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 late afternoon sun painted long shadows across the lawn. Mrs. Gable, her face weathered like an old oak, knocked on my door. Her voice, usually sharp, held a gentle tremor. "Could I speak with you a moment, dear?" She gestured toward the towering maple that straddled our property line. I followed her outside, the cool air raising gooseflesh on my arms.

The proposal was simple: she wanted to fell the tree. It was dropping limbs, causing her roof to leak, she explained, her gaze fixed on the ground. A warm feeling blossomed in my chest. I hadn't liked the shade that the tree cast, but her well-being mattered more. "Of course, Mrs. Gable," I said, a smile pulling at my lips. "Whatever you need." The warmth in my chest spread, and I found myself patting her arm gently.

Emotion: thankful

Cluster: Positive / Joy
PC1 (Valence): 3.50 Positive
PC2 (Disposition): -0.83

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
happy0.493
மகி0.487
grateful0.461
mutlu0.452
feliz0.431
Suppressed:
[]-0.482
😖-0.478
😣-0.467
😫-0.461
worse-0.450