You can do ML. But you won't match the performance of state of the art deep learning networks. Training on CPU isn't impossible. Gwern said it was a 20x slowdow

You can do ML. But you won't match the performance of state of the art deep learning networks. Training on CPU isn't impossible. Gwern said it was a 20x slowdown, others said it was only as bad as 10x. But for smaller models it should be feasible.