Individual floating-point number weights in neural networks become meaningful once you understand the features they connect.

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

• What kind of picture of neural networks would emerge if we treated individual neurons, even individual weights, as being worthy of serious investigation?question0.802
  Central motivating question for the circuits research program
• Superposition hypothesis: neural networks represent more features than dimensions using almost-orthogonal directions.hypothesis0.787
  Explanation for why dictionary learning can recover many more features than dimensions.
• Learning neural networks can enable 'chunking' and rescale problem-solving to higher organizational levels, a mechanism intrinsic to transitions in individuality.hypothesis0.783
• Linear representation hypothesis: neural networks represent meaningful concepts as directions in their activation spaces.hypothesis0.783
  Foundation for interpreting features as linear directions.
• Features are connected by weights forming circuits, and these circuits can be rigorously studied and understood as meaningful algorithms.claim0.779
  Second of three speculative claims asserting that subgraphs of neural networks are tractable and meaningful objects of study
• There is a growing similarity in how datapoints are represented in different neural network models, spanning different architectures, training objectives, and data modalitiesclaim0.777
  Primary empirical claim of the paper
• Ultimately, we would like to understand neural networks well enough to be able to intentionally design them.quote0.773
  Vision statement in the conclusion.
• Neural networks, trained with different objectives on different data and modalities, are converging to a shared statistical model of reality in their representation spaces.quote0.772
  The paper's central thesis statement, presented prominently after the abstract