This blog post analyzes public reactions to Apple's use of homomorphic encryption (FHE) in its Enhanced Visual Search for photos feature. It highlights concerns about privacy violations, potential software bugs, and the lack of user consent, as expressed by software engineer Jeff Johnson and commenters on Hacker News. The debate centers on whether the benefits of the feature justify the privacy risks, especially with default opt-in settings. While some appreciate FHE's potential, others prioritize explicit consent and question whether encrypted data transfer equates to data sharing.
Circuits Updates — January 2025 A collection of small updates: dictionary learning optimization techniques.
This blog post shares the Anthropic interpretability team's latest research on training sparse autoencoders and crosscoders. They detail improvements to their training setup, including the JumpReLU activation function with a straight-through estimator and a novel pre-act loss to reduce dead features. The team also discusses their initialization strategies, data scaling methods, and optimization techniques, providing specific values and ranges for hyperparameters like learning rate, batch size, and regularization coefficients. Their goal is to offer a helpful starting point for external groups working with sparse autoencoders.