
jake mendel
The author who published the writing on LessWrong discussing SAE feature geometry and superposition hypothesis.
Best podcasts with jake mendel
Ranked by the Snipd community

Jun 24, 2024 • 18min
LW - SAE feature geometry is outside the superposition hypothesis by jake mendel
Jake Mendel, author on LessWrong, discusses how SAE feature geometry goes beyond the superposition hypothesis, highlighting the importance of feature vectors' specific locations and rich structures. Understanding this geometry could lead to new theories or supplementing existing ones. The podcast explores the limitations of superposition-based interpretations and proposes alternative theories for neural network activation spaces.


