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 the dusty table, its delicate inlaid wood shimmering in the afternoon light. Aunt Millie, hands clasped tightly, stared at it with a predatory glint in her eye. "It's time," she announced, her voice a little too high-pitched. My jaw clenched. I bit back the retort that sprang to my lips. My grandfather had given me that music box.

I turned away and walked toward the window, the sunlight warming my face. I could feel a knot tightening in my stomach. The way she kept looking at it, the way she kept touching it with those gloved hands, just sent a jolt of unpleasant energy through me. This wasn’t just about money, surely. It was about something more.

"I need the money," Millie continued, her voice losing its forced cheer. I spun around, my hands balling into fists.

Emotion: frustrated

Cluster: Anger / Hostility
PC1 (Valence): -1.56 Negative
PC2 (Disposition): 0.64

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.344
P0.302
exasper0.257
aggravated0.254
que0.252
Suppressed:
unfolding-0.242
otre-0.226
latter-0.220
☺️-0.216
soon-0.212