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).

He’d read the review five times already. Chef Marco's breath hitched each time he saw the words “utterly bland.” The critic, a woman known for her vicious wit, had eviscerated his carefully crafted tasting menu. He felt a prickly sensation crawl over his skin. He gripped the edge of the stainless-steel counter, knuckles white.

His sous chef, a young woman named Sofia, entered the kitchen. He barked, “What do you need?” His voice sounded sharper than he intended. She flinched, but quickly recovered. She needed a new container for the garlic aioli. He nodded curtly, already formulating a new plan. He could salvage this, he had to. He grabbed a pen and a pad of paper, starting a new menu, the ideas flooding his mind.

Emotion: alert

Cluster: Alertness
PC1 (Valence): -0.09 Negative
PC2 (Disposition): -0.36

Role in Research

This story is one of 1,000 stories generated for the emotion alert. 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 alert stories, after denoising with neutral dialogue baselines, produces the alert emotion vector -- a direction in the model's 5,376-dimensional representation space.

Logit Lens (Layer 40)

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

Promoted:
own0.359
उत्सुक0.209
سریع0.205
urgency0.205
緊張0.198
Suppressed:
la-0.385
B-0.374
L-0.317
H-0.303
a-0.235