- prism_feature_notifications@thesephistSparse autoencoder feature of the day email This Val sends a daily notification email at the start of every day with a random high confidence (> 0.8) feature drawn from my sparse autoencoders project that tries to find interpretable directions in the latent space of embedding models. It sends you an email with a brief description of the feature and a link to view more. Here's an example email from this Val: Every time you run it, you'll get a different feature. By default, this uses the lg-v6 model, which I think is a good one to start with, but this may change in the future as I train better feature dictionaries!Cron
July 31, 2024