An artificial intelligence specialist known by the pseudonym Angel presented a demonstration of the classic shooter Doom running on a processor architecture entirely created by the GPT-5.6 Sol language model without the use of any ready-made technical solutions. The circuit designed by the neural network was designated Codex-R32 and was implemented inside the virtual environment of the educational puzzle game Turing Complete, available on the Steam digital storefront.
The demonstration video shows classic gameplay overlaid on a visualization of the working processor circuit, where logic gates, registers, memory modules, and arithmetic units activate in real time. Technical parameters are displayed alongside: a clock cycle counter, simulation speed, and current values in memory cells and registers. The Turing Complete simulator allows users to build computing systems literally from basic transistors and logic gates, while the free sandbox mode imposes no restrictions on the complexity of the architectures created — which explains the scale of the presented circuit.
The technical implementation is based on the PureDOOM port, written in the C programming language and representing an adapted version of the original shooter for running on non-standard platforms. This code is compiled into the native machine format RV32IM — a 32-bit instruction set of the RISC-V architecture with an integer base module and an extension for multiplication and division operations. Crucially, this is not emulation on top of existing hardware, but the direct execution of real machine code on a circuit built from scratch.
The performance of the resulting construction corresponds to the specifics of simulating logic elements in a virtual environment and does not suggest practical suitability for actual gameplay. However, the value of the experiment lies in confirming the correctness of code execution, the successful initialization of the application, and the rendering of graphical frames. The author of the experiment described the result as extraordinary, noting the language model's ability to design not only software but also hardware architectures.
The 1993 shooter Doom has long acquired the status of a universal benchmark for testing the computational capabilities of a wide variety of devices — from oscilloscopes and medical tests to Lightning-to-HDMI adapters. In this case, the neural network's ability to create processor solutions served as an analogous evaluation criterion. In response to the traditional question about running the demanding game Crysis, the neural network replied with humor, noting the need to first build a graphics processor, add several gigabytes of RAM, and construct a circuit visible from orbit.