show you anything, because technically it’s still behind a bit of an NDA. But on December, mid December, we’re going to have a reshaping of the look and feel of BitTensor. Really excited about that. I would say this is within this year. So this is kind of talking about things that I think we can say that are coming out in this cycle.

(16:13):

But the big one for us is the Parachain launch. We have picked a date. We funded this launch for the auction. We’re highly confident that we’ll be able to get this. And our next network is coming out on January 10. So that’s our new network that’s going to be called Fini. And it will have subnetworks, as well as interaction between the greater Polkadot ecosystem.

(16:41):

So this is BitTensor 2.0 for us, where we are beginning to relinquish a lot of control over the running of the chain. Growing up, letting it run outside and be independent of ourselves, mostly the foundation. So this will allow things like the DAO, the beginning of you guys taking more and more control of the system, allowing voting mechanisms on the chain, allowing people to stake into a shared validator and make returns, even if they don’t want to run Node. And that’s how we’re going to fund a lot of the development. That’s Q2. That’s Q3.

(17:24):

I think the big one, obviously, for Parachains will be the connection to DEXs. There’s a variety of different DEXs out there that people will be able to use, building smart contracts on different chains and using TAO as the reserve currency in different forms of contracts. It’s going to be really exciting to see what the development community can build here, as well as the foundation.

(17:53):

Subnetworks. I think this is going to be later than Q2, or maybe late Q2, where we’re pushing to a second network. We want to work on images. This is obviously a stability diffusion image. This is stable diffusion, a generation that we should be able to create. On BitTensor, we should be able to create it through our Discord channel by connecting into BitTensor network. And this is going to allow us to really peer into the workings of the chain. So we’re really, really, really looking forward to this innovation. The reason we stuck to text for this long is because we needed that time as a foundation to really get a hold on what these hyperparameters were doing, and how we could tune them as a foundation to make the network really click. And as I showed you, it’s beginning to really walk.

(18:50):

Iron rule of nature is, you get what you were worth for it. You want ants to come, you put sugar on the floor. We, I think, more than any project in the entire ecosystem, are playing with incentives. Accurately, at a high dimension. And that was what the original paper for BitTensor outlined as its goal, to use incentives. Decentralization is not our killer app. It’s using this technology of incentive to gather compute, to organize ourselves in a multiplayer game to create this digital commodity.

(19:41):

I think we’ve created the first true incentivized neural anything, truly incentivized artificial intelligence. And we’ve followed through with that over this last year to really make it a reality. So it’s very exciting. And I think that’s really the closing remarks. We’ve showcased what we can do in just one year, from 5 million difficulty and the most stressful months of our lives, staying up late, to a technology which is alive, thriving, working, truly.

(20:16):

And I’m so glad that everyone here has come to this talk. So glad that everyone is still with us on this journey and into the future.

Video Description

SUBSCRIBE THE BITTENSOR HUB FOR EVERYTHING BITTENSOR!

This Clip was recorded on Bittensor’s Youtube Channel on November 8, 2022

Let’s celebrate one year into Bittensor! The past. The present. The Future!

Subscribe to the The Bittensor Hub for everything Bittensor!

Bittensor is an open-source protocol that powers a scalable, globally-distributed, decentralized neural network. The system is designed to incentivize the production of artificial intelligence by training models within a distributed infrastructure and rewarding insight gained through data with a custom digital currency.

Discord: https://discord.gg/Qv3fxVaXyE Website: https://docs.bittensor.com/ Whitepaper: https://drive.google.com/file/d/1VnsobL6lIAAqcA1_Tbm8AYIQscfJV4KU/view Network Map: https://bittensor-explorer-staging.netlify.app/

Socials: Bittensor: https://twitter.com/bittensor_ Ala Shaabana: https://twitter.com/shibshib89 Const: https://twitter.com/unconst1

DONATIONS: 5DHYPY9P8VBSpp6Q8m8hHrq4ME5UTiXfXVFbH32MNV4JcSxH

HASHTAGS: #BITTENSORTAO #BITTENSORMINING #BITTENSORCRYPTO #BITTENSORNETWORK #AI #artificialintelligence

Sponsors

Podscript is a personal project to make podcast transcripts available to everyone for free. Please support this project by following us on Twitter.