Sparse Autoencoders for Interpretability Quiz

Answers and explanations appear as you take the quiz.

1. What is the main purpose of training a sparse autoencoder on a model's activations?
   • To decompose tangled activations into interpretable, mostly-isolated features
   • To compress the model so it runs faster
   • To translate text between languages
   • To increase the model's accuracy on benchmarks
2. How does an SAE’s sparsity penalty affect hidden-unit activations?
   • It guarantees each unit represents exactly one concept
   • It encourages only a small subset of the wider hidden layer to activate for an input
   • It makes every hidden unit activate equally
   • It removes the reconstruction objective
3. Roughly how many features did Anthropic extract when scaling SAEs to Claude 3 Sonnet in 2024?
   • About 500
   • About 2,000
   • About 34 million
   • About 100 billion
4. What did the 'Golden Gate Claude' demonstration show?
   • That SAEs can only be trained on images
   • That the model could no longer produce text
   • That features have no effect on outputs
   • That clamping a single feature high causally steers the model's behavior
5. Why is a hidden layer in an SAE typically much WIDER than the activation it reconstructs?
   • Because concepts are superposed in fewer dimensions than they truly need, so more units help disentangle them
   • To make training cheaper
   • Because wider layers are always more accurate
   • To reduce the number of parameters
6. Which is a known limitation of current sparse autoencoders?
   • They cannot be trained with gradient descent
   • Feature splitting, where one concept fractures into many separate features
   • They only work on convolutional networks
   • They eliminate the need for any human labeling
7. Which two objectives are commonly combined when training a sparse autoencoder?
   • A policy gradient and value loss
   • A next-token loss and a contrastive loss
   • Reconstruction error and a penalty on hidden activations
   • A calibration error and classification accuracy
8. What does a TopK sparse autoencoder do to encourage sparsity?
   • Keeps the K largest input pixels and discards the decoder
   • Sets all encoder weights to zero
   • Trains K different tokenizers
   • Keeps only the K largest hidden activations for an input