Reproducible case study of pitfalls in contrastive SAE discovery and steering for "consciousness" features (GemmaScope SAEs, Gemma 3 4B/12B): reconstruction confound, delta-steering fix, matched controls, and false-positive scaling law vs dataset size.

Explore limitations of contrastive SAE steering in identifying causal consciousness features and introduce delta-steering to improve experiment validity.

🍳 Gemma's Test Kitchen — click a "vibe" to live-steer gemma-3-4b-it with GemmaScope SAE features, right in your browser. A beginner-friendly + power-user playground for SAE feature steering; no mechanistic-interpretability background needed.

Finding SAE features in Gemma 3 vision-language model — autonomous AI vs human-guided AI comparison using GemmaScope 2

Use GemmaScope and Neuronpedia features to steer Gemma 3 4B IT locally

Using LLM interpretability through middle-layer sparse auto-encoders to detect spelling, grammar, and word-level errors from internal model activations.