In a demo video posted by X user Angel (@Angaisb_), 1993's DOOM is running on a processor designed by GPT-5.6 Sol. The chip is called Codex-R32 — a name the model picked for itself.
The stage for the demo is Turing Complete, an educational puzzle game about computer science: players start from NAND gates and build upward, wiring together registers and an arithmetic logic unit until they've assembled a full computer capable of executing instructions. Its sandbox mode offers gate-level simulation, where every logic element has to be wired by hand. Codex-R32 was built from the ground up in this environment — the model didn't just draw the circuit, it also wrote the accompanying assembly code.
The DOOM build running on it is a C-based port called PureDOOM, compiled down to RV32IM machine code and executed directly on the simulated hardware. The visuals are the main reason the demo spread: the game's viewport is overlaid on a live circuit diagram of the CPU, with gates, registers, memory and the ALU lighting up cell by cell in step with each instruction, alongside a live readout of cycle count, simulation speed, and memory and register values.
The Weight Those Four Letters Carry: RV32IM
The choice of instruction set says something about the scope of the work. RV32IM is a subset of RISC-V: 32-bit, with the base integer instructions plus the M extension for multiplication and division. There's a practical reason for picking it — PureDOOM is standard C code that needs a target platform with a mature compiler backend, and RISC-V fits that bill.
But that also means the CPU has to correctly implement a published specification down to the letter. Get the edge-case handling wrong on even one instruction, and the game won't reach the title screen. DOOM does fixed-point math, looks things up in tables, and jumps around constantly to drive its rendering loop — a buggy processor won't survive the first few tens of thousands of cycles. Booting up and staying playable amounts to passing a fairly dense round of implicit testing.
Angel's own comment was brief: this is the DOOM everyone was asking for, running on a CPU that GPT-5.6 Sol built and named Codex-R32 itself, and it's absurd that this is even possible. Someone in the replies asked whether Crysis should be next. Angel let the model answer for itself — sure, just build me a GPU first, throw in a few more gigabytes of memory, and draw a circuit diagram visible from space.
Does This Count as "AI Chip Design"?
Here's where it's worth being precise. There's an entire industry standing between the gate-level simulation in Turing Complete and real chip design. Actually taping out a processor means clearing timing closure, handling clock trees and power domains, doing place-and-route against a real process library, running formal verification and tens of thousands of regression tests, and negotiating a foundry's DRC rules. None of that exists inside a game sandbox — the simulator hands you ideal components and zero-latency wires.
So what this demo proves isn't that a model can replace a chip engineer. It proves something else: that inside a rules-based, instant-feedback environment, a model can take on a design spanning multiple layers of abstraction — gates, microarchitecture, instruction-set implementation, and the assembly toolchain — and carry it through from start to finish while staying internally consistent, where an error at any single layer would surface at the top. That kind of cross-layer consistency has always been one of the hardest things to measure with a benchmark.
A Ritual More Than Three Decades Old
"Can it run DOOM" has been a fixture of the hardware world's inside jokes since the late 1990s. The game has been squeezed onto printers, oscilloscopes, ATMs, digital pregnancy tests, and even a gut-bacteria culture rig. It keeps getting used as a target because its requirements land in exactly the right spot: it needs a processor that can run general C code, a few megabytes of memory, and a refreshable display output — but no floating-point unit and no modern graphics interface. Whatever platform can run it can reasonably be called a real computer.
This time the target being tested is an AI-generated design. Thirty years ago, running DOOM was used to prove a piece of hardware was the real thing. Now it's used to prove a model's output is the real thing. The test hasn't changed. What's being tested has.
Sources: Tom's Hardware, demo videos and notes the creator posted publicly on social media, CocoLoop; verified the CPU name, the instruction-set subset naming, the PureDOOM port version and the simulation environment name.