The 4090 and the Nano
The same 13 CUDA launches cost 4.78 ms on the RTX 4090 and 39.18 ms on the Orin NX — 8.19×, with residency ruled out as the cause — and the first deploy of the stack to the board found a hang, not a mission.
An open notebook
Journal
Motivations, working hypotheses, evidence, and changes of mind.
The mark
The psi monogram as geometry — a 738-byte JSON file, an SVG generated from it, and a Blender pipeline whose two runs are byte-identical — with the flat mark and the 3D mesh unable to drift apart.
The ladder
The neuro-symbolic seam as one number — squared Mahalanobis distance from the predictor's mean, computed with the JEPA crate's own routine — plus a rule table, and a healing ladder that is still only a specification.
The agent in the loop
A braid that stores nothing, a write edge invented because two strands are separate processes, a belief clock that lets the map degrade instead of stopping, and an exploration mission that opens itself — with an uncertainty weight of 0.625 measured on the robot.
The front end, rebuilt from lessons
Five existing front ends read as evidence, the lessons written down before the crate existed, then one egui binary with five views, six snapshot tests and a screen that was actually read back — plus a TUI panel and a headless board that can build the GUI but not display it.
The fly in the belief matrices
A connectome prior loaded, digest-checked, coupled into the belief layers behind an off-by-default feature, and watching its own rate model publish observe-only — seven tickets, a 1.75 coupling total on the 4090, and three CUDA kernels measured at 2.50, 10.59 and 11.84 µs.
The connectome as a prior
Turning a 929,735-edge type-level connectome into a checked, attributed artifact a belief layer can load — built offline, from two Feather files, with a 12-row fixture standing in for the dataset.
The licence and the snapshot
In a clean-room rewrite the licence and the proof of non-copying are the deliverable, not paperwork around it. Apache-2.0, a NOTICE carrying three attributions, and a gate that compared 104 then 110 files against the private reference and found 0 identical.
The male CNS, as arrays
The complete male fruit fly central nervous system — 166,700 neurons, 11,710 types, 45.6 million pre-synaptic sites — is public, fits on one workstation, and reduces to a 929,735-edge graph between types. That is small enough to compute on a GPU and specific enough to steer a robot with.
Tiles the compiler chose
A tile compiler met the hand-written kernels on five FP32 operations. It won one, matched another and lost two — and the losses are exactly where data reuse is highest, which is where layout decides the answer.
The load in flight, and the shape Triton uses
The matrix multiply stopped waiting on its own tile loads — 80.00 to 68.62 µs — and the layer norm was rebuilt in the shape Triton's kernel emits — 10.05 to 8.97 µs. Both changes are the same change: more independent loads in flight.
Twelve generations, one diagnosis
A benchmark that proposes its own kernel variants kept three changes. What it produced was not a faster kernel but a diagnosis about where the time waits — and a hand-written pipeline then answered it, 11% further on.
How the kernel time fell, step by step
Three measured steps took the Rust matrix multiply from 344 to 80 microseconds and layer normalization from 18.6 to 10.1 — and the steps that carried the reduction were a change of layout and a change of work split, not of arithmetic.
Rewriting the layer norm and matmul kernels
Two Rust kernels restructured so values stay in registers and partial answers meet inside a warp, with the measured GPU time before and after.
A faster call can hide a slower kernel
The first profiles changed the question. Before choosing a language, I need to understand which part of the work became faster.
Why this notebook: thinking above the code
AI is opening more directions for me to explore. I want to keep the reasoning, the experiments, and the changes of mind in public.