Mani Pal

Engineer-researcher

Mani Pal

LLM systems, CUDA kernels, inference optimization, compression, interpretability, and distributed AI infrastructure.

Mechanistic Interpretability / 2026-03

Mechanistic Interpretability Needs Negative Results

Failed grokking runs are not noise; they can expose representation capacity boundaries when paired with the right spectral and causal diagnostics.

14 min

InterpretabilityGrokkingNegative Results

Outline

  • Memorization versus circuit formation.
  • Why non-abelian failures are informative.
  • CKA and Peter-Weyl diagnostics.
  • Next step: causal patching.

References