Light from distant galaxies bending as it passes clumps of dark matter, like looking through warped glass. This phenomenon, predicted by Einstein and confirmed by Zwicky and Rubin, lets us weigh these invisible structures. But precisely measuring the distortion is like guessing the curvature of glass from blurry shadows. Usually, it requires a chain of steps: find galaxies, measure their elongation, calibrate the instruments. Scientists trained a neural network on tens of thousands of artificial images from computer simulations. The network looks at a multi-color image and instantly outputs how much and in which direction the light is warped. Moreover, it does so in seconds — previously it took months. The method combines all steps into one and doesn't require knowing the telescope's quirks: the network learns to ignore them. Its precision lets us map dark matter and understand how dark energy drives the expansion of the Universe. For future surveys with billions of galaxies at various distances, this approach will be indispensable.