Member-only story
What a Tensor Really Is
Why Gravity and Your GPU Run on the Same Object
Ask a physics student in the first week of general relativity what a tensor is, and if the course is honest, they will recite the definition the textbooks actually use: a tensor is something that transforms like a tensor. The sentence sounds like a joke at the reader’s expense, or a circular trap with no exit. Ask one of the millions of people who type torch.tensor every day, and they will tell you something that sounds nothing like a joke at all: a tensor is a multidimensional array, a box of numbers with a shape — three of them stacked is a colour image, a few thousand of those stacked is a training batch. Two enormous audiences collide on the same word and walk away holding two different objects, each convinced the other is either overcomplicating something trivial or missing something deep.
Neither picture is wrong. Neither is quite the real thing. A tensor is a geometric object whose components — the array of numbers you actually store — change precisely, lawfully, and predictably under a change of coordinates, so that the object itself does not. That single inversion is the whole subject, and it is the same idea that holds up both Einstein’s theory of gravity and every deep-learning framework running on Earth.
