papers/hanzo-consensus-ai/sections/01-introduction.tex

\section{Introduction}

\label{sec:intro}
% ============================================================================

Decentralized AI compute markets face a verification problem absent from
traditional blockchains. In Bitcoin, verifying a transaction requires checking a
digital signature---a deterministic operation costing microseconds. In a
decentralized AI network, verifying an inference result requires re-executing
the model forward pass---a stochastic operation costing milliseconds to seconds
on expensive GPU hardware.

Full verification (re-executing every request) doubles compute cost and negates
the economic benefits of decentralization. No verification allows rational
operators to return random outputs, collect fees, and save GPU cycles. The
challenge is designing a verification protocol that:

\begin{enumerate}[leftmargin=1.4em]
    \item Makes cheating economically irrational under the HANZO staking model.
    \item Imposes minimal overhead ($<$10\% of total compute).
    \item Handles the non-determinism of GPU floating-point arithmetic.
    \item Scales to millions of inference requests per day.
\end{enumerate}

This paper describes the three consensus mechanisms deployed on Hanzo
ACI~\cite{hanzoaci} that address these requirements.

% ============================================================================