| ▲ | How an AI agent beat Universal Paperclips in 1:21:23 (WR was 1:33:42)(universal-paperclips-ai.netlify.app) | |
| 2 points by arthur-G a day ago | 1 comments | ||
| ▲ | arthur-G a day ago | parent [-] | |
Hi HN , 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... | ||