Archive
Every post, newest first. Reading in order instead? Start from the series or the map.
2026
- I Removed Every MLP from Gemma 4 12B
- A Network Made of Parts Kernel Instruments · 4
- How to Interrogate a Kernel Network Kernel Instruments · 3
- The Concept That Would Not Die Kernel Instruments · 2
- The Trained Network, Under Mercer's Microscope Kernel Instruments · 1
- How Many Random Neurons Buy a Trained One? Prototype Network · 9 JAX companion Lazy Training a Yat Network in JAX/Flax NNX
- The Geometry of Attention Is a Choice of Kernel Attention Kernel · 6 JAX companion Measuring Attention's Geometry in JAX/Flax NNX
- The Kernel Between the Roles Attention Kernel · 5 JAX companion Softmax-Free Attention in JAX/Flax NNX
- Depth on Demand Integrators · 5 JAX companion An Error Controller for a Trained Net, in JAX
- Backprop Without the Memory Integrators · 4 JAX companion Reversible Backprop as a custom_vjp in JAX
- A Network Built from Hamiltonian Steps Integrators · 3 JAX companion Building the Hamiltonian-Step Net in JAX/Flax NNX
- The White-Box Survival Model on Trial Prototype Network · 5 JAX companion Running the Survival Trial, in JAX/Flax NNX
- Transformers With a Velocity Ledger Integrators · 2 JAX companion A Velocity Ledger for Transformers, in JAX/Flax NNX
- One Kernel Family, Fitted Two Ways Prototype Network · 8 JAX companion Solving It and Descending It, in JAX/Flax NNX
- When 80% Should Mean 80% Prototype Network · 4 JAX companion Calibrating a Bounded Net, in JAX/Flax NNX
- How Far Down Can You Build? Prototype Network · 3 JAX companion Building the Second Layer by Hand, in JAX/Flax NNX
- Distillation Is a Geometry, Not an Answer Key Representation Geometry · 6 JAX companion Distillation as Kernel Transfer, in JAX/Flax NNX
- Edit One Operator, Edit Every Depth Prototype Network · 7 JAX companion Editing a Deep Equilibrium Network, in JAX/Flax NNX
- Your Skip Connection Is Half of Newton Integrators · 1 JAX companion Skip Connections With Inertia, in JAX/Flax NNX
- Your Network Is a Stack of Layers. It Could Be a Fixed Point. Prototype Network · 6 JAX companion A Network That Is a Fixed Point, in JAX/Flax NNX
- The MLP Block Can Be a Kernel Memory Kernel Weights · 4 JAX companion A White-Box Kernel FFN in JAX/Flax NNX
- What Can a Weight Be? Kernel Weights · 3 JAX companion A Kernel's Price List, in JAX
- Where Does a Weight Live? Kernel Weights · 2 JAX companion Where a Weight Lives, in JAX/Flax NNX
- How Much of a Fashion-MNIST Network Can You Build by Hand? Prototype Network · 2 JAX companion Constructing the Fashion-MNIST Network, in JAX/Flax NNX
- Your Network Is a List of Pictures. You Can Edit It. Prototype Network · 1 JAX companion Editing a Network by Hand, in JAX/Flax NNX
- What a Finite Kernel Buys an MLP Attention Kernel · 2 JAX companion The Yat-Kernel MLP in JAX/Flax NNX
- The Three States of Information Representation Geometry · 5 JAX companion The Three States of Information, in JAX
- Latent on the Spectrum: Why Cats Sit Closer to Dogs Than to Cars Representation Geometry · 4 JAX companion Latent on the Spectrum, in JAX
- Why Attention Needs Q and K Projections Attention Kernel · 4 JAX companion Q and K Projections in JAX/Flax NNX
- The Readout is a Convex Combination of Prototypes Kernel Weights · 1 JAX companion The Prototype Readout in JAX/Flax NNX
- What Makes a Good Latent Space? The Welch Bound and the Simplex Representation Geometry · 3 JAX companion Auditing Latent Space Geometry in JAX
- Cheap Attention: Linear-Time Kernel Approximation Attention Kernel · 3 JAX companion Cheap Attention in JAX/Flax NNX
- Untangling the Moons: A Visual History of Contrastive Learning Representation Geometry · 2 JAX companion Organizing Randomness: Contrastive Learning in JAX
- What Attention Weights Can Explain Attention Kernel · 1 JAX companion Self-Attention as Kernel Regression in JAX/Flax NNX
- What Activations Do to Geometry Representation Geometry · 1