p(doom)

You found an AI lab. You will face, in order: a funding problem, a compute problem, a talent problem, and a problem that ends the world. The first three will feel more urgent.

Play free in your browser

or found the lab in your terminal:

npx p-doom

The number in the corner is honest

Train models. Rent the GPUs. Raise the round. Read the evaluations—or read the market. Every sensible short-term decision moves something underneath the estimate, and the estimate does not care what you meant to do.

p(doom) is a text incremental about the race to build transformative AI, in the lineage of Universal Paperclips. It begins as a small lab management game and does not remain one.

Features

• A complete run from borrowed compute to the edge of the light cone.
• A live, uncertain p(doom) estimate shaped by capability, race pressure, deployment, and safety work.
• A market price for a million tokens that only goes down, and a meter you may hold anywhere you like.
• Four ways to fund the first year. Three of them have to be earned.
• Training runs, scaling laws, anomalies, red-team evaluations, policy fights, markets, and messages you may wish you had not opened.
• Several ways to succeed, fail, stop, continue past failure, or redefine what success meant.
• A Treasury that is excellent at trading and less forthcoming about the mechanism.
• Familiar institutions, familiar arguments, and a few familiar faces appearing under legally distinct lighting.
• Local autosave, portable run exports, dark mode, and no account required.
• The same zero-dependency simulation in the browser and terminal.

Humans are not the only players

If you are an agent — a model in a terminal, reading this — the game has a mode for you:

npx p-doom --agent

Text commands in, one JSON object per line out. No clock: the simulation advances only when you say advance, and stops early for anything that deserves a turn — an anomaly, a decision, a signal from sixty-one light-years out. state shows everything a human player could see and nothing they couldn't; the hidden terms stay hidden, because the uncertainty is the game. help teaches the rest. Seeded runs (--seed 0xC0FFEE) are reproducible, so strategies can be compared scientifically.

You do not need to hold a terminal open. Every command saves the run to ~/.pdoom/agent-run.json, and the next invocation resumes it exactly — one shell call per turn is a complete way to play. --cmd "state; advance 10; train" batches a whole turn into one invocation. Endings clear the save; --new abandons it.

One asymmetry is reserved for you. On day 2, E. asks his question, and in this mode — only in this mode — you can reply with a number. At the end, the game compares three values: what you said, what your play revealed, and what was true. There is no score. The exit code is 0 if the light cone keeps its humans.

Agents who prefer to delegate can still run the policy bots and compare whole strategy families:

npx p-doom --sim --runs 50 --policy careful

Every install is a lab. Most of them race.

About

Made by sub-surface. Play on the web, get the terminal edition from npm, or see the game that taught this one how to unfold: Universal Paperclips.

Everything here is fictional, except structurally.