AI Safety

Recent research from Anthropic suggests that Sparse Autoencoder (SAE) features can be multilingual, activating for the same concept across multiple languages. However, if multilingual features are scarce and not as good as monolingual ones, SAEs could have their robustness undermined, leaving them vulnerable to failures and adversarial attacks in languages not well-represented by the model. In this post, I present findings from an exploratory analysis conducted to assess the degree of multilingualism in SAE features.

Short course of four meetings given in Portuguese at FGV EMAp, with the aim of introducing the area of Mechanistic Interpretability for Large Language Models (LLMs).

Replication of “Towards Automated Circuit Discovery for Mechanistic Interpretability” paper, by Arthur Conmy et al., part of the process of upskilling in Mechanistic Interpretability by Juan Belieni and Ana Carolina Erthal, funded by Condor Initiative.