In 1967, graduate student Jocelyn Bell Burnell discovered the first pulsar — a rapidly rotating neutron star predicted by Fritz Zwicky back in the 1930s. These objects are unimaginably dense: a teaspoon of their matter weighs as much as all the cars on Earth. Understanding their internal structure has been hindered by equations with hundreds of parameters. Previously, scientists adjusted them by hand. Now, NNStar, a program with machine learning, does it on its own. Like a kitchen robot, it reads the recipe (the mathematical description of the star), mixes ingredients (types of nuclear forces), sets the mode (numerical simulation), and tastes the dish — checking the model against data from telescopes and the LIGO detector (which catches gravitational waves using ultra-precise laser measurements). The equations for calculations are based on Einstein's ideas about space curvature. The agent figures out how matter is structured inside and how heavy elements like gold are born in the process of nucleosynthesis. If the prediction diverges from observations, it changes the "ingredients." The most surprising skill of NNStar: it extracts formulas directly from PDF articles and double-checks the conclusions, like a chef reconstructing a recipe from a photo. The error stays under 1%, and improving the theory takes minutes. The whole "kitchen" is open for replication.