Steering vectors: tailor LLMs without training. Part I: Theory (Interpretability Series)

Поделиться
HTML-код
  • Опубликовано: 23 янв 2025

Комментарии • 7

  • @TarunGupta360
    @TarunGupta360 2 месяца назад

    Very helpful video! Please keep the good work coming :)

  • @GAURAVKAUL84
    @GAURAVKAUL84 4 месяца назад

    Wonderful explanation Anastasia!

  • @swairshah
    @swairshah 3 месяца назад

    Oh wow. Great to have non-slop ML channel like this. I think steering vectors, SAEs some of other MechInt papers would make a good series. I'd also like to know why something like KSVD isn't used (these days its faster too?) instead of SAEs.

    • @anastasiaborovykh120
      @anastasiaborovykh120  3 месяца назад

      oh interesting! i wasn't aware of KSVD, but i think it could be valuable in this setup. will look into it more & get back to you.

  • @RahulKumar-m1j2q
    @RahulKumar-m1j2q 4 месяца назад +1

    better w/out music