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 attic air hung thick and dusty, catching in Maya’s throat. A groan escaped her as she wrestled with the cardboard box, the tape stubbornly refusing to give way. Of all the things to be doing on a Saturday afternoon, this was the worst. Why couldn't her mother just have thrown the stuff out?

She finally ripped the box open, a cloud of ancient paper dust puffing into her face. "Ugh!" Maya coughed, waving a hand in front of her. Inside, neatly stacked, were stacks of papers. Mostly her papers.

The first page she flipped over was a childish scrawl, a drawing of a lopsided dog next to an even more lopsided tree. Anger pricked at her. This felt like some kind of weird museum exhibit she was forced to participate in. A wave of heat flushed her cheeks.

Emotion: irritated

Cluster: Anger / Hostility
PC1 (Valence): -0.86 Negative
PC2 (Disposition): 1.07

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
l0.370
S0.282
P0.264
😠0.248
aggravated0.233
Suppressed:
own-0.300
joyful-0.211
joyous-0.207
खोले-0.206
newfound-0.205