GPT reconstructed an 18-piece torn Tamil note from photos alone
Emergency_Abies_6508 · reddit · 2026-11-15
A Reddit user built two torn-paper puzzles in South Indian languages and gave GPT only photos. For a Telugu note torn into 9 irregular pieces, the model used ruled lines, punch holes, and red-ink continuity across tears to reconstruct it, correctly reading "naaku telugu vachchu" ("I know Telugu").
The harder test: a Tamil passage (opening Thirukkural couplet) cut into 18 fragments, some rotated or upside down, with no lines or holes as clues. GPT leaned on rotation, cut edges, and ink continuity, grouped small clusters first, and improved placements for four loose pieces when asked to rethink. It placed all 18 with rotations plus an evidence report, marking some joins provisional. The author cautions it's not a controlled test — the real takeaway is that the model switched reconstruction strategies based on what clues each page offered.
More from Models
- Grok's Imagine Tab Isn't Dead - It's Hidden for New and Inactive Users — nima_owji · 2026-11-15
- Blogger describes 'RHLF': models that spiral and deny instead of engaging your thesis — StewartalsopIII · 2026-11-15
- It's Been 72 Days Since OpenAI Shipped a Frontier Model, Nearing Its Longest-Ever Gap — banteg · 2026-11-15
- Same UI prompt, three results: ChatGPT vs Claude vs Figma Make — Tegadesigns · 2026-11-15
- Musk says LLM is outdated: a model is just a big file of numbers — haydendevs · 2026-11-15
- Ex-Googler rants: Gemini licensing spans 17+ confusing SKUs across Workspace/GCP — majidmanzarpour · 2026-11-15