SauersML/watching

A library for mechanistic interpretability of GPT-style language models

Companion code for the global workspace interpretability paper

To know what models don't say out loud.

Keeping language models honest by directly eliciting knowledge encoded in their activations.

AI Systems as Scientific Instruments for Understanding the Mechanisms of Intelligence

Extract residual-stream activations and apply steering vectors (including activation oracles) to any vLLM model during inference.

Mapping out the "memory" of neural nets with data attribution