AlgZoo.pdf - Google Drive

In the wake of recent debate about pragmatic versus ambitious visions for mechanistic interpretabili‐ ty, ARC is sharing some models we've been studying that, in spite of their tiny size, pose a challenge for any ambitious interpretability vision. The models are RNNs and transformers trained to perform al‐ gorithmic tasks, and range in size from 8 to 1,408 parameters. The largest model that we believe we more-or-less fully understand has 32 parameters; the smallest model we've failed to fully understand despite substantial effort has 432 parameters. The models are available at the AlgZoo GitHub repo. We think that the "ambitious" side of the mechanistic interpretability community has historically un‐ derinvested in "fully understanding slightly complex models" compared to "partially understanding in‐ credibly complex models". There has been some prior work aimed at full understanding, for instance on models trained to perform paren balancing, modular addition and more general group ope

saved by

• Asher P