ARC2-CHIMERA

GPU compute-shader playground exploring OpenGL-based primitives for symbolic reasoning. Accuracy on ARC is a research baseline, not a competitive submission.

Please read PROJECT_STATUS.md before anything else.

TL;DR: this repository is a research proof-of-concept. Current benchmark accuracy on ARC-AGI is ~2-5% on 100-task subsets — well below the 85% target of the ARC Prize. This is not an ARC Prize submission and should not be described as one. The reusable value is the GPU compute-shader architecture, not the reasoning engine.

What the package is

arc2-chimera exposes the v10 OpenGL-based architecture as an installable Python package so the shader primitives can be reused:

• Matmul / GELU kernels implemented as GLSL compute shaders.
• Cellular-automaton texture update kernels.
• Diffusion-style update passes.
• Holographic-memory read/write passes.
• A GPU-texture state loop (no CPU round-trips between frames).

All of this runs on any OpenGL 4.3+ GPU via moderngl. There is no PyTorch or TensorFlow dependency.

Install

pip install arc2-chimera          # core (numpy, Pillow)
pip install arc2-chimera[gpu]     # add moderngl + glfw for real GPU runs
pip install arc2-chimera[dev]     # pytest / build / twine

CLI

arc2-chimera info                                      # prints status + disclaimer
arc2-chimera evaluate --task-file task.json --output out.json

Repository layout

What works / what doesn't

See PROJECT_STATUS.md for the full list. In short: the GPU kernels work and are worth reusing. The symbolic reasoning layer on top of them does not yet perform on ARC.

License

Apache-2.0. Author: Francisco Angulo de Lafuente agnuxo1@gmail.com.