Series · ongoing
A series on verifiable machine learning: how zero-knowledge proofs, optimistic schemes, and crypto-economic incentives can make AI inference trustworthy on-chain.
Zero-knowledge proofs promised to make machine learning trustless. A field survey of where zkML actually stands — proving systems, quantization tradeoffs, and what's deployable today.
Zero-knowledge proofs aren't the only path to trustworthy on-chain AI. Optimistic schemes trade latency for a 1000x cost reduction — here's how dispute games over inference actually work.
zkML costs 1000x, optimistic schemes cost a challenge window. Hardware attestation verifies AI inference at under 7% overhead and ~$0.26 on-chain — if you're willing to trust Intel. Part 3 weighs the third leg of verifiable inference.
FHE left the lab: confidential ERC-7984 tokens settle on Ethereum for ~$0.09 in gas, with 48,000 transfers since December. We dissect a live encrypted transaction, the 13-node MPC committee underneath, and why encrypted inference is still 11 seconds per token.
zkML, optimistic, TEE, FHE all prove the computation. Restaking takes the other road: bond it and slash liars. We do the cost-of-corruption math behind EigenLayer's $18B AI-AVS security, the overloading attack that breaks it, and the probabilistic-audit tax.
Bond-and-slash verifiable AI assumes someone audits. But checking costs money and pays nothing, so a rational verifier won't — the verifier's dilemma. Proof of Sampling fixes it with a bounty, and the math says honest inference holds at a 0.74% sampling rate where nobody ever arbitrates.
DeepSeek-R1 emits 8,400 thinking tokens before a 341-token answer on hard math. Every existing on-chain verification scheme — TEE, optimistic, ZK, sampling — was designed when computation and output were proportional. They aren't anymore.
An SP1 proof costs 327k gas on Ethereum mainnet. At a 30M-gas block target, a naive AI inference market maxes out at 91 verifications per block. SnarkPack's O(log N) aggregation collapses 1,000 proofs to a single 880k-gas verification — here's the mechanism, the math, and the latency tradeoffs.
LLM non-determinism has been the blocker for output-based verification. DiFR commits to a random seed before inference — and activation fingerprints built from random orthogonal projections detect 4-bit quantization in just 2 tokens at AUC > 0.999.
DiFR's seed-commitment round-trip is the last synchronization barrier to async verifiable inference. TOPLOC removes it: a locality-sensitive hash of intermediate activations compresses to 258 bytes per 32 tokens — 1,000× smaller than raw embeddings — and validates faster than inference ran.
Zero-knowledge proofs can slash a double-signer but not a lying AI oracle. EIGEN's intersubjective fault model closes this gap via social consensus-driven forking — and it's live on EigenAI mainnet.
MoE routers let inference providers silently halve the experts they activate — saving 31% of compute while staying undetected by every current fingerprinting scheme.