NDIF hosts Thinking Machines Lab’s Inkling 2 Mini and TML-d24-varloop-research looped transformers for remote white-box inference. This post explains how the looped architecture changes where the residual stream lives, and how to read, patch, steer, and differentiate its activations across loop iterations with NNsight.
This post is published on the NNsight blog: read it on nnsight.net.