A Lisp coding agent is beating Codex in token efficiency by 10x (twitter.com)

A new AI coding agent, NODECODE, has been demonstrated to significantly outperform Codex in token efficiency, achieving results that are 7-14 times more efficient while also completing tasks 2.5 times faster. In a head-to-head testing scenario where both agents were tasked with creating a playable 3D pelican kart racer, NODECODE not only required fewer input tokens but also finished the project well ahead of Codex. However, while it performed faster and more efficiently, the final product was simpler and less polished than that produced by Codex. NODECODE stands out as an open-source agent designed to rewrite itself as it works, utilizing LISP—the AI-native programming language. This innovation reflects a vision from AI pioneer John McCarthy, enabling a self-evolving program that can iteratively improve through user interactions. The implications of NODECODE's capabilities extend beyond mere performance metrics; it signifies a shift toward more adaptable and efficient coding agents within the AI/ML community, encouraging further exploration of self-improvement mechanisms in AI development.