“In recent work we have been arguing that predictive processing in the brain is best understood as taking place within a larger agent-environment system (Bruineberg, Kiverstein and Rietveld forthcoming; Bruineberg and Rietveld 2014). We’ve suggested that prediction-error might be thought of as a measure of the disattunement between internal dynamics (on the side of the agent) and external dynamics (in the environment). The process of prediction-error minimisation so-understood takes place within an agent-environment system as a whole with the aim of improving the “grip” of the agent on its environment. Must this process of prediction-error minimisation be described in representational terms? One might think so based on the role of generative models that carry information about long-term stable regularities in the environment. Prediction-error is computed relative to such long-term models of the environment, which look to have the status of representations if anything does. In this talk I will attempt to defend a non-representational interpretation of grip. This will require offering an alternative theory of how generative models work in dynamical, ecological terms.”