Andrej

The complexification of the engine is worrying in this case. I haven't fully thought it through yet but I'd wish to support this only if the changes are very minimal,...

"and that would run faster" I think that ship has sailed a long time ago :) I want to emphasize cleanliness and interpretability over everything else. Let me stare at...

Ok preliminary thoughts: - advise we get rid of flags and make the above behaviors default - did you perhaps mis `__pow__` changes? Still uses `self.data`

Ok, hmm I still quite dislike that we have to `x.grad = 0` explicitly.

Also using `y = x*x*x` instead of `y = x**3` now gives an error

yes, similar to PyTorch, a good feature request. I'll see if I can find a way to do it without sacrificing the clarity, cleanliness and compactness.

because pow currently requires the second argument to be an int/float.

I appreciate the effort, I reviewed the full diff, but I don't like it. But I do like the `itertools.cycle` call, which I was not originally aware of and I...

Actually on closer look I'm a bit worried about [itertools.cycle](https://docs.python.org/3/library/itertools.html#itertools.cycle), as it caches the individual elements in memory to yield through them on second pass. I am worried this would...

Thank you for the PR, I will be looking at adding words next as well and will reference it.