# Proof of Talk 2025 keynote: "Satanic to Sublime: Aligning Human Potential Through Open-Ownership Protocols"

- Title: Satanic to Sublime: Aligning Human Potential Through Open-Ownership Protocols | Proof of Talk 2025
- Channel: Proof Of Talk
- Event and date: Proof of Talk 2025, main stage, the Louvre, Paris, 10 June 2025
- Length: 19:25 (per the caption service; the host introduction is about 1:25 of it, Jacob's talk about 15:30, then a 2-minute Q&A)
- URL: https://www.youtube.com/watch?v=3FDo1Cpnt74
- Jacob's start offset: about 1:30 (estimated from word count; the host speaks about 210 words first, then applause). The assignment note said "around 0:44"; that could not be confirmed because no timing data was obtained.
- Method: OBTAINED 20 Sep 2026. YouTube's auto-caption track (English, ASR) for the full video, pulled through the third-party transcript API at kome.ai (`POST https://kome.ai/api/transcript`, body `{"video_id": <url>, "format": false}`), which returned the complete text (about 3,000 words, `hasMore: false`, `length: 19m 25s`). Every direct route failed again: yt-dlp with a fresh bgutil PO-token server and every player client (android_vr, ios, mweb, tv, web_safari, web_embedded); yt-dlp and youtube-transcript-api through an SSH SOCKS tunnel via the validator box (same "sign in to confirm you're not a bot" wall); youtube-transcript-api direct; all live Piped and Invidious instances (dead, Cloudflare, or YouTube-blocked); tactiq (App Check token), notegpt (login expired), youtubetranscript.com and savesubs (Cloudflare), downsub, lemnoslife. No audio was obtained, so this is the YouTube ASR text, not a Whisper pass.
- Timestamps: ESTIMATED. The caption service strips timing. Each block's time is computed from its word position against the 19:25 length (about 2.6 words per second, with 5 s of music at the start and 10 s of applause before Jacob speaks). Expect an error of up to about 30 seconds anywhere; the two anchors that can be checked are the 57-second X clip (starts at "Bitcoin is 450 times bigger", placed here at ≈3:12) and the video end (19:25). To get exact timestamps, download the video on a home connection and drop it in `media/`; the Whisper pipeline from the other transcripts will then realign this text.
- Text: verbatim YouTube auto-captions for Jacob and the one audience question, including fillers and repeats. Corrections of obvious caption errors, applied throughout: Steves → Steeves; Betensor / Potensor / Bitensor / Bit Tensor → Bittensor; louv → Louvre; Cerebrus → Cerebras; shoots / Shootes → Chutes; Gradians → Gradients; three gorgeous dam → Three Gorges Dam; AS6 / F FPGs → ASICs / FPGAs; DTO → dTAO; news research → Nous Research; Afine → Affine; "8 billion premium run" → "8 billion [parameter?] run" (uncertain); two censored words restored as [expletive]. Left as heard: "Subnet 20" at ≈16:33 (he means 120; the slide had been taken away), "lobby" at ≈16:33 (a person's name the captions did not catch), "Celium" (the compute subnet).

## Transcript

**[≈0:05] HOST (paraphrased):** The MC thanks the room for its energy, points people to the second stage inside the Louvre itself (art, paintings, sculptures), then introduces Jacob: co-founder of Bittensor, but better described as co-creator of the Bittensor community. She notes that many subnet developers are meeting each other in person for the first time today, and asks for a big round of applause.

*[≈1:27] (applause, walk-on music)*

**[≈1:32] JACOB:** Hey everyone. Thank you. Love you all. What an incredible venue. What an incredible venue. Um the introduction is correct. Um I always like to say that Bittensor is a federation. Um, we're trying to create a thousand-year federation. So, I'm not the CEO. I'm just one of the people here that really believes in this technology. And actually, there's a whole bunch of people in this room that should be up on the stage. I see JJ. I see Cole.

**[≈2:03] JACOB:** I see Barry. They're going to come up later. Um, we're we're a movement. And we we started this movement a long time ago. And one of the first things we did was write it down. We wrote down our mission to unify Bitcoin and AI. And at the time, we really didn't even really know what that meant. Um, we've learned a lot. And I'm and I'm actually going to talk a bit about what we've learned about what the unification of Bitcoin and AI really is and what it means for the world.

**[≈2:39] JACOB:** In this um paper that's on the bittensor.com, we talk about how what Bitcoin really showed was that a digital commodity market, a pure market could create anything almost anything in case of Bitcoin is hashes with such efficiency and power that we could compete with nation states. We're talking about a technology that built the largest supercomputer in the world. And but what is that thing? It's a pure market where miners compete to produce hashes and add them to the network to secure the network.

**[≈3:12] JACOB:** Just to give you guys an example of uh or an idea of how amazing that is. Bitcoin is 450 times bigger than Google's cloud, IBM cloud, Cerebras cloud, AWS, and Nvidia. 450 times when measured in terms of hashes than all of those data centers combined. And if you were to sell Bitcoin at the price of Bitcoin every single day since its inception in 2009, you would have spent a thousand times less money to build that data center than those corporations.

**[≈3:44] JACOB:** So if we can get this right, if we can get the power of digital currency markets to focus on artificial intelligence, we can be a thousandfold larger than any centralized organization in the world. And that's a big deal and that's really scary. So we've been focused on how to do that for a while now. I just want to caveat this. I'm going to I'm going to talk through a couple of the projects that are currently on on Bittensor right now, but none of them are mine.

**[≈4:18] JACOB:** I didn't make any of them. So, I'm really just kind of showcasing other people's work. Um, amazing work. Forgive me for for stealing that if this is one of your one of your projects, but here are some things that we've done. Um, we've managed to commoditize compute. A lot of a lot of companies in the space have been trying to do this for a while. We've been doing it very well. Um, we we get the cheapest B200s, we get the cheapest H200s when when you measure in terms of emission to the actual machine that you can SSH into and use, we can we're driving down the cost of compute and it's way way way way more powerful than what you can achieve with a centralized organization.

**[≈5:08] JACOB:** So when this really gets used, the use that it deserves, we will have in Bittensor the biggest data center bigger than OpenAI. And this is just one subnet. This is uh there's actually a couple subnets that are doing this, but I'm I'm just going to showcase one of them. We can do inference. Um this is uh Chutes. This is the number of of actual organic queries into the network over the last month. Um we're we're serving a trillion tokens.

**[≈5:40] JACOB:** In terms of the speed at which this inference usage is growing, we're growing faster than Google did in terms of uh their inferences of their machine learning models, which is kind of incredible, right? We're growing faster than Google over one year. And we're driving down the price of inference. We're commoditizing inference. We're doing decentralized training. This is Templar. We can take compute and inference and we can turn that into a training run. So this is um this is Templar.

**[≈6:11] JACOB:** This is the latest 8 billion [parameter?] run. Um I just want to just take a moment to uh try to express like the significance of building a market where the market pushes participants to lower the loss. In order to do that, you need to make the the actual convergence of a machine learning model a market equilibrium. It's not like, oh, we're just training with their friends across the web. Um, you're training with individuals that are maybe they're using GPUs next to the Three Gorges Dam and maybe their entire goal, in fact, we know that their entire goal is to destroy your training run.

**[≈6:53] JACOB:** We know that they don't give a [expletive] about the loss going down, but we compel them to do that. And when we when we create these permissionless systems, they grow incredibly fast as I was saying. And this is really like the the pinnacle of what Bittensor was trying to achieve. And we've we're proving that we can achieve it at better and better every day. I'll talk a little bit about how how how long it took to get to that point in in a second.

**[≈7:26] JACOB:** Um this is Gradients. Gradients is is commoditizing the training of machine learning models. Gradients makes a market game for the ability for miners to lower the loss on any machine learning model that you you ask them to. It's a functioning product and it just re recently released a paper showing that we're better better than hugging face. We're better than Google AutoML. We're we're better than together AI in AutoML. We don't even know what the miners are doing here, which is kind of incredible to think that we can actually take the concept of Bitcoin, which is commoditizing hashes, something very well defined, and turn that into the commodit commoditization of training a machine learning model.

**[≈8:11] JACOB:** So, it's been a long way getting to this point, and we've really um excuse my language. We've [expletive] up a lot. Um the first couple years of Bittensor were actually just us trying to wield this incredibly powerful machine. I would I would say that the first era of Bittensor going back a little bit here first era of Bittensor was the missionary era where people were trying to take this vision and make it real. And honestly, we were using CPU machines and and people were making exorbitant amounts of money um off doing nothing.

**[≈8:48] JACOB:** And and honestly, it's kind of embarrassing how how there's many people in the room. There's like a minor here that was mining when he was 13. It wasn't it wasn't very industrial. Um and we learned a lot of things. We couldn't just point a market at a machine over an over an Axon connection, which is over a an RPC connection, and say, "We're going to pay you until you create AGI." That doesn't work actually. Um, it's too ill-defined.

**[≈9:19] JACOB:** And so, over time, what we've done is gotten much better at defining commodities. And every time, every time we coalesce these commodities into being better, more well-directed, they get more powerful and we get closer to unlocking the supercomputer power of Bitcoin. That was the missionary era. I think after that, we had a bit of like a gamer era where people showed up and they were video game players and they said, "Oh my god, we're going to just crush." And they they did crush.

**[≈9:52] JACOB:** Uh they went from um you know, similar to the way in which Bitcoin evolved early early days. It was missionaries. It was people that just believed in Bitcoin. They wanted to see the Fed destroyed. And so they were mining on their CPUs. And then shortly after that, we have ASICs. We have FPGAs. Um, and we we have the development of more of a competitive game. And then finally, we have people like Barry on the stage coming next who industrialized Bitcoin.

**[≈10:24] JACOB:** They said, "Okay, this is no longer a game. this is something that's very serious and we're going to get very clear about the economics and the margins that we can make from these digital commodities. And it was that it was at that point when you really see the exponential rise in the power of Bitcoin. I think that that's where we are right now too. We're at a point where Bittensor has stopped being a game. It's no longer gamers.

**[≈10:56] JACOB:** It's no longer a faux competitiveness. It is industrial scale. What you're looking at here is industry where these teams are measuring pure margin gains on their ability to produce and extract as much value out of the commodities that they can. And so this is why Bittensor today actually looks like this. Here we have um I this is a bit of a mock. I couldn't figure it out entirely. There's some things in Bittensor we don't know for sure, but these lines are everywhere.

**[≈11:28] JACOB:** I know a Bittensor subnet is mining another Bittensor subnet as well as a subnet where they're getting revenue from an external source commoditizing their their themselves. First one obviously is is Chutes. Chutes is in in talks with Nous Research about mining their sub their their new network. I think that's positive. I think that the future of this industry is an interconnected web of digital well-defined digital commodities that push the prices and the efficiencies to their limits and scale beyond anything we've seen in the classic corporate enterprise.

**[≈12:03] JACOB:** A lot of these subnets are mining each other because I really believe the best market for intelligence is intelligence itself. I used to say the best measure of intelligence is me measure is intelligence itself but I think the best market is and that's where the connective tissue of this whole ecosystem really shines. Where is the best place to take a computer that's not in the data center but run by a um some homegrown individual in in their basement, right?

**[≈12:35] JACOB:** We're talking just a single like 8x H200 box. Where's the best market for that? Well, the best market for that is actually decentralized training. And that's this is really interesting. So, we need it tells us that we need to build this ecosystem from the from the bottom up. If we're going to make decentralized artificial intelligence, which is what we've been on a mission to do for a long time, we need this to happen. We need to have co-mining.

**[≈13:06] JACOB:** We need to have integration of the different products on Bittensor. And this is where really the the system begins to shine. So this is all underneath dTAO. I don't know if you guys know much about dynamic TAO but it's a market for the these individual commodities and this is really accelerated and exponentiated the speed at which people are moving towards this level of efficiency. Okay, it kind of looks like this. This is what Bittensor is building.

**[≈13:36] JACOB:** We're we're building the full stack. Um, and I got this idea when I was working at Google because I was so fascinated by that corporation. They are a full stack company. They are sub teams on sub teams working on small aspects of the larger problem from compute all the way up to intelligence. We're refining these digital commodities and as we go along this path, we're learning how to well define them. We're getting very very specific about what what exactly is that thing that you're doing when you create intelligence, right?

**[≈14:12] JACOB:** Is it just compute plus data? Well, we are figuring that out and and we think it is actually. We think it actually in many cases is just compute plus data but also inference. And so over time I suspect that the way in which we will reach pinnacle pinnacle pinnacle artificial intelligence is as an interleaved connective tissue of refined intelligence all the way up to God. Um so that's where we are today. And I want to note that I've been speaking a lot about something that I'm not really running at all.

**[≈14:48] JACOB:** None of those subnets um I run. Um, none of this structure I organize it all happens organically. It's sort of out of my control now. and and um that's a great thing as as a leader and uh of of this project for for many years. As someone that's been putting my lifeblood into making sure that it's well organized, I consider a success that I'm not needed anymore. And I had a bit of a hard time writing this presentation because I was like, why am I on stage?

**[≈15:23] JACOB:** What I'm just a missionary, but I'm not really doing anything. What's I'm just sitting there. So, um I registered a subnet uh about 35 minutes ago. Um um it's subnet 120. Um it's called Affine. Uh and and I'm back in the trenches, guys. Um I I I I want to I I want to lead from the back. I want to show I want to showcase what we can do on Bittensor. And my vision um that I just presented here on the stage is what I'm going to make this this subnet about.

**[≈15:59] JACOB:** It's about interlocking um the different commodities on Bittensor and focusing on I think the pinnacle element of RL. Uh so our earlier design the design that we're using right now is is using compute from Celium. Um we're using the ability to train machine learning models on gradients and we're actually serving those models that we're training on Chutes. We're co-mining and connecting Bittensor together to focus on these higher order commodities, these pinnacle elements. So that's why I named my my talks uh satanic to sublime.

**[≈16:33] JACOB:** Uh satanic obviously being compute um and sublime being uh the the pinnacle of of intelligence. So anyways guys, yeah, it's called Oh, they took the slide away, but it's called Affine and it's on Subnet 20. So anyways, that's thanks for for coming everyone. It's a pleasure to see everyone's faces and uh love you all. Cool. Um, I don't know if I have more time or if Barry and and uh lobby are coming up. Either way, I can answer questions perhaps if people if people want to know.

**[≈17:07] JACOB:** Um, you know, technically you can just invest right now. Um, I'm pitching you. Um, it's obviously no premine, so I don't have any allocation. Is there any questions in the crowd?

**[≈17:20] AUDIENCE (Atlas):** Yeah. Uh, could you be more specific on the architecture? I'm quite interested.

**[≈17:24] JACOB:** Right. That's a great question. Let's go back. I think I got kicked out. Um what we've seen over time is that the original design of Bittensor was run a node, we'll query you and we'll measure how close you are to AGI now. And we'll pay you if you're Oh, here it is. We'll pay you if you're closer to AGI. So, it turns out that that doesn't really work that well. Um and you have to be very very specific about the way you have to constrain the problem.

**[≈17:59] JACOB:** You have to constrain the what what the degrees of freedom of the miners are. So this is where Bittensor really shines. We can use Bittensor's ability to train models as the basis for another incentive mechanism. We can use Bittensor's ability to host GPUs as the resource that those training runs use and we can define the final landing place of a machine learning model as a place where that model can be inferenced. So what this means is that like in a second I can build a chat interface with high class sorry high fidelity inference speeds.

**[≈18:37] JACOB:** I can define my my incentive mechanism as just train a model by giving me the data and the the uh the Python script and you can't do anything else. So this doesn't mean this means that machine machine engineers don't they don't need to worry about the architecture. Maybe that's predefined. They don't need to worry about how they're going to produce a an inference server. They don't need to worry about the productivity of their inference server.

**[≈19:07] JACOB:** That's all taken care of by another commodity on the network. So, we're stitching together commodities. Does that answer your question, Atlas? Cool. Great. Well, if there aren't any more, without further ado, thanks guys. Uh, enjoy the talk. Thank