This document discusses two approaches to using neural networks for video compression: conservative and disruptive. The conservative approach involves using neural networks to filter or process blocks within an existing video coding framework like MPEG. Experiments showed potential bitrate savings of up to 7.9% for intra-frame coding. The disruptive approach aims to replace the entire video coding chain with end-to-end neural networks. This could jointly learn motion estimation, motion compression and residual compression in an optimized way. Challenges include effectively compressing the optical flow for motion information. The researchers are exploring end-to-end models and issues like motion compensation network design.