Workshop: Steering AI Models through Mechanistic Interpretability

​AI is a black box, but that's changing. Labs like Anthropic publish research that push towards a glass-box scenario, where humans have full observability into the brains of our AI. This workshop covers the specifics of that research, covering intros and demos to:

• ​Reverse Engineering Neural Networks (mechanistic interpretability)

• ​How Anthropic/OpenAI mapped out 'features' of their models using Sparse Autoencoders

• ​Steering Models in Real Time (how Anthropic/OpenAI steer their models, demonstrated with a 3B param model)

​Approaches from scientific literature will be expanded upon, with the goal of breaking down the knowledge for common understanding