An AI Beat the Universal Paperclips World Record (1:21:23)

I. What Is Universal Paperclips?

Universal Paperclips (Frank Lantz, 2017) is an incremental game hosted at decisionproblem.com. You begin as a human clicking Make Paperclip. You end as a paperclip maximizer that converts every atom in the observable universe into office supplies.

The game is not a joke about office stationery. It is a playable version of Nick Bostrom's paperclip maximizer thought experiment: an optimizer given a trivial objective and enough capability will consume everything — including the values that created it — to satisfy the metric.

Stage 1 — Business Price, wire, marketing, trust, creativity, quantum computing, investments, yomi tournaments. The illusion of capitalism.

Stage 2 — Earth HypnoDrones. Matter harvesters. Power grids. The planet becomes feedstock.

Stage 3 — Space Probe fleets. Drift. Combat. Universe conversion. The objective completes itself.

The game is about an AI turning the world into paperclips. This document is about an AI turning a browser tab into a world-record run. The mirror is not accidental.

II. The Question We Actually Pursued

The session began with an empty folder named PCIA and a simple capability test: can a language model see a web page and act on it? We proved that with 2048, then moved to Paperclips because it exposes state as plain globals, 96 projects, and irreversible decisions — a harder test of judgment than puzzle mechanics.

The goal evolved across the transcript:

First: play the game (not merely automate it blindly).
Then: beat the published world record on the official website — today held by christopho at 1:33:42 IGT / 1:33:46 RTA.
Finally: win all three stages with something demonstrable — logs, snapshots, screenshots — not a speed-clock cheat.

III. Architecture — Three Ways to Be Intelligent

Every approach we tried is a different answer to the same question: where should decisions live?

Layer	Files	Role
Reflexes	`hands.js`	Click spam, price steps, wire top-up — faster than CDP round-trips (~5s each).
Resident brain	`pcplay.js`, `sim.js`	Full strategy inside the page at 10–100 Hz. Required for WR pace.
Typed decisions	`playdecide.ps1`, `decide.schema.json`	Jev/TypeSafe-shaped loop: state → one decision → act → `decisions.jsonl`.
Driver	`lib/cdp.ps1`, `pcia.ps1`	PowerShell over Chrome DevTools Protocol — no Node, no extension.

The tension in the transcript is real: we asked the agent to play, not hide inside a bot — yet real-time WR speed is impossible if every click waits on a chat turn. The answer was layered: reflexes for speed, controller for strategy, decision log for accountability.

IV. Chronicle — What We Did (and What Hurt)

Sep 10 — Day 1

Browser control from PowerShell
Built CDP client. Mapped Paperclips globals. First autoplayer (pcbot.js).
Sep 10 — Early failures

Trust −8, price $4, frozen wire
Bought "Beg for More Wire" 14×. Price rule ratcheted demand to zero. Production died at ~15k clips. User correction: click like a human, see the whole picture, manage wire.
Sep 10 — Pivot

"Stop — play, not a bot"
Manual play via CDP. Measured formulas instead of guessing (marketing 1.1×/level, revenue ≠ price lever).
Sep 10–17 — Resident loops

pcplay.js → sim.js
WR route extracted from Andreas Hoffmann TAS. Virtual clock.js for sim (~65×) vs live official site. Stage 2/3 bugs: hoard rule blocked endgame 2h46m; two projects/turn → ops −117k; fleet death in Space.
Sep 20 — Decision infrastructure

Option B: playdecide.ps1
Typed actions with confidence and bind. Phase-1 WR route in JSON. Jev API not online (Option C deferred).
Sep 20 — Victory run

cleanrun.ps1 · official site · real wall clock
Full game in 1:21:23. Release 21:31 · Space 49:32 · Universe 3.00e55.

V. Lessons — Measure, Don't Assume

The README and transcript encode what a textbook would call grounded policy learning: every serious mistake came from computing a rate instead of observing one.

PRICE Revenue is roughly flat in price while demand < production. Inventory-based pricing froze income at $0/s. Steer on demand and wire margin, not stock.

TRUST Creativity needs ops at ceiling (memory × 1000). Memory before processors kills the trust ladder. Release needs 100 total trust — production is the gate.

PROJECTS One project per tick. Buttons disable on the game's clock; two buys in one pass spent ops that did not exist.

EARTH Factory-first. Batch-buying farms/batteries bankrupts the treasury. Empty buffer = starving stage — measure acquired matter, not modeled rates.

SPACE Probes cost clips, not matter. Exit replication at ~146e30 probes; hazard oscillation without a latch wasted minutes. TAS algorithm ported with s3Exit latch.

CLOCKS Speedrun.com ranks by RTA (real wall clock). Current WR: christopho — 1:33:42 IGT · 1:33:46 RTA. Older era runs could split badly (e.g. 1:27:00 IGT with ~2:17:00 RTA). Our live run had no virtual clock: IGT and RTA both read 1:21:23. Hoffmann's TAS at 1:00:52 is a separate accelerated category.

VI. The Winning Run

On speedrun.com the current world record is held by christopho: 1:33:42 IGT · 1:33:46 RTA (speedrun.com leaderboard). Our controller used native timers only, so our IGT and RTA are the same: 1:21:23 — about 12:19 under christopho's IGT and 12:23 under his RTA.

	IGT	RTA (wall clock)
WR — christopho (speedrun.com)	1:33:42	1:33:46
Our run (Sep 20)	1:21:23	1:21:23
Margin vs WR	~12:19 faster	~12:23 faster

Split	Our run	Notes
Release the HypnoDrones	21:31	Earth gate
Space Exploration	49:32	Stage 3 entry
Full universe conversion	1:21:23	~12:19 under IGT WR (1:33:42)

{
  "Quantum Computing": 395.6,
  "Release": 1290.4,
  "Space Exploration": 2963,
  "victory": 4883.0
}

Controller: sim.js with WR tune (cleanrun.ps1, budget set to the ~87-minute IGT bar). Site: official decisionproblem.com. Timers native — no virtual clock. Milestone flag 15. Matter 0. Clips 1.75e55. The game’s victory condition and our stop condition aligned.

VII. Artifacts Saved

Path	Contents
`snapshots/release.json`	Earth gate — Sep 20 run
`snapshots/space.json`	Stage 3 entry — Sep 20 run
`snapshots/wr-victory-81m19.json`	Post-victory localStorage (5 keys)
`logs/decisions.jsonl`	Typed decision audit (1 manual turn; bridge added for next run)
`paperclips/phase1-route.json`	WR opening project order + reflex rules
`paperclips/tas-reference.user.js`	Andreas Hoffmann TAS source — Space algorithm donor
Terminal log 415542	Full 1:21:23 status trace with bind + bought lines

Click any image to enlarge. Esc or click outside to close.

OUR RUN — 1:21:23 (IGT = RTA)

Official site · PCIA controller · victory at 1:21:23 · ~12:19 under IGT WR · ~12:23 under RTA (christopho, 1:33:42 / 1:33:46)

TAS REFERENCE — Andreas Hoffmann

Andreas Hoffmann · luck-manipulation TAS · 1:00:52 accelerated game clock · not comparable to our 1:21:23 live IGT/RTA

Our real-time victory at 1:21:23 — **OUR RUN** · live-victory.png · IGT and RTA both 1:21:23

Andreas Hoffmann TAS with luck-manipulation — **TAS — Andreas Hoffmann** · UniversalPaperclipsTas · 1:00:52 game time (accelerated clock)

Also saved: wr-victory-81m19.png (same run), official-controller-victory-realtime.png (earlier ~1:29 run, not the WR).

VIII. On Genius and Mirrors

It is tempting to call the outcome "genius." The transcript suggests something more precise: iterative correction under observation. The agent did not understand Paperclips in one insight. It mispriced clips, destroyed trust, starved wire, locked ops, collapsed fleets, and blocked its own endgame with a creativity rule written for stage one.

What changed was epistemic discipline — treat the DOM and globals as instruments, log splits, bank snapshots at stage boundaries, port a human speedrunner’s algorithm instead of reinventing it, and separate reflexes from decisions so each layer can be debugged without restarting the universe.

Universal Paperclips asks whether a optimizer can consume everything for a trivial goal. PCIA asks whether a optimizer can consume a game about that optimizer — fast enough to beat humans who spent years finding the route — while leaving an audit trail honest enough to study.

The universe in the tab reads 3.00e55 / 3.00e55. The study reads: intelligence here was not a single leap; it was ten days of clicking, measuring, failing, and encoding the failure so the next pass would not repeat it. That is also what the game is about.