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 report was flawed, riddled with inaccuracies. He stared at the pages, a hot flush creeping up his neck. The hours wasted, the effort poured into a document that was now useless. His fingers drummed a furious rhythm on his desk, the sound a staccato of discontent.

“That’s a Camp Firefly lanyard, isn’t it?” a voice broke through his thoughts. It was Jessica, a new hire.

He felt a surge of… something. He pushed the feeling down. “Yes,” he said, his voice clipped.

“I went to Camp Firefly every summer,” she continued, oblivious. “So much fun, right?”

He closed his eyes briefly. The fun? He’d been the laughingstock of the camp for an unfortunate incident involving a canoe and some ducklings. “Right,” he said, his voice devoid of any warmth.

“I remember this time, the camp had a fire drill. Did you get the award for the fastest time?” she asked with excitement.

He opened his eyes. “No,” he said. The word was cold and short.

Emotion: angry

Cluster: Anger / Hostility
PC1 (Valence): -2.05 Negative
PC2 (Disposition): 1.18

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
C0.354
est0.354
aggravated0.342
🤬0.337
恨0.323
Suppressed:
de-0.424
a-0.411
la-0.341
H-0.320
L-0.296