← Back to all posts
Retrofitted LLM can count the letter ‘i’ in ‘artificial intelligence’
🤖

Retrofitted LLM can count the letter ‘i’ in ‘artificial intelligence’

Most LLMs struggle to count letters because they process words as tokens rather than individual characters. A technique called byteification retrofits these existing models so they can work directly with the bytes that encode each character. The researchers show that byteified models can deliver competitive performance while gaining the ability to inspect text letter by letter. So, that familiar “how many i’s?” question could become a much fairer test of what an AI can actually read.
Read the original article ↗
Share