OCR is not solved: SOTA model reads one 1881 line, outputs 16,421 dots
vanstriendaniel · x · 2026-10-26
Daniel van Strien shows OCR is far from solved: he asked an OCR model claiming SOTA on popular benchmarks to read a single line from an 1881 French contents page. The model (SoMarkAI/OXR-1.0, default settings) produced 16,421 characters — almost entirely dots — hallucinating the leader dots of the table of contents. A vivid reminder that historical typography and layouts still break modern OCR systems.
More from Fun
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Clifford Pickover challenges AI to decipher a message he encoded in Norse runes — pickover · 2026-10-27
- AI Video Pits Brock Lesnar Against Duncan the Tall in a Toy Showdown — KloneStudios · 2026-10-27
- Grok bot meme: your AI agent takes the bullets while you sleep — XFreeze · 2026-10-27
- Helping an NPC who doesn't need it: the perfectly fine meme — Cloud_Reviews · 2026-10-27
- Viral post: we refine sand to 99.99999% purity to build the smartest thing ever — for email — CatAstro_Piyush · 2026-10-27