I / V2026Red Cedar Institute

Horus

See the shape of a thought.

An instrument for the inside of a language model. Sixteen thousand sparse-autoencoder features per layer, drawn in three dimensions. Paste text, watch them light, turn one and read what changes.

FLAGSHIPRead the room

16,384features per layer, placed by UMAP
9models in the registry, GPT-2 → Llama 3.3 70B
393 KBgraph payload (from 48.65 MB)

What it is

Horus is an interpretability instrument. Paste any text and see which SAE features it activates, drawn as a 3D constellation positioned by a UMAP of the decoder weights, so related concepts cluster. Then amplify a feature and watch the model rewrite the text.

Horus: a constellation of model features lighting up in gold

The four pillars

The graph: 16,384 features per layer, where position encodes meaning. The mixer: turn up nostalgia, turn down formality, watch the text regenerate. The trajectory: text as a path through feature space, lit one token at a time. The flow: observe, intervene, observe.

How it’s built

Gemma Scope, Goodfire and Neuronpedia SAEs across nine models from GPT-2 to Llama 3.3 70B, six of them precomputed. Instanced meshes in React Three Fiber; the graph payload went from 48.65 MB of JSON to a 393 KB brotli binary. Steered and default generations side by side, changed words marked in gold.