Open-source visual lab Model-Booster enables LLMs to directly code neural network weights

Open-source project Model-Booster has released a visual experimental environment that uses Claude Opus to directly code weights for small neural networks without gradient descent training. Users can adjust sliders in the browser to blend weights and observe real-time behavioral changes in an untrained multilayer perceptron (MLP) with roughly 170k weights across mini-games like catching, dodging, and maze navigation. Local deployment drives the Claude Code CLI to write weights and execute test evaluations, with each generation typically taking several minutes.