Probing ChatGPT's Memory: Asking It to Guess Shortcuts for a Game It Never Played
saijanai · reddit · 2026-09-15
A Reddit experiment asked ChatGPT to guess command names and keyboard shortcuts for Pax Romana without web access. One version honestly flagged its output as low-confidence convention-based reconstruction—a 'useful negative result'—while another corrected the game to Anno 117 and produced plausible WASD-style bindings, illustrating how models separate retrieved facts from convention-based guessing.
More from Models
- GLM 5.3 Flash EXL3 adds TP=3 mode: 28% faster decode across three DGX Sparks — HankYeomans · 2026-09-15
- Occamy-1.0 Open-Sources a 35B Cost-Efficient Agent Model for Co-work Tasks — Accio-Lab · 2026-09-15
- Meta's Muse Voice Transcribe tops Pipecat benchmark with lowest semantic WER for voice agents — bowenc0221 · 2026-09-15
- ChatGPT Voice Mode Wows User With Tornado Re-enactments and Black Hole Fall Simulations — VoidStateKate · 2026-09-15
- Doubao launches consumer phone agent; Anthropic reportedly targets Nasdaq IPO in October — APPSO · 2026-09-15
- Opus 5.2 rumor prompt works only in Claude Code, not on web or app — CtrlAltDwayne · 2026-09-15