Research & Innovation
A New Design Framework Wants You to Edit an AI's Thoughts, Not Just Its Prompts
Nearly every interaction with an AI system today runs through prompts in, outputs out, with the model's actual internal workings kept completely opaque. A new MIT Media Lab framework called Neural Representation Interaction argues the internal representations models already compute are rich enough to interact with directly — and sketches design principles for interfaces that let ordinary users, not just interpretability researchers, work with what a model is actually 'thinking' rather than only what it says.