Research

Watch · narrated whiteboard episodesL3

Zero-Knowledge Proofs and Verifiable AI

A zero-knowledge proof lets one party convince another that a computation was performed correctly without revealing the inputs — which is exactly what verifiable AI inference needs. This series builds the idea from interactive proofs and the simulator argument through Fiat-Shamir, SNARKs, transparent STARKs, and closes with zkML: proving a model produced an output on a committed input. Grounded in Goldwasser-Micali-Rackoff, Groth16, and Ben-Sasson STARKs.

Murali Chillakuru·5 episodes
  1. 24 min Episode 1Interactive Proofs and Zero-Knowledge, Defined: Completeness, Soundness, and the SimulatorA moderator and a cryptography expert introduce one of cryptography's most beautiful ideas — the interactive-proof model, the zero-knowledge property of proving a statement while revealing nothing else, a concrete example that builds intuition, the simulator argument that defines it rigorously, the difference between proofs and arguments, and why it matters for AI agents.
  2. 24 min Episode 2From Interactive to Non-Interactive: Fiat-Shamir, Commitments, and the Random-Oracle CaveatA moderator and a cryptography expert show how zero-knowledge proofs become practical — why removing the interaction matters, commitments as the key building block, how they bind and hide, the Fiat-Shamir transform that self-generates the challenge, the random-oracle caveat behind its security, what non-interactivity unlocks, and the stakes for AI agents.
  3. 25 min Episode 3SNARKs: Succinct Arguments, the Arithmetic-Circuit Pipeline, and the Trusted-Setup QuestionA moderator and a cryptography expert unpack succinct proofs — what 'succinct' means and why it's remarkable, the arithmetic-circuit pipeline that turns any computation into something provable, how the proof becomes tiny and fast to verify, the trusted-setup question and its risks, the trade-offs of the technology, and the stakes for AI agents.
  4. 24 min Episode 4STARKs and Transparency: Hash-Based, Post-Quantum-Plausible Proofs Without Trusted SetupA moderator and a cryptography expert examine the transparent alternative to SNARKs — what transparency means and why no trusted setup is valuable, the hash-based construction, why building on hashing makes STARKs quantum-safe, the size and time trade-off against SNARKs, when to choose a STARK, and the stakes for AI agents.
  5. 25 min Episode 5Proving an ML Inference: zkML, the Committed-Input Guarantee, and the Honest CostIn the zero-knowledge finale, a moderator and a cryptography expert reach the frontier — proving that an AI model ran correctly on an input without revealing the input, the model, or both, the committed-input guarantee that pins down what was run, the honest computational cost, when it's worth it, the series synthesis, and why it's foundational for AI agent trust.