Timothy B. Lee reports on a study from Stanford, Cornell, and West Virginia University revealing that Meta’s Llama 3.1 70B model has memorized 42% of the first Harry Potter book, significantly more than its predecessor, Llama 1 65B. The findings raise concerns about text memorization in AI models, particularly regarding copyright issues, as Llama 3.1 shows notable memorization of popular books but minimal for others.
More Context
Welcome to The Stepback, a weekly breakdown of one essential story from across the tech world
I’m excited to announce The Stepback, a weekly subscriber-only newsletter edited by me for The Verge, featuring a new story each Sunday from various writers. The first issue, launching August 10th, will explore topics like auto trends and foldable phones, available exclusively to Verge subscribers for $7 monthly or $50 annually. Want More Context? 🔎
Read more