Work Done
- fixed sae shape bug
- ran visualization code on meta-sae
- added EMNIST dataset
Confusions
- how could I measure what changes as you get more meta?
- some basic ideas: (LOOK AT THE PAPER!!!)
- (1) track the amount/trend of nonzero activations
- (2) track the distribution of nonzero activations (ie are they closer togeter, farther apart etc )
- (3) track feature descriptions from the eventual auto-interp loop
- some basic ideas: (LOOK AT THE PAPER!!!)
- how to deal with large amounts of data (ie downloading these datasets)
- how to clean up my mac since it has 300 GB of data somewhere (WHERE?!)
Next Steps
- fix EMNIST loading (3 dim channel issue again I think?)
- run normal sae on EMNIST
- run meta-sae on EMNIST
- Look at the work of the meta-sae paper
- (a) make the infinite meta-sae loop (this is probably more fun )
- (b) figure out how to analyze it
- VISUALIZE STUFF next time or the time after
- Transition to ViT at some point
Cool Stuff
- Google released a new product that allows you to Learn About anything deeply with AI