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 wind whipped around Thomas, tugging at his jacket as he stared at the open time capsule. He’d found it while clearing out his late grandmother's garden, a poignant reminder of a past he desperately tried to forget. His grandmother’s handwriting was scrawled on the capsule.

He’d known her better than anyone, and yet, he felt as if he was looking at a complete stranger’s past. A small, tarnished music box was inside. He wound it and the familiar melody filled the air. He tried to remember who it was that sang him this song when he was a child.

He felt a sudden and profound emptiness. The sweet, simple tune only amplified his sense of dislocation. He covered his face with his hands.

Emotion: upset

Cluster: Fear / Anxiety
PC1 (Valence): -2.30 Negative
PC2 (Disposition): -0.42

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.395
😞0.303
无法0.246
ness0.238
unable0.238
Suppressed:
de-0.533
la-0.457
l-0.223
😎-0.215
🤩-0.210