Clojurians Log v2
clojurescript 2026-10-15
Cheng Lou’s Handwired turns ordinary JavaScript control flow and arithmetic (if-else, loops, inequalities, selections) into an exact, fully differentiable neural circuit (matrix multiplies + ReGLU). No training required for the original behavior; runs in the browser (CPU or WebGPU). Demos include a mini OS, live-trained layout, and millisecond JSON search over 100k documents. This is particularly interesting for the kind of bounded arithmetic inequality / constraint reasoning used in layout systems. Idea: pair it with a modern CLJS interactive environment (Figwheel-style hot reload), expose DOF handles and slack cues on UI objects. Exact constraints stay frozen and valid; selected parameters can be trained or optimized live. Conversation exploring this: https://x.com/i/grok/share/674018a8800c40b28848242546d7de89 OP: https://x.com/willdepue/status/2110588918999707998?s=20
> This is particularly interesting for the kind of bounded arithmetic inequality / constraint reasoning used in layout systems. I'm not in the know, so all I can ask is "why?" And I don't really follow how that's relevant to CLJS, especially the "expose DOF handles and slack cues on UI objects" bit.
DOF = degree of freedom slack = how much latitude remains for a given UI element dimension, under existing constraints So, handles (say, visible upon hover) indicate manipulability, while cues indicate allowed range. I've always preferred declarative constraints to purely procedural space allocation. A simple default slack distribution procedure (e.g., midpoint of next tightest position/length interval) can initialize.