Interpretability in the Wild: A Circuit for Indirect Object Identification in GPT-2 Small (Wang et al., 2023)

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• GPT's ability to simulate text automata is the source of its most surprising and pivotal implications for paths to superintelligence.claim0.766
  Importance of recursive generation.
• GPT, insofar as it is inner-aligned, is a simulator which can simulate agentic and non-agentic simulacra.claim0.758
  Central thesis of the post.
• GPT-2 implements at least one induction head using pointer arithmetic on positional embeddings rather than K-compositionhypothesis0.751
  Observation of an alternative induction head implementation algorithm in larger models with positional embeddings in the residual stream
• The field of interpretability has focused mainly on understanding model activations, not the computations themselvesclaim0.746
  Motivation for VPD's parameter-focused approach.
• Interpretability as technical grounding: activation patching and mechanism-finding validate the reflective/care/aliveness concepts.claim0.744
• The key challenge for active inference is finding the generative model that best explains observable data.claim0.742
  Identifies an outstanding problem, Section 10.
• Causal abstraction is not enough for mechanistic interpretability because it becomes vacuous without assumptions about how models encode informationclaim0.741
  Central thesis of the paper
• Automated interpretability using LLMs can usefully score feature specificity.claim0.741
  Claude 3 Opus ratings aligned with human judgment of feature descriptions.