Neural nets are unstable during training and dynamic weights amplifies the problem. Thus, the neural networks could end up in totally unusable states at inference time.