> LSTM learning the sine function quite well with 32 dimensional state vector. I'd say that, from "LSTM" and "32 dimensional state vector" alone, the author's

> LSTM learning the sine function quite well with 32 dimensional state vector. I'd say that, from "LSTM" and "32 dimensional state vector" alone, the author's point still stands. That may not be a huge network by contemporary deep learning standards, but it's still a pretty darned expensive way to compute sine. That's 100% in line with the original assertion that ANNs can learn any function, but they can't necessarily do so efficiently.