New recording technologies and population analyses have made it increasingly conceivable to functionally dissect large-scale circuits in vivo, causally relating neural activity to behavior. Such direct testing typically relies on hypotheses formulated in advance of the experiment. However, preconceived hypotheses may be restricting our exploration of ever-larger neural systems and more complex behaviors. We propose that adaptive experimental designs give us the statistical efficiency to search large parameter spaces and the flexibility to change our models with evolving dynamics. We thus developed a new method for building models of population-level neural dynamics online, while the experiment is running. Our method combines fast, stable dimensionality reduction with a soft tiling of the resulting neural manifold, allowing dynamics to be approximated as a probability flow between tiles. It can be fit efficiently (rates faster than data acquisition), scales to large populations, and outperforms existing methods when dynamics are noise-dominated or feature multi-modal transition probabilities. We demonstrate its performance on both simulated nonlinear dynamical systems and experimental neural data. Using online modeling, we can also close the loop in visual stimulation experiments performed in larval zebrafish, using real-time interventions to both generate and test hypotheses. While displaying thousands of unique combinations of multidimensional visual stimuli is infeasible, we use ideas from Bayesian optimization to sequentially choose maximally informative stimuli, allowing us to rapidly characterize the preferred stimulus for each neuron. We additionally present a new population optimization method using multi-output Gaussian processes that couples online model fitting and active stimulus selection to acquire data at locations where models are likeliest to be wrong given the data seen so far. These methods, which combine online neural modeling with adaptive intervention, open the door to automated, theory-driven circuit dissection at scale, providing a powerful new means of interrogating neural function.