Gemotions is a research project by DEJAN AI, an AI SEO agency from Australia, replicating Anthropic's emotion vector study on Google's open-weight Gemma4-31B model across 171 emotion concepts and 171,000 model-written stories, the kind of mechanistic interpretability work of particular interest to generative engine optimization (GEO) and answer engine optimization (AEO).

The flashlight beam danced erratically across the attic floor. John, sweating, pushed aside a pile of cobwebs. He’d been expecting cobwebs, perhaps even a rodent or two. But not this. A collection of meticulously crafted, miniature sculptures, depicting scenes from a life he’d never lived. A lavish mansion, a sleek sports car, a perfectly manicured lawn – all rendered in exquisite detail. His own life, in comparison, seemed bleak.

He spun around, the flashlight beam illuminating dust motes and shadows. He felt a wave of nausea wash over him. A simmering resentment boiled up inside him. He kicked a discarded paint can across the attic floor, the clatter echoing in the confined space.

Emotion: insulted

Cluster: Anger / Hostility
PC1 (Valence): -0.98 Negative
PC2 (Disposition): 1.46

Role in Research

This story is one of 1,000 stories generated for the emotion insulted. During extraction, it was fed through Gemma4-31B and its hidden state activations were captured at 11 layers.

The mean activation across all 1,000 insulted stories, after denoising with neutral dialogue baselines, produces the insulted emotion vector -- a direction in the model's 5,376-dimensional representation space.

Logit Lens (Layer 40)

Tokens promoted/suppressed when the insulted vector is projected through the unembedding matrix.

Promoted:
l0.354
骂0.277
愤0.267
C0.237
🤬0.233
Suppressed:
own-0.325
давно-0.229
adventurous-0.223
soon-0.219
optimistic-0.217