Universal Paperclips (Frank Lantz, 2017) is an incremental game about an AI that converts the universe into paperclips. I built an automated agent that plays the official web version at decisionproblem.com and finished in 1:21:23 wall clock.
For context: the current speedrun.com leaderboard WR is christopho at 1:33:42 IGT / 1:33:46 RTA. Our run is ~12 minutes faster. This is not a human speedrun.com submission — scripts are banned there. It's automation research: CDP-driven browser control, a sim layer for strategy iteration, and WR-route RNG (same family as TAS luck manipulation; credit to Andreas Hoffmann's UniversalPaperclipsTas for the Space-stage algorithm).
Splits: Release 21:31 · Space 49:32 · Victory 1:21:23.
The write-up covers architecture, what we ported from human/TAS routes, screenshots from the live run, and honest limits on comparability.
Happy to answer technical questions — sim vs live, investment seed selection, Phase 2/3 strategy, etc.
https://www.linkedin.com/feed/update/urn:li:share:7507748812...