Optimizing a C collision detection 100x with an LLM

A developer used an LLM to optimize a C collision detection implementation, achieving over 100x speedup from the baseline at a cost of ~125 million tokens. The optimization targeted Kevin Tracy's differentiable collision detection code for convex primitives.

Using an LLM to optimize code: I created a reference implementation of @kevintracy48 https://x.com/kevintracy48 's collision detection in C, then used gpt-5.5 to optimize it and managed a 100x speedup from that baseline. Cost ~125M tokens Code and details:I’m excited to finally show off fast and robust differentiable collision detection between pairs of convex primitives paper: arxiv.org/abs/2207.00669 https://arxiv.org/abs/2207.00669 code: github.com/kevin-tracy/Di… https://github.com/kevin-tracy/DifferentiableCollisions.jl 00:00