# Towards Data Science podcast 123: "Ala Shaabana and Jacob Steeves - AI on the blockchain (it actually might just make sense)"

- Title: Ala Shaabana and Jacob Steeves - AI on the blockchain (it actually might just make sense)
- Channel: Towards Data Science (host Jeremie Harris)
- Date: 12 May 2022
- Length: 54:43 (podcast audio)
- URL: https://www.youtube.com/watch?v=sRn-Q_mSjAM&t=792s
- Jacob's start offset: 13:12 (given). In the podcast audio that lands in Jacob's passage on representations as the digital commodity (12:18–13:09) and the host's recap. Jacob's first answer is earlier, at 4:22.
- Method: YouTube blocked from this VM. The show's podcast RSS feed (Anchor) carries the same episode; the audio was downloaded and transcribed locally with faster-whisper `medium`. Timestamps are from the podcast audio and accurate to a few seconds. The YouTube version may be offset by a few seconds.
- Format: research transcript. Host and Ala paraphrased; Jacob's key lines quoted verbatim. This is the earliest recording in the set (four years before the others) and shows the thesis before subnets and dTAO existed.
- Speakers: HOST = Jeremie Harris; ALA = Ala Shaabana; JACOB = Jacob Steeves.

## Outline with timestamps

0:00 HOST intro. "AI on the blockchain sounds like a winning entry in startup buzzword bingo," but the founders have good reasons. Themes: decentralising AI research, rewarding labs for models not papers, robust benchmarks, concentration of power.

1:45 HOST: backstory. Ala was at FOR.ai (Aidan Gomez's group); Jacob was at Google.

2:39 ALA tells how they met: he pitched at FOR.ai; Jacob messaged, "I've been working on this problem for a lot longer than you have, come join me," with a prototype running. "Aidan actually introduced us."

3:30 HOST asks for the basic ingredients of blockchain. JACOB: beyond decentralisation and censorship resistance, Bitcoin invented "this concept of digital trust and the ability for computers to reach consensus." "blockchains are basically tech tools for having computers reach consensus about something." The bridge: "Bitcoin is this market for hashing power. It's the largest in the world. And we're going to take that same technique of building these global markets, which grew Bitcoin to this global size, and apply it to the other computational problem, which also has its own digital commodity, digital intelligence."

5:15 HOST asks for the explain-like-I'm-five version. JACOB: "a blockchain is just an append-only database where it's distributed across many computers and they effectively vote on what the next thing in the database is going to be." Votes are compute in Bitcoin, wealth in proof-of-stake. Why vote at all: "IPs are basically infinite, they can be spoofed." Explains block and chain plainly.

7:51 HOST: how does this connect to ML? JACOB: "the largest supercomputer in the world is Bitcoin. It's like 500 times bigger than all of Google's data centers." Personal detail: "I worked at Google and I remember when they finished building a data center, they would just start building another data center beside it." Bitcoin "created a very efficient market where anybody in the world could work on the arbitrage opportunities." Mining moved from laptops to "a freezing river in Finland or using geothermal energy in Iceland." The claim: "we can get to Google and OpenAI size in terms of neural networks... if we can use the same tooling."

10:36 HOST guesses: replace random-number guessing with useful computation? JACOB: "Yeah, exactly." The commodity in 2022: representations. "that representation is valuable. And so that's the digital commodity that we're working on. It's literally the representation of inputs." Hashes are "immediate to validate"; "the value of a representation of input is not so obvious." "our consensus mechanism is designed to work with that sort of fuzzy data." Representational knowledge is "the most foundational thing in AI."

14:02 HOST: how is value assessed? JACOB: two roles, miners producing representations ("maps of the territory") and validators testing them on next-token prediction. "machine learning by its very nature sifts out signal from noise. So are you producing signal?" Bittensor as "an incentivized benchmark... a very high resolution benchmark."

16:35 HOST confused about who gets paid. JACOB: block production is a standard blockchain mechanism; consensus about value runs as a smart contract above it, "we're like layer two." The intersection of validators' judgments decides "who gets inflation."

18:24 HOST: how do you make money? ALA: fair launch, no ICO, every TAO mined; one TAO per 12-second block to the best performers; the block carries a ranking matrix.

20:06 HOST: how do you measure value? JACOB: Fisher information, "an approximation of how different or worse off the validator would be if a particular node was removed." Salience methods; LeCun's old paper; entropy via KL divergence. HOST: like an ablation study or permutation importance. JACOB: "and then the ones like Shapley values, similar."

21:48 HOST: why use Bittensor instead of AWS? ALA: a large shared dataset on IPFS; models learn from each other; information "compounds" instead of being retrained. JACOB adds: conditional computation lets validators learn "you're valuable for this particular type of data," e.g. a German model or an arXiv model. Aim: "a corpus or a library of machine intelligence that is continually growing," "a continuous machine learning library." Each endpoint "is like the company OpenAI." "we're the protocol that stitches together all of those AI companies... the neural internet, as we like to say, in a kind of a cliché way."

27:47 HOST: mechanism for learning from each other? ALA: a protocol, an API, a decentralised mixture of experts, like a DNS lookup for experts. HOST: mixture of experts, not ensemble.

31:15 HOST: how does this differ from SingularityNET? ALA on contribution barriers and human bias. JACOB: "SingularityNet is like a human mediated market" with little composition between models; Bittensor is "designed around the language torch"; "it's a machine to machine market. It's neural networks learning which other neural networks are valuable and rewarding them for such." Then the shared vision: "Are we in 1983, where it's going from mainframes to the internet?" "The only thing bigger than the OpenAIs and the Googles is all of them combined." Bitcoin "gets to that size by being permissionless open over the internet."

35:39 HOST on the calibre of people in crypto (OpenSea anecdote).

37:23 HOST: how big is the network? ALA: launched November 2021; 2,048 active nodes; a sybil-resistant proof-of-work entry funnel; a wait list. JACOB: parameter count is a conservative estimate, "about 110 billion," because "we have no idea what it is that is actually running on the system." Scaling by requests processed and sequence lengths; "at the token price, we can calculate how much computational power we could be extracting in a perfect market," heading toward "hundreds of thousands of dollars per day of compute."

41:39 HOST: is this environmentally wasteful and uneconomic? JACOB: "there's a lot of waste in the machine learning world as is today... tons of researchers training the same models and publishing them to NeurIPS... why don't we do it collaboratively?" "We're getting the commodity we want," not papers. Compute "tends to arbitrage for cheap electrical sources." "75% of Google's energy costs go to machine learning." "we're definitely gonna be using electricity, there's no question about it, but hopefully we're gonna use it in an effective way."

44:13 HOST: malicious use and alignment risk seem worse when you decentralise. JACOB's answer is the earliest recorded version of the core message: malicious use is what "centralized organizations focus on because the more glaring problem is the fact that they're deciding who gets to decide what is good and where AI should be moving and who owns it." Decentralisation means "the question of what can or cannot be done with AI is answered collectively in a democratic way instead of the ivory tower people saying I don't like this." Caveat: "Ivory tower people are really smart and they maybe have a really good view from their tower." Then: "our thesis here is that the bigger issue for AIs is not the AI fire the nukes and they take over the world. It's more that like a small group of people come to own it and use it for themselves rather than that collectively it's aligned with humanity." "decentralization is all about decentralization of power." "there's not just five big players, there's tens of thousands of people that control AI." Close: "making sure that the AI's bottom line is us and not some small group of people."

47:46 ALA: AGI needs more teamwork than one company; "I find it a little bit more terrifying that a single company... would own AGI compared to all of us owning a piece of it."

48:41 HOST pushes back: he thinks AGI is intrinsically dangerous and alignment unsolved; is there consolation here? JACOB: it depends "what is it tethered to... what is it working for?" "We're tying that objective to tokenization and decentralization... it's got to be tied to the holders of these tokens. And so then that's its leash for AGI." HOST gives the paperclip argument. JACOB: "I do actually buy that argument, I believe it." Then the economic-environmentalism analogy: capitalism is an AI that only cares about GDP and cuts down forests; alignment means the bottom line must see the forest. "we need to make sure that the humans are owning the things that it wants. So it doesn't kill the humans." "Number go up means actually help all those people that are holding these tokens." Honest limit: "that symbiosis is a really difficult problem. And we don't really have the answers to that... we're not claiming that we have the perfect answer. But we think that tokenization allows us a lot of maneuverability."

52:58 HOST: the metagame remains a problem, but "do we want the world to go to shit in the hands of a specific company?" JACOB: "that's the problem with alignment. This is a beer problem. We just sit down and have a beer and talk about it."

53:48 ALA gives links (GitHub, bittensor.com, Discord). Close.

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

- "blockchains are basically tech tools for having computers reach consensus about something." (4:40)
- "I worked at Google and I remember when they finished building a data center, they would just start building another data center beside it." (8:10)
- "machine learning by its very nature sifts out signal from noise. So are you producing signal?" (15:35)
- "It's neural networks learning which other neural networks are valuable and rewarding them for such." (34:10)
- "Are we in 1983, where it's going from mainframes to the internet?" (34:30)
- "The only thing bigger than the OpenAIs and the Googles is all of them combined." (35:10)
- "the bigger issue for AIs is not the AI fire the nukes and they take over the world. It's more that like a small group of people come to own it and use it for themselves." (46:40)
- "making sure that the AI's bottom line is us and not some small group of people." (47:35)
- "that's its leash for AGI." (49:45)
- "we're not claiming that we have the perfect answer." (52:20)
- "This is a beer problem." (53:48)