AI safety research papers or announcements from AI safety organizations

OpenAI Enhances GPT-4 Interpretability with 16 Million Human Interpretable Features Using Sparse Autoencoders

Jun 6, 2024, 06:04 PM

OpenAI has introduced a new technique to enhance the interpretability of its language model, GPT-4, by breaking it down into 16 million human interpretable features. This advancement leverages sparse autoencoders (SAEs) to disentangle the internal representations of GPT-4, making it easier to understand the neural activity of the model. The new methods show promise in improving the trustworthiness and controllability of AI models. This development is part of the final work from the Superalignment team, which has also introduced new metrics for evaluating SAEs. The approach scales better than existing methods and operates completely unsupervised, marking a significant step forward in AI interpretability.