We played around with neural ODEs a bit earlier this year but found them very slow to train and quite unstable, really excited about the new research direction!