AI x Crypto

Episode Summary

Executive Summary: The episode frames AI and crypto as complementary forces: AI centralizes power through data, compute, and models, while crypto decentralizes control through verification, ownership, and incentives. The guests explore how crypto can help AI become more private, trustworthy, and community-owned, and how AI can help crypto with code generation, security, moderation, and identity verification.

Main Topics: AI vs. crypto as opposing system designs (Priority: 5/5): Ali argues AI tends toward centralized, top-down control and privacy loss, while crypto enables decentralized cooperation, sovereignty, and user control. This is presented as a broad meta-framework for understanding their relationship. Zero-knowledge proofs for machine learning (ZKML) (Priority: 5/5): Dan explains how zero-knowledge techniques can verify model inference and, eventually, aspects of training, allowing users to confirm a model was actually run as promised without exposing private data or model internals. Decentralizing AI compute, data, and models (Priority: 5/5): Ali breaks AI into three prongs—compute, data, and models—and describes how crypto-based marketplaces and verification systems could distribute each across communities instead of central platforms. AI-generated code and crypto security (Priority: 4/5): The guests discuss both the promise and danger of LLMs writing Solidity and cryptography code. AI can help generate and debug code, but current models can produce insecure implementations and must be paired with formal verification and static analysis. Proving authenticity, humanity, and media provenance (Priority: 5/5): They explore blockchain, trusted hardware, and ZK proofs for fighting deepfakes and proving a person is human, which becomes more important as bots and synthetic media dominate online spaces. AI for crypto trust, moderation, and MEV defense (Priority: 4/5): AI can help identify suspicious transactions, warn users, and potentially reduce harm from MEV and other adversarial behavior in crypto, improving safety in permissionless systems. NFTs, creators, and community in an AI-saturated media world (Priority: 4/5): They argue that as AI creates abundant media, crypto can preserve human provenance and enable creator communities, with examples like sound.xyz and collective, on-chain art generation.

Key Arguments: AI and crypto are natural counterweights: AI concentrates decision-making and data power, while crypto distributes control and ownership. Zero-knowledge proofs can make ML more trustworthy by proving a model was evaluated correctly and, in the future, that training was executed as claimed. A decentralized AI stack needs marketplaces for compute, data, and models so contributors can share upside rather than value accruing only to incumbents. Current ZKML can prove some medium-sized model classification tasks, but not large frontier models or full training at scale yet. Distributed compute for AI still faces two big problems: cryptographic verification and distributed systems coordination. Data marketplaces for AI risk fake or poisoned data, so authenticity checks may require trusted hardware, reputation systems, and improved attribution methods. LLMs can generate useful code, but current code generation is unsafe for smart contracts and cryptography unless paired with formal verification and testing. AI can improve crypto security by flagging suspicious transactions, detecting bugs, and helping users avoid MEV-related harm. Blockchain plus ZK proofs can help prove media provenance and authenticity, including edited images and potentially deepfake defense. Proof of humanity will matter for governance, voting, and online participation because bots can otherwise overwhelm one-person-one-vote systems. In an abundance-of-content world, crypto can help creators build direct communities and sustain value around human-authored or human-led art.

Data Points: LLM foundation: 2017 - Ali references the transformer paper 'Attention Is All You Need' as the basis for modern LLMs. Peter Thiel AI/crypto framing: 2018 - Ali cites Peter Thiel’s line that AI is communist and crypto is libertarian. AI trend scale: thousand X more - Ali describes a future where AI-generated content may outnumber human-made content by roughly 1,000x. AI trend scale: a thousand to one or a million to one - Ali estimates bots and synthetic content may vastly outnumber humans online. Model scale currently provable with ZKML: medium-sized models - Dan says current ZKML can prove classification for medium-sized models, not GPT-3/4-scale systems. Current frontier model limitation: GPT-3 or 4 not provable today - Ali says current ZKML performance is nowhere near enough to prove training or inference for large LLMs. Number of AI prongs: 3 - Ali breaks decentralized AI into compute, data, and models. Music platform example: sound.xyz - Used as an example of NFT-based creator/community infrastructure. Identity hardware example: WorldCoin orb - Ali cites the biometric orb as an example of proof of humanity using secure hardware and ZK privacy. Camera provenance standard: C2PA - Dan discusses camera signatures and provenance metadata as a way to authenticate images and video.

Pivotal Quotes: "AI is very much a technology that thrives and enables top-down centralized control, whereas crypto is a technology that's all about bottom-up decentralized cooperation." — Ali Yaya: Opening framework for why the two technologies act as counterweights. "One of the things that will become important in a world where anyone can participate online is to be able to prove that you are human for various different purposes." — Ali Yaya: Discussion of proof of humanity, governance, and bot resistance. "If we're going to incentivize people to contribute data, basically we're going to incentivize people to create fake data so they can get paid." — Dan Bonet: Explaining the core data-authenticity challenge in decentralized AI marketplaces.

Implications: Listeners should expect AI and crypto to converge around verification, ownership, and incentives. Near-term opportunities are in secure code generation, data provenance, proof of humanity, and decentralized infrastructure; the biggest blockers remain scalable verification and distributed coordination.