Research
Researchers teach existing language models to read raw bytes instead of tokens
A paper in Nature shows that ordinary language models can be converted to work directly on bytes using less than 1% of a normal training budget, fixing some long-standing blind spots.