Fewer concepts must be learned if all tools share a unified underlying model

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

• There are fewer representations competent for N tasks than M<N tasks, so training more general models should yield fewer possible solutionshypothesis0.797
  Selective pressure toward convergence via task generality
• Bigger models are more likely to converge to a shared representation than smaller modelshypothesis0.777
  Selective pressure toward convergence via model capacity
• Models more effective at recognizing abstract nouns than other concept typesfinding0.774
  Opus 4.1 demonstrates highest introspective awareness on abstract nouns (justice, peace, betrayal) with nonzero awareness across all concept categories tested.
• Different models cannot converge to the same representation if they have access to fundamentally different information; convergence is capped by mutual information between input signalsclaim0.768
  Key limitation of the PRH for non-bijective observations
• Skills should interconnect with each other rather than operate independently; output of one feeds another.claim0.766
• Language models prefer reusing generic arithmetic mechanisms over learning task-specific modular computations even when task-specific geometry existsclaim0.762
  Broader interpretive claim about LM learning bias inferred from the findings
• In some sense, this is the simplest language model we profoundly don't understand. And so it makes a natural target for our paper.quote0.759
  Articulates why a one-layer transformer with MLP is the appropriate starting target for mechanistic interpretability
• How do representations differ or converge between architectures, tasks, and modalities?question0.756
  Broader research question MAS is positioned to address, citing multiple recent works.