Papers · Alternatives to ZK · opML

opML

opML: Optimistic Machine Learning on Blockchain

KD Conway, Cathie So, Xiaohang Yu, Kartin Wong

Venue
arXiv:2401.17555
Date
2024-01-31
Affiliation
Hyper Oracle
Numbers from
primary
Paper
https://arxiv.org/abs/2401.17555

Approach

Interactive fraud proofs, optimistic-rollup style

Notes

The strongest argument against zkML as a default. Runs 7B-LLaMA on a standard PC with no GPU, because it never generates a proof unless challenged. Cites zkML at >1000x overhead vs plain computation. The tradeoff is latency-to-finality (a challenge period must elapse) and a weaker, economic security model. Note the inversion: when the model is large enough that zk proving exceeds the challenge period, opML reaches finality FIRST.

Our reading

Citation neighbourhood

Builds on
  • none recorded
Cited by, in this corpus

Edges are a proxy: paper A's text mentions B anywhere (body or bibliography). See the full graph.

Discussed in

When verifiability is not worth the priceOptimistic verification, fraud proofs and challenge windowsEvery relaxation, and what an adversary can still do
Other recorded fields
trust_assumption: AnyTrust -- any ONE honest validator can force correct behavior.
  Economic + game-theoretic, not cryptographic.
tradeoff_vs_zk: 'Loses: public verifiability without watching the chain, privacy of
  model/inputs, non-interactivity. Wins: ~1000x cost, no trusted setup, runs on commodity
  hardware.'