# The big talk. Tsinghua University, Beijing: "From Bitcoin to Bittensor: Building Open Markets for Intelligence"

- Title: From Bitcoin to Bittensor :: Building Open Markets for Intelligence // Tsinghua university, Beijing
- Channel: The Opentensor Foundation (YouTube). The same full presentation was posted by @opentensor on X on 26 January 2026: https://x.com/opentensor/status/2015773331778158762 ("Full presentation // @HackQuest_")
- Event: Build on Bittensor, HackQuest developer meetup at Tsinghua University, Beijing
- Date: filmed January 2026 (X post 26 Jan 2026; YouTube upload 9 March 2026)
- Length: 33:14 (X copy, 1994 s)
- URL: https://www.youtube.com/watch?v=ZZAbk_NfJBg
- Jacob's start offset: none given; Jacob speaks from 0:00. This is a solo lecture with slides. No host.
- Method: YouTube blocked from this VM. Audio taken from the Opentensor Foundation's own X post of the full presentation (yt-dlp), then transcribed locally with faster-whisper `medium` (int8, CPU). Timestamps are from Whisper and are accurate to about a second.
- Format: research transcript. Full outline in plain words with timestamps, plus Jacob's key passages quoted verbatim. Jacob calls this "my big talk." It is the reference for the Montreal structure.
- Speaker: JACOB throughout. Audience asides noted.

## Structure at a glance

| Time | Beat | Function |
|---|---|---|
| 0:00–1:54 | Who I am, and what this is not | Credentials in 60 seconds; disarm the sales frame |
| 1:54–6:52 | AI before and after 2012 | Establish the pattern: state, objective, feedback, adaptation |
| 6:52–9:35 | Slime molds, trees, lightning, rivers, us | The pattern is nature's pattern |
| 9:35–13:58 | Bitcoin is that pattern too | Reframe Bitcoin as a self-adapting computer; the scale numbers |
| 13:58–16:38 | Why it works: the six qualities | Borderless, continuous, autonomous, permissionless, no HR, pure market |
| 16:38–19:53 | Hashes are useless, so generalise it | Name the idea: incentive computing; define Bittensor in one sentence |
| 19:53–23:33 | Proof 1: coding agents (Ridges) | "An AI lab without any engineers" |
| 23:33–25:29 | Proof 2: training a 70B model across the internet (Templar) | "Bitcoin mining to training machine learning models" |
| 25:29–30:28 | Proofs 3 to N: GPUs, inference, robotics, the long list | Breadth |
| 30:28–31:23 | Apply it to ourselves: dynamic TAO | "reinforcement learning to itself" |
| 31:23–33:08 | Why: the 3,000-employee company | The moral close, kept short for an academic room |

## Outline with timestamps and key lines

0:03 Opens flat, with biography. Studied AI and mathematics in Canada, bachelor only, 2015. "AI was a lot different in 2015 than it is today." Promises "the journey from that year to where we are now." DARPA contractor, neuromorphic chips; Bitcoin in 2016; "started building Bittensor almost 10 years ago"; Google Brain; quit in 2021 to start "a digital currency network." "I live in Peru and this is really my life."

1:23 Disarms the room: "I'm not here to sell you guys a digital currency. I'm not going to be talking about prices and bullishness." He wants to talk about "the fundamental ideas of why this intersection of artificial intelligence and digital currency exists and why you should care about it, whether or not it's Bittensor or any other project. If they're doing it properly, it's interesting."

1:54 The 2010 world. MNIST; asks the room who knows it. A 2010 paper detected digits by "curved straight line analysis" and got 97%, "which is actually really shitty." A two-layer network today gets almost 100%. Joke: "If we'd done a lecture on AI in 2010, none of you would come. It was like boring."

3:45 AlexNet, 2012. The innovation in one sentence: "instead of pre-encoding the representations, we let the model learn what to look for inside of the image." Then the loop: weights, loss, gradient, adapt. "So this is like a loop, right? And you just let it run."

5:26 The same loop in RL (environment, agent, reward), in GPT-5's RL fine-tuning, in genetic algorithms. "There's a feedback loop. And we can run this feedback loop a billion times." The AI winter: "Now we're in where it's hot outside right now."

6:52 Nature. Slime mold in a maze following the gradient of smell to oats ("You probably eat it in China"). "It's like the simplest form of intelligence in biological systems that we study."

8:26 Goes deeper: trees, leaves, lightning bolts, river deltas: "structure that gets energy and adapts itself to maintain its energy." Humans too: "We're basically energy-adapting structures that try to find energy so we can sustain our structure, and if we can't, we die." The line that ties the first ten minutes together: "There's a pattern. All these things are one thing. State, objective, feedback, adaptation, a loop, and things move quickly."

9:35 The turn: "OK, but why am I telling you all this stuff? Because we're not talking about just AI." Bitcoin, seen while studying neuromorphic chips with the US military: "Bitcoin was also an example of this phenomenon, not as a currency, but as a network, that Bitcoin is a self-adaptive computer, a network that produces hashes." Miners produce hashes; the network checks and pays in BTC.

11:09 The scale passage, delivered with repetition: "Bitcoin is the largest supercomputer in the world. Let me repeat that. It is the largest supercomputer in the world." Claims: 1,000 times the compute of the six largest US compute providers; 10^21 hashes; "23,000 megawatts of continuous draw... the same electrical draw as Thailand." "Just a network." Aside: it has to be that big "because if it's not large enough, it can't be resistant to governance, but that's a different conversation." "As the price of Bitcoin goes up, the miners produce more hashes." Cost comparison: about $1 trillion for 1,000 exaflops at the big six versus $50 to 300 billion for "450,000 exaflops" in Bitcoin; "somewhere between 709,000 times more efficient at producing hashes." (Caution for Montreal: hashes are not FLOPs. The thesis already flags this. Keep the "same draw as a country" image; drop the exaflop equivalence or say "hash operations.")

13:06 Why that framework is right: the same "exponential efficiency and scale and improvement" seen after AlexNet, "but we're seeing it again in a network for computation." Bitcoin is "kind of like a program... an optimization algorithm... applied to creating this thing that's called a digital commodity."

14:08 "Bitcoin is the first ever example of this in nature. We've never really produced anything like this." Then the six qualities, each a short sentence:
- Borderless: "Doesn't matter where you are, you can mine Bitcoin with your laptop."
- Continuous: "It doesn't have a start or a stop time and it doesn't close on weekends." Compared to seminars, school seasons, the US stock market, sports leagues. "It's 24-7."
- Always optimising: "It's very, very competitive and it's always optimizing itself."
- Autonomous and permissionless: "you don't have to ask anybody to participate in this system, it simply draws as much electricity from the grid as it can."
- No HR: "A normal corporation can't achieve that form of efficiency because it needs to have an HR department, it needs to see who you are, there's going to be bias, where did you go to school. Bitcoin doesn't care that you guys went to the best university in the world. Sorry. Bitcoin doesn't care. It just cares if you can produce hashes."
- Pure market: produce more hashes, get paid, "know exactly what your return is going to be."

16:08 Names the idea: "this new type of computing which I'm going to label incentive computer." Places it beside ML, RL, genetic programming: "Incentive computing is also a type of computing that we should study and we should think about. How can we do it properly? How can we do it generally, actually?"

16:42 The pivot line of the talk: "Because we don't want to produce hashes. We want to produce anything. Hashes in Bitcoin are useless. But what if we could take the same premise and we could apply it to storing files, generating compute or do machine learning itself." "What if we could apply that same premise... to, hey, training machine learning models, for instance, or really anything actually."

17:14 Bittensor defined by decomposition. Bitcoin: "The miners produce hashes, the validators check the hashes, the network emits to miners. We built the general version of this. Miners do any work, validators check any work, and the network emits a digital currency to those miners. We called that thing Bittensor." Then the framework analogy: "it's like a torch. It's like a language for creating incentive computing. It's a language."

17:58 Mechanics: register a miner under a key, do work, validators check it against predefined rules, chain reaches agreement on who did well. Slide with green dots (miners) and a ranking curve. "every mechanism on the network of Bittensor has some sort of curve." Over time "we churn out the bad miners and replace them with new miners. It's an adaptive system."

19:53 Proof 1, a coding-intelligence subnet (Ridges): "Bitcoin for being good at programming with Cursor basically." Miners produce agents, agents are scored on a benchmark, reward paid in a token. SWE-bench chart over three months. The key repetition: "the mechanism didn't define the solution to the agent. They simply encoded the incentives under which the agents could be evolved. Let me say that again. We did not make a machine learning agent. We simply defined the incentives under which those AI agents could be evolved. And in a short period of time, we beat the best in the world." Claims the agents beat Claude, OpenAI, any closed or open model on coding challenges. "Who did it? We don't know. Some random person in the world wrote a 7,000 line AI agent." Personal note: "For me, it's very similar to what it felt like in 2012, when we were able to get neural networks to finally work." Then the sentence that should open a Montreal slide: "This is an AI lab without any engineers. This is an AI lab that doesn't define any specific problem, doesn't define the way in which problems are solved, but in fact solely defines the incentive."

22:42 Invitation, no permission needed: "You don't need to ask me any questions. It's an open network. It's a permissionless network." Waves at an ex-miner in the crowd ("Hey, Nick"). Economics: "instead of raising all this venture capital and then hiring HR departments and doing marketing, we just put all of the capital, all of the money straight through the network to the people that are actually doing the optimization." "if you're a top miner on this subnet for one day, you can make $60,000... And you don't need to ask anyone for permission."

23:33 Proof 2, decentralised training. "without tens of thousands of GPUs at our disposal, it's basically impossible to compete with closed source labs. But if we can combine resources from you and me... we don't all have to carry the same economic load." The idea: "a market for the gradients." "an optimization problem for an optimization problem." "We did that, and this is the result." 70 billion parameter model trained across the internet, permissionless, "no whitelists, we don't ask questions, and we pay you for being able to produce a gradient that reduces the loss faster than somebody else. This is Bitcoin mining to training machine learning models." "it's never been done before... this is something I care deeply about."

25:29 Proof 3, GPUs. Verify the GPU is real, pay out, rent it back. "some Chinese miners for instance, contribute GPU resources." "These prices are the cheapest in the world. Because we built this network in a permissionless, borderless, and incentivized way."

26:32 Proof 4, inference. Miners provide endpoints; speed is checked. "we are the largest provider of open source models in the world on OpenRouter." "At one point, it was providing more inferences for DeepSeek than DeepSeek itself."

27:30 Proof 5, robotics. Why incentive computing reaches the physical world: "we're dealing with real, real incentive, not like a reward from a machine learning perspective, where it's just kind of a number on a computer, we're talking about real money, real, real physical money, maybe humanity's most abstracted form of energy." Drone in simulation running a network-provided model, rewarded for fast course completion. "So we're also optimizing things in the physical world."

29:19 The long list, fast: stock-market signals, sports betting, vision models, AutoML, drug discovery, weather, Bitcoin prediction, quantum computing, Bitcoin mining, 3D generation, data understanding, inference speed, commodities trading.

30:28 Reflexive move: "We applied this idea of monetary optimization to ourself." Dynamic TAO: reward mechanisms with liquidity if they are valuable to a market of other mechanisms. "Let's apply reinforcement learning to itself." 128 mechanisms. "All of this is basically us applying ourselves to ourselves."

31:23 The why, said plainly and briefly. "We're moving into an era, probably for the first time in history, that some of the largest companies to ever be created only have one, two, three, 4,000 employees." OpenAI: 3,000 employees, "controlled basically by one person. It's a closed source lab. You'll probably never work there. And for the rest of your life, you'll have a subscription." "All the power for a very small group of people that they make decisions. And ultimately, you don't even know what's happening. You don't know where your data's going. There's some shady shit there. You're not participating in it. You don't own any OpenAI, because it's a closed round." The dystopia: "This kind of top-down world, where artificial intelligence gets pulled into a small set of really small companies, and they control all of the compute and all of the data, and there's nothing that you can access. That's where everyone wants us to go."

32:33 The answer, one breath: "as a consequence of us approaching these problems with this new and very powerful technology of monetary optimization, we also are distributing ownership, making the games transparent, everyone can participate, you can own, you can contribute, and you can access these digital commodities in a different way, in a more fair and open way. So that's why we're really doing this."

33:01 Undercuts his own close: "But I didn't really talk too long about that, because here we are in an academic setting. Anyways, yeah, so that's the end of my presentation. I hope that that was interesting for people."

## Notes for Montreal (see the analysis file for the full breakdown)

- The spine is one pattern repeated five times: loop in ML, loop in RL, loop in nature, loop in Bitcoin, loop in Bittensor. Each repetition earns the next. Keep this.
- The moral argument gets 100 seconds at the end and he apologises for it. In Montreal it should be the frame, not the coda.
- Every claim is shown as a chart from a live subnet. Proof before philosophy.
- He never says "decentralised AI" as a slogan. He says "incentive computing," "monetary optimization," "digital commodity," "market for gradients."