HackerStreak maps neural loss landscapes

A useful teaching tool, but not a truth machine: it makes loss geometry legible without pretending 2D and 3D slices fully explain million-parameter optimization.

• –The strongest value is intuition-building, especially for people who have only seen static contour plots and need to understand why those analogies can mislead.
• –Filter-normalized visualizations are the right choice here because raw weight scaling can create fake flatness and distort comparisons between models.
• –The ability to sweep architectures from tiny MLPs to ResNet-8 and LeNet-5 gives it real educational range, not just a toy demo.
• –The main limitation is structural: any low-dimensional projection can invent surfaces and relationships that do not exist in the original parameter space.
• –Best use case is debugging mental models and explaining optimization behavior, not making hard claims about generalization on its own.