Conventional climate models must treat clouds statistically because their numerical grid is too coarse to resolve them. Predictions of global warming of a few degC for a doubling of CO2 have not changed for half a century since the first pioneering modelling studies by Manabe and Wetherald (1967) and Manabe and Stouffer (1976), despite the increasing complexity of global climate models. A key issue is that the range of these predictions from various global models has not changed (2 to 4 K). This uncertainty has been traced to treatment of clouds and specifically to deep convective clouds. The climate system consists of many feedbacks, both positive and negative, the sum of which determines the climate sensitivity to a given forcing (e.g. 2 x CO2). Two decades ago, to reduce the uncertainty there was the idea of embedding a high-resolution cloud model inside each gridbox of a global model. Since the cloud model domain would span only a small fraction of each gridbox, computational expense would be limited. However, this ‘super-parameterization’ (SP) approach was nevertheless expensive [1]. More recently, a new idea of compressing the cloud model with an AI neural network was attempted [2]. The presentation describes a fresh approach to create an AI representation of clouds in a global model, with use of a temporal convolutional network (TCN) to replace the cloud model. A TCN is a type of deep learning network, designed to be applied to time series of data [3]. It can treat the time evolution of natural systems. A time-window slides along the time series. To predict the system at any given time, with neurons in each layer are assigned to all times in the window. In any call to the TCN, information propagates only forward in time through the layers of neurons to the output layer. Here, the full expensive SP global model is used for training of the TCN. Results from off-line testing of the scheme are shown and various approaches for training are illustrated.