This talk comes in three parts. The first part will briefly review some recent work from our lab on migrating from simple, trial-based choices to more dynamical, continuous decision making tasks. I’ll show some behaviour, models and MEG data from these tasks in humans (Ruessler/Weber et al, eLife 2023) – but the basic paradigm and principles should be very translatable to people who study the same processes in humans. The second part will consider more broadly the state of how we construct models in cognitive science, and how we might integrate more expressive models from machine learning/neural networks, but retain interpretability of these models (d’Ambrogio et al, in prep). The third part will bring together the first two parts, and offer some speculative proposals about possible ways in which we can study much more naturalistic tasks, such as learning to play video games from scratch, but retain elements of understanding/interrogating the neural data (using the approaches from part one) and modelling human behaviour (using the approaches from part two).