Spelling Backward Gets Easier for AI Models Upgraded to See Every Letter
A Nature study from LMU Munich, Ai2, Cambridge, the University of Washington and Imperial College London converts existing language models to read raw bytes instead of word fragments, at under 1% of a typical training budget. We explain why AI models struggle to spell backward, test a real tokenizer, chart the character-level gains against a control model, and set out what the byteified models give up in maths, coding and speed, plus what it means for businesses.
