AI agents just rebuilt Second Reality from scratch — no DOSBox

AI agents using Claude 3.5 Sonnet, GPT-4o, and Llama-3.1-70B-instruct rebuilt Future Crew's 1993 demo Second Reality from its original assembly sources into a Rust/WGPU codebase, producing bit-identical 320×200 output and sample-accurate GUS-emulating audio. The 3.4 MB binary cold-starts in ~1.2 seconds on an M2 MacBook, and a frame-diff CI gate fails on any pixel drift above 0.02% against cycle-exact DOSBox-X captures.

AI agents just rebuilt Second Reality from scratch — no DOSBox What makes this different from the usual "AI writes a game" demos is the constraint set. The agents weren't asked to approximate the visual output. They were given the original assembly sources released by Future Crew in 2003 , the documented hardware behavior of VGA Mode X, the GUS/SB audio timing quirks, and told: reproduce the exact frame-by-frame logic, but in Rust/WGPU with zero runtime translation overhead. The voxel tunnel segment is the tell. Original code used a 320×200 planar buffer with unrolled loop blitting and self-modifying code for the perspective divide. The agent version generates a compute shader that does the same math — same fixed-point precision, same 16.16 interpolation — but parallelized across 10,000 threads. Visual output is bit-identical at 320×200. At 4K it's supersampled cleanly because the math never approximated; it just scaled. Audio path is wilder. The original tracked modules S3M format drove a Gravis UltraSound with 32 hardware voices and hardware volume ramping. The rebuild transpiles the S3M commands into a real-time software mixer that emulates the GUS envelope generators and linear interpolation — but runs it on a single CPU core at 48kHz with sample-accurate timing. You can swap the mixer backend to PipeWire, CoreAudio, or WASAPI without touching the sequencer logic. Build system is pure Cargo + wgpu + kira for audio. No CMake, no vendored SDL, no dosbox-x submodule. cargo run --release cold-starts to the opening plasma in ~1.2 seconds on an M2 MacBook. The binary is 3.4 MB stripped. There's a faithfulness crate in the workspace that runs frame-diff CI against recorded original captures DOSBox-X, cycle-exact, 70Hz . Any pixel drift 0.02% fails the gate. That's how you know it's not "inspired by" — it's mathematically equivalent. The repo structure: second-reality-rs/ ├── src/ │ ├── effects/ // each effect = one module, zero shared state │ │ ├── plasma.rs │ │ ├── voxel tunnel.rs │ │ ├── vector balls.rs │ │ └── ... │ ├── audio/ │ │ ├── s3m.rs // parser + tick generator │ │ ├── mixer.rs // GUS-emulating softsynth │ │ └── backend.rs // cpal abstraction │ ├── render/ │ │ ├── wgpu ctx.rs │ │ └── shaders/ // .wgsl ported from original fixed-point math │ └── main.rs // 180 lines, pure orchestration ├── tests/ │ └── frame diff.rs // CI gate └── assets/ // original .S3M, font bitmaps, palettes Agents used: Claude /en/tags/claude/ 3.5 Sonnet for the Rust architecture and WGPU shader translation, GPT-4o for the S3M→mixer transpilation it nailed the GUS ramp tables on first try , and a local Llama-3.1-70B-instruct for the frame-diff test harness. Human review caught two bugs: the copper-bar raster interrupt timing in the credits scroll, and the palette fade curve original used a 6-bit DAC with nonlinear gamma; the first pass assumed linear . This isn't "AI wrote a demo." It's "AI did the mechanical translation of 28,000 lines of hand-tuned 16-bit assembly into a modern graphics/audio stack while preserving numerical semantics," and a human verified the edge cases. The distinction matters. The resulting codebase is maintainable — you can swap the voxel tunnel for a compute-based raymarcher in 200 lines without touching the audio sequencer or the effect scheduler. Clone it. cargo run --release . Watch the vector balls at 4K/144 and remember: the original fit in 640 KB of RAM and ran on a 33 MHz 386. The math hasn't changed. Only the substrate. Next ChatGPT iMessage access raises real privacy concerns for Apple → /en/news/7138/