The Arimoto algorithm computes the Gallager function $\max_Q {E}_{0}^{}(\rho,Q)$ for a given channel ${P}_{}^{}(y \,|\, x)$ and parameter $\rho$, by means of alternating maximization. Along the way, it generates a sequence of input distributions ${Q}_{1}^{}(x)$, ${Q}_{2}^{}(x)$, ... , that converges to the maximizing input ${Q}_{}^{*}(x)$. We propose a stochastic interpretation for the Arimoto algorithm. We show that for a random (i.i.d.) codebook with a distribution ${Q}_{k}^{}(x)$, the next distribution ${Q}_{k+1}^{}(x)$ in the Arimoto algorithm is equal to the type (${Q}'$) of the feasible transmitted codeword that maximizes the conditional Gallager exponent (conditioned on a specific transmitted codeword type ${Q}'$). This interpretation is a first step toward finding a stochastic mechanism for on-line channel input adaptation.