Coherent oscillations of neuronal activity are ubiquitous across brain spatial and temporal scales. Gamma oscillations (30-120Hz), detectable with Local Field Potentials (LFPs), are classically analyzed through massive trial averages. This approach leaves us with one average oscillation shape, one mean power and a view based on rigid frequency bands. Consistently, mechanisms based on biophysical models of neural networks typically predict periodic oscillations (e.g. PING models [1]). The functional role of oscillations have been largely investigated in the hippocampus, where, different gamma-frequency bands, timed at different phases of the ongoing theta oscillations (10 Hz), would mediate information from different sources [2]. To characterize theta-gamma oscillations, LFPs were recorded in the dorsal hippocampal CA1 during a spatial navigation task in mices. A huge variety of gamma bursts with different amplitudes, frequencies and phases of occurrence relative to the theta cycle was observed (Fig.1a). To investigate if this diversity was functional or just noise, we trained classifiers to decode the position of the mouse, based on the features of gamma bursts. Decoding was possible even from only those gammma burts at weak amplitudes (Fig. 1b), considered noise until now and neglected in classical approaches. But what is the origin of gamma bursts diversity? We hypothesised that it stems from the dynamic regime of operation of the local networks in the Hippocampus. Such diversity, natural consequence of randomness in recurrent synaptic connections combined with balanced excitation and inhibition, would then be shaped by the detailed firing of excitatory and inhibitory neurons in the source population. We constructed a spiking model for balanced excitatory-inhibitory populations with random recurrent connectivity. We found that gamma diversity robustly emerges for most parameter combinations (Fig. 1c), provided the network remains not too far, but still below a transition to strongly synchronized oscillatory firing (i.e. at the “fringe of synchrony”). Finally, to prove that gamma bursts variability is shaped by neuronal firing patterns, we succeffully trained a classifier to predict, from different gamma elements’ features, the firing or silence of specific neurons (Fig. 1d). All together, widespread gamma diversity, beyond randomness, may thus reflect complexity, likely functional but invisible to classic average-based analyses.