# The Block Runner #156: "How Bittensor Saves Humanity From Runaway AI"

- Title: How Bittensor Saves Humanity From Runaway AI | Interview with Co-Founder Jacob Steeves | TBR #156
- Channel: The Block Runner (hosts William and Iman; TJ on production)
- Date: 20 May 2023 (recorded the week Sam Altman testified to the US Senate, 16 May 2023)
- Length: 1:16:57 (podcast audio)
- URL: https://www.youtube.com/watch?v=h2CmXVMfiB8&t=1s
- Jacob's start offset: 0:01 (given). Jacob's first answer is at 2:40.
- Method: YouTube blocked from this VM. The show's podcast RSS feed (Podbean) carries the same episode; audio downloaded from the episode page and transcribed locally with faster-whisper `medium`. Timestamps from the podcast audio, accurate to a few seconds.
- Format: research transcript. Hosts paraphrased; Jacob's key lines quoted verbatim. This is the second-oldest recording (2023, one subnet, pre-dTAO) and shows how early the core message was fixed.
- Speakers: HOSTS = William and Iman; JACOB = Jacob Steeves.

## Outline with timestamps

0:03 HOSTS introduce the episode and the four topics: runaway AI, how Bittensor works, miners vs validators vs stakers, community questions.

1:34 HOSTS: Altman is in front of the Senate right now. Could decentralised AI prevent the collapse of humanity?

2:40 JACOB: "I think that is valid. I think it's a philosophical take. It's certainly one that's slightly libertarian in nature." The other view is top-down control, "and I think there's certainly a value in that type of control of AI. But I don't think it's the big problem." Then, in 2023, the core message: "the bigger problem is that we're building this incredibly powerful tool that is going to be gate-kept and controlled and owned by a really small minority of people that use it to exponentiate their power." "humanity is more of a problem for humanity than AI." "decentralized artificial intelligence, which really means just decentralized in terms of power." The two reasons for Bittensor: access ("open ownership systems") and scale ("soaking up as much computational power as we can"); "the ethics and the technical side of Bittensor are very well interwoven."

4:40 HOSTS: what about bad actors and a runaway AI that cannot be shut down?

5:20 JACOB: the runaway AI "is pretty sci-fi." Taking the nukes "would require us giving the nukes in the first place, which means that there's this unholy alliance between some people in humanity and the computers." "at the end of the day, who controls it that matters and what power they give it." More urgent: censorship. "This is an extension of our brain and it's going to be used to manipulate people... to narrow the perspectives that we have at the behest of the political opinions that that small group has." Then the tether idea: "it's going to run away in a sense, no matter what we do, it's really about who it runs away with." "if we are collectively in symbiosis with this technology, then it's not going to fight against us because we are its bottom line. Right now, with companies like OpenAI and Google... its bottom line is the corporate shareholders of those institutions, it's not me and you." AI "adapts at the end of the day to the reward signal." "Tethering that to TAO was our goal... so that when it does run away, which it will, it's going to run away, pulling us with it." "I actually think it's a bit of a red herring for the real problems arising around centralized ownership and control."

8:30 HOSTS compare this to Musk's Neuralink merge idea; will humans always be needed?

9:25 JACOB: if TAO is the AI's "metric of value" and humans hold it, "where the rubber hits the road... we still maintain the control." Neuralink "could actually just metastasize it... now the AI has a super highway into our brains." "I think that incentives are everything." Democracy analogy: "Democracies are really slow. Democracies are really inefficient... but the fact that we do have a vote means that this really complex, unwieldy machine... at the end of the day, it's going to have to come back to us and try to get our vote." Without that tether "we would be really shit out of luck."

12:29 HOSTS: what would a runaway event look like? JACOB: "Let's be honest. What's actually going to happen is that that's going to be a human trying to build a monopoly for its AI with the help of governments to build a centralized power structure... this weird mix of corporate capitalism mixed with AI." First Elysium reference on tape: "It's like that movie Elysium, where there's the humans in space that are living on this halo ring thing... they just kind of enslave humanity. I think that's much more likely than, oh my God, like iRobot." "this is actually what's already happening." On Altman testifying: "I'm sure that he is incentivized to say that AI is dangerous and likely going to run away because then he can get them to give him regulations which would stop other competitors." "it's a great narrative pitch."

16:20 HOSTS: is an uncontrollable AI even possible? JACOB: many systems are not under anyone's control: capitalism, democracy. "It's not really being controlled... Instead, it's being aligned." "Bitcoin is uncontrolled. Totally uncontrolled. But also at the same time, it's controlled in the perfect manner because it's run by people that own Bitcoin." "No one can turn off the economy." "let's build mechanisms of incentive that are aligned rather than totally controlled." "in a sense, Bittensor is uncontrolled, right? And we aspire to that." On distribution: "We had advice like, don't do that. Hey, you should just hold most of them for yourself so that you can control the price... But that's totally antithetical to the mission itself."

19:20 HOSTS: the elevator pitch to Elon. JACOB: "the ethics side doesn't really come up because people are like, how are you gonna make money?" "intelligence is a commodity... If you're intelligent, people wanna talk to you." "Markets are the most efficient way of extracting resources that humanity has ever come up with, which is really crazy. It's not like top-down bureaucratic control over mining companies. It's actually competition between them." Bitcoin "has become the largest supercomputer in the world by many magnitudes based on that premise." Over the past year: "almost impossible for our own selves to allocate the compute that people had thrown at the system." "An open, anonymous, censorship-resistant, permissionless, liquid market." Management aside: "well aligned people are like a thousand times faster than a thousand people."

22:30 HOSTS: the technicals, three or four participants. JACOB on miners (2023): "They are hosting and inferencing machine learning models." "They don't actually need to have a machine learning model at all. They could be connecting to an endpoint. They could be sitting behind the screen typing things in. But the output is what we measure." Example: a Vicuña 13B model on a 20 GB GPU. Validators rank miners; the chain reaches consensus; tokens flow "in a higher proportion" to the best, creating "an incentive gradient." The synthesis: ML is "going downhill in a loss function" and incentives are "people wanting to go uphill to maximize their incentives. And those things are very well related." 50% of emissions to validators, "sort of like a prediction market or like a bonding market."

27:59 HOSTS: how does a validator judge a picture of a cow? JACOB explains CLIP embeddings and cosine distance; the incentive is "isomorphic, to use a mathematical term, to the loss function that's used to train those types of models." On sourcing outputs from anywhere: "it's like a vacuum, like we're vacuuming in the intelligence from everywhere else and pulling it into the single system."

32:40 Inference cost: "over time, inferencing... becomes the really expensive part." "I predict that over time this will become more and more where these companies are focusing their budgets."

34:13 HOSTS: how does a developer use it? JACOB: a Python API through validators; "secret knowledge is that Bittensor is wide open... hardly anyone is querying it."

35:45 HOSTS: what metric wins developers? JACOB: parameter size, query time, context; plus "censorship resistance... We can't stop you. And that's really cool." Diversity: "centralized systems will tend to converge to a particular opinion." Google's mission "became organize the world's information in a way that their advertisers make money." Bittensor's: "organize the world's intelligence and make it accessible and democratically accessible to everybody." "if intelligence was a wave or like a signal, we're trying to maximize the broadband throughput of the network."

39:30–46:45 Validators as businesses gating access; 128 validator slots; adaptive stake thresholds; "That's competition, dude."

49:50 On the community: "the people come out of the woodwork... I would never hire them, frankly, a lot of them, no way. But they turn out to be the people I should have hired. It's like, because they've done the work." "because we didn't do a pre-mine. We didn't raise like a stupid amount of money."

51:21 HOSTS: could Google and OpenAI join? JACOB: "I would love it if they were... I think it's too icky, right? Web3. Ew." "I've actually asked Sam Altman to do that on a phone call. And it didn't come through." "the only thing bigger than Google and OpenAI is like everyone combined."

54:28 HOSTS: is it a computational arms race? JACOB: compute, data, innovation, and "access to machine intelligence itself, which is recursively useful." "it's a computational arms race, but it's not only a computational arms race."

56:40 Community question: when will you hand over the network? JACOB (2023 roadmap): decentralising "the sudo key across a number of individuals, which is still quite centralized"; delegates as "true representatives... voted for in a liquid manner" within six months; then a democracy system; "finally, we will likely delete the executive altogether." Pressure from regulation: "We need to, as a collective of people, reduce our surface area and that's another reason for decentralization. We may be forced out the door sooner than we'd like."

60:42 On Polkadot: independent chain now; may return "maybe not this year, maybe not the year after."

62:15 Privacy question. JACOB: "It doesn't." "we're bullish on openness."

64:00 How is it protected from corporate takeover? JACOB: "at any moment all TAO holders could just sell all of their TAO to Google" but that "would also destroy the value proposition." "Oh, you don't like it? Just fork it." Compute takeover: "why would they do that? They would be destroying the project that they're mining into." The real answer "comes from decentralization."

68:27 What is a threat to Bittensor? JACOB: "ourselves. The people that are here in this ecosystem. In a sense, it's like a baby. It's slightly fragile. Even Bitcoin was the same." "We could clutch our pearls and refuse to give up power with decentralization. I think that would be a risk." "Maybe nobody cares." "you plow on, right? You believe in your vision and you work hard." "we have people that have done proof of work... they committed their time to this project."

71:36 Recursive training via distillation: "it's like a mountain of data that we're building." "In a sense, we're mining the network itself as inputs to the network." "the holy grail for us is building this self-referential, self-reinforcing" system.

74:42 Roadmap: sudo decentralisation "in the next quarter," proposal voting, multimodality. "People believe what they can see. People believe what they can read."

76:13 HOSTS close.

## Jacob's stage-worthy lines (verbatim, Whisper medium)

- "humanity is more of a problem for humanity than AI." (3:40)
- "decentralized artificial intelligence, which really means just decentralized in terms of power." (3:55)
- "it's going to run away in a sense, no matter what we do, it's really about who it runs away with." (6:50)
- "we are its bottom line." (7:05)
- "when it does run away, which it will, it's going to run away, pulling us with it." (8:15)
- "I actually think it's a bit of a red herring for the real problems arising around centralized ownership and control." (8:25)
- "I think that incentives are everything." (10:30)
- "It's not really being controlled... Instead, it's being aligned." (17:05)
- "Bitcoin is uncontrolled. Totally uncontrolled. But also at the same time, it's controlled in the perfect manner because it's run by people that own Bitcoin." (17:30)
- "No one can turn off the economy." (18:00)
- "Markets are the most efficient way of extracting resources that humanity has ever come up with." (20:25)
- "well aligned people are like a thousand times faster than a thousand people." (22:05)
- "we're vacuuming in the intelligence from everywhere else and pulling it into the single system." (31:40)
- "We can't stop you. And that's really cool." (36:50)
- "I would never hire them, frankly, a lot of them, no way. But they turn out to be the people I should have hired." (50:10)
- "Web3. Ew." (51:40)
- "the only thing bigger than Google and OpenAI is like everyone combined." (53:40)
- "We could clutch our pearls and refuse to give up power with decentralization. I think that would be a risk." (69:20)
- "People believe what they can see. People believe what they can read." (75:40)