We use deep neural network to predict SPX prices using history minute intraday data.The model tested includes: 1.recurrent neural network 2.transformer and its variants 3.convolutional sequence learning using residual network 4.xgboost The conclusion is that: Although deep learning provides an approach to build models with large enough capacities,it does not solve the to problem of ensuring a 'correct' inductive bias,or it does not have the power to build 'theorem' using logic.It does not have the ability to 'reason'.This agrees with what Yoshua Bengio proposed in NIPS 2019,that now deep learning models are just giant curve fitters,with no 'intelligence'. In summary,deep learning, the current theories development,will not surpass human 'intelligence' in the forseeable manner during the next decade.