A Mechanistic Interpretability Analysis of Grokking - AI Alignment Forum
www.alignmentforum.org
A significantly updated version of this work is now on Arxiv …
  • @maccam912@beehaw.orgEnglish
    2•10 months ago

    The takeaway for me on this though was how it started off essentially memorizing the sums, until eventually “getting it” and finding an algorithm.