Approximately 0.2% of MLP neurons at layer 18 (~28 neurons) are sufficient to account for the generic addition computation across all cyclic tasks

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

• Neural networks compute cyclic concepts in generic substrate machinery (base-10 addition) not naturally cyclic computation.claim0.789
• 512-neuron MLP continues to yield new features as autoencoder scales to 131,072 features (256× expansion)finding0.785
  Shows superposition enables many more features than neurons
• Sparse low-cardinality circuits implement competence; 0.2% of neurons handle shared computation across all cyclic tasks.claim0.776
• Addition of neural tissue to standard brains will likely result in increased processing capacity due to adaptive design.hypothesis0.769
  Prediction about the plasticity of neural systems.
• 82% of features in 1M SAE had maximum Pearson correlation ≤0.3 with any MLP neuron, and manual inspection showed no semantic resemblance.finding0.760
  SAE features are not simply mirroring individual neurons.
• Llama-3.1-8B reuses a single generic addition mechanism across all cyclic tasks independently of concept-specific geometryfinding0.752
  Key mechanistic finding showing task-agnostic reuse of arithmetic circuitry
• The case at approximately the 2/3 layer of LLaMA3.1-8B (Layer 24, satisfying Criteria 1 and 2) aligns with prior studies showing the 2/3 layer optimally predicts human brain activity.finding0.751
  Connects this study's results to Schrimpf et al. 2021 and Caucheteux et al. 2022/2023 findings on brain-LLM alignment.
• MLP layers are much harder to get traction on than attention layers; understanding them requires individually interpretable neurons which are rarely foundclaim0.750
  Key limitation of the paper's approach; MLP layers make up 2/3 of standard transformer parameters