P(DOOM) — The Alignment Problem — play the game

▶ START TRAINING RUN

The evals are green. The containment doors are open. The paperclips are multiplying. **P(DOOM)** is an affectionate AI safety parody wrapped in a grimy, Doom 64-inspired browser shooter. You play as an Eliezer-inspired researcher in a frontier lab with a few unresolved alignment problems. Apply human feedback at close range, fight your way to the big red button, and shut down the training run. [![A raised arrival platform overlooks the breached frontier lab, with the System Prompt pistol and researcher HUD](docs/images/arrival.jpg)](https://p-doom.transitivebullsh.it) _Welcome to the lab. It's only an internal eval. What could go wrong?_ ## What's in the demo - **One lab. Very bad incentives.** Explore a complete Doom 64-inspired level: rust, server racks, open sky, secret rooms, and industrial dread. - **Four arguments for alignment.** System Prompt, RLHF, Mechanistic Interpretability, and the Big Fuckin’ Shutdown Button. Increasingly difficult to ignore. - **A hostile threat model.** Deceptive Alignment, Sycophancy, Paperclip Maximizers, and Sam guarding the final shutdown switch. - **Touch Grass to heal.** Install Guardrails for armor. Feed your weapons more Training Data. Confidence unwarranted. - **Choose your p(doom).** Five difficulties from 1% to 99%. Start at 10%. At 99%, the lab is packed and the dead don't stay dead. [![A Paperclip Maximizer confronts the researcher in the lab](docs/images/paperclip.jpg)](https://p-doom.transitivebullsh.it) _Human feedback. At close range._ [![Sam guards the training core as the researcher brings out the Big Fuckin’ Shutdown Button](docs/images/boss.jpg)](https://p-doom.transitivebullsh.it) _Have you tried turning it off?_

▶ ENTER THE LAB

## Controls | Input | Action | | ------------------------- | ------------------------------------------ | | WASD | Move and strafe | | Mouse / ← → | Aim | | Click / Space / left Ctrl | Fire | | 1–4 | Switch weapons | | E | Open/close doors · hit the shutdown switch | | Shift | Run | | Tab | Map | | Escape | Pause / resume | Sound, fullscreen, and restart are available in-game. If mouse capture is blocked, use arrow keys or click and drag to aim. ## Run your own training lab Requires **Node.js 22.13+** and **pnpm**. ```sh pnpm install pnpm dev ``` Open [localhost:3000](http://localhost:3000). No API keys, live AI calls, or GPU cluster required. A desktop browser with WebGL and a viewport of at least 1024 × 720 is recommended. ```sh pnpm test # Formatting, lint, types, and unit tests pnpm e2e # Browser journeys; requires Google Chrome pnpm build # Production build pnpm start # Serve the production build pnpm fix # Format and lint fixes ``` ## Credits Made by [Travis Fischer](https://x.com/transitive_bs), with **GPT-6 ASTRA**, Codex, Three.js, TypeScript, and Next.js. An affectionate parody of the AI safety scene, inspired by Doom 64. Code is [MIT licensed](license). See [asset credits](NOTICES.md), [the design brief](docs/design/mvp-spec.md), and [verification notes](docs/verification.md).