I really like the figure showing the nearest-neighbour graph of MNIST being stretched as it goes through the hidden sigmoid layer. Really helps build an intuition for those of us who think visually or geometrically.
He quotes Bret Victor at the end:
"When Hamming says there could be unthinkable thoughts, we have to take that as “Yes, but we build tools that adapt these unthinkable thoughts to the way that our minds work and allow us to think these thoughts that were previously unthinkable.”"
This is a fabulous article in general, and I love the Bret Victor quote as well.
And while I love the analogy -- and I think is applicable to what we're calling data science -- thoughts, speaking broadly, are fundamentally different than sounds and smells and wavelengths of light. The essence of thought is that we think it (it is thunk?). All of those other examples are subjective representations of physical phenomenon, while constitutes a thought, again broadly speaking, is less well agreed upon (understood?).
Still a terrific article, my pedantic nitpicking aside.
He quotes Bret Victor at the end:
"When Hamming says there could be unthinkable thoughts, we have to take that as “Yes, but we build tools that adapt these unthinkable thoughts to the way that our minds work and allow us to think these thoughts that were previously unthinkable.”"
Great work!