Hand-built 16-neuron network without training hits 96.0% vs 95.3% for trained original
DragggonZ · x · 2026-10-18
DragggonZ reports identifying each neuron's main job in a 16-neuron model and rebuilding it entirely by hand with Claude's help — no training: 96.0% accuracy vs 95.3% for the ARC-trained original. Attached images show the hand-built weights relabelled and rescaled alongside the original weight matrices, realizing the same function. More details to come.
Related event: Researcher Hand-Rebuilds 16-Neuron Network, Beating Original Accuracy(3 posts)→