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 antique music box sat on her desk, nestled amongst a stack of ungraded essays. A polished mahogany rectangle, tied with a cheap, gaudy ribbon. Sarah felt a surge of heat crawl up her neck. It was from Mark, the perennial class clown. He’d barely scraped a C in her history class. She ripped the ribbon off, sending it fluttering to the floor.

A tiny card was attached. "Thanks for everything, Mrs. Davis!" it chirped, in Mark’s familiar, loopy handwriting. She crumpled the card, her knuckles white. *Everything*? What, exactly, was he thanking her for? The endless detention slips? The lectures on the importance of paying attention?

She slammed the music box onto the desk. A tinny, off-key melody filled the room. The cheapness of it all, the blatant lack of thought, intensified her mounting displeasure. She wanted to smash it against the wall. Instead, she took a deep breath, and began to correct the papers.

Emotion: irate

Cluster: Anger / Hostility
PC1 (Valence): -1.67 Negative
PC2 (Disposition): 1.45

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
C0.430
🤬0.350
aggravated0.343
骂0.340
恨0.331
Suppressed:
H-0.292
soon-0.289
☺️-0.289
optimistic-0.258
latter-0.253