jaap van till

300 years of statistics said bigger models should get worse. More parameters means more memorization, more overfitting, more garbage on new data. Then in 2019, researchers pushed past perfect memorization and test error went down again. They called it double descent - the phenomenon that broke classical learning theory and justified every LLM scaling decision on every GPU cluster since. The answer is the lottery ticket hypothesis. A huge network hides countless tiny sub-networks, each with its own random initialization. The number of tickets grows exponentially with size - the bigger the model, the more likely one sub-network finds the simplest solution. Up to 96% of weights can be deleted and performance stays the same. A model trained to predict the next word isn't a parrot - it's searching for the simplest program behind the text. And that program might be a model of the world. 2 minutes. Bookmark & watch today. The reason bigger AI models find simpler solutions.

In 1843, Hamilton realized complex numbers needed not one more dimension but two. One real part, three imaginary - i, j, k. He carved the equation into a Dublin bridge on October 16th. That number system now runs inside every 3D graphics engine, every robotics controller, and every LLM embedding space that maps language into high-dimensional geometry on GPU clusters today. Quaternion multiplication spins two perpendicular circles simultaneously - something that can only happen in four dimensions. i times j is k, but j times i is -k. Rotations don't commute, like a Rubik's cube. That's exactly why game engines and spacecraft use quaternions instead of matrices - no gimbal lock, no numerical drift. The same rotation math that every transformer uses when computing rotary position embeddings to encode token order in the complex plane. 2 minutes. Bookmark & watch today. The number system that powers every 3D engine and nobody taught you about.