Simon Crowe
LightNews LightNews
simon-crowe.bsky.social
Simon Crowe
@simon-crowe.bsky.social
42 followers 180 following 7 posts
Software Engineer -- Python / Platfom. Tinkering with Rust, Go, Electronic music.

simoncrowe.blog
https://fosstodon.org/@simoncrowe
Posts Replies Media Videos
simon-crowe.bsky.social
This weekend project is getting interesting now.

#webgl #wasm #vibecoding #ai #simulation
A screenshot of a full-page WebGL canvas on a web page served from my local machine. The canvas shows some sinuous red and purple organic forms twisted around each-other and receding into a dark blue background. To the top left of the canvas is a heads-up-display with some settings to control a reaction-diffusion simulation as well as several parameters controlling graphics.
January 11, 2026 at 10:55 PM
simon-crowe.bsky.social
And I was right. GPT 5.2 was fine with complex graphics maths but it tripped up with basic refactoring. This is a relief as it's just as stupid as I am in this area. Rather than wasting API credits I just ran `sed -i 's/positionition/position/g' wasm/src/meshing.rs`. #aicoding #vibecoding
error[E0609]: no field `positionition` on type `Vertex` --> wasm/src/meshing.rs:37:37 | 37 | let ib = self.push_vertex(b.positionition, b.normal, color); | ^^^^^^^^^^^^^ unknown field | = note: available fields are: `position`, `normal`
January 11, 2026 at 2:47 AM
simon-crowe.bsky.social
And there's a screen grab for the visually inclined.
A 3d scene displayed on a webgl canvas: A number of blue blob-like volumes are arrayed against a dark background.
January 10, 2026 at 8:02 PM
simon-crowe.bsky.social
I'm kind-of glad to be late to the "vibe coding" party. The current batch of models is quite impressive. I fed it some LLM-generated design documents and it produced a working 3D reaction-diffusion simulation in almost one shot.

#ai #vibecoding #wasm #webgl
Some tiled terminal widows. The window on the right shows some computer graphics focused rust code the defintiion fo a function called ... The left-hand terminal shows the OpenCode agentic coding tool, displaying a message from CPT 5.2: That behavior is mostly “simulation is collapsing to a near-uniform state”, plus two visualization choices that make the end-stage look especially weird: - HUD V[0.000,0.000] was likely just rounding hiding that V had a tiny remaining range. - The viewer was using an adaptive isovalue iso = (vMin + vMax) * 0.5 (web/main.js:375 before), so as the scalar range collapses toward ~0 it keeps chasing it. That produces: - very “blocky” surfaces (iso near min/max, lots of near-degenerate crossings), - almost-zero gradients → unstable/flat normals, - and eventually no surface once the field becomes effectively constant. I made two concrete fixes to separate “rendering artifact” from “sim actually went uniform”: 1) Fixed isovalue + more informative HUD - Mesh generation now uses a constant iso = 0.5 (web/main.js:375 and web/main.js:447). - HUD now prints 6 decimals and dV (range) so you can see when it’s really flat vs just rounding (web/main.js:437). With this, if the sim truly decays toward V≈0 everywhere, the mesh will cleanly disappear (because nothing crosses iso=0.5). That’s expected, and it tells you the simulation reached the homogeneous equilibrium. 2) Stop “normal flipping” from degenerate gradients - Gradients are now stored unnormalized (raw central differences), and the shader continues to normalize per-fragment (wasm/src/meshing.rs:62, wasm/src/meshing.rs:76). This avoids the previous behavior where near-zero gradients turned into an arbitrary fallback normal, which can look like normals “flip” or go flat/blocky as the field gets smooth. Also - Perlin seeding now uses v_amp = 1.0 so iso=0.5 can actually produce a surface (web/main.js:335).
January 10, 2026 at 7:47 PM