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 relentless sun beat down on the blacktop driveway, and I was in the middle of washing my car. Beads of water, reflecting the bright light, ran down the hood. My neighbor, Sarah, a whirlwind of energy, approached, her hands gesturing enthusiastically.

“Hey, John!” she called out, already smiling. She wanted to plant some sunflowers. Large ones. Her eyes were bright with expectation.

I paused, leaning against the warm metal of the car. My back arched, and I took a deep breath. She went on, detailing the variety and where she’d like them to grow.

“Sounds lovely, Sarah,” I replied, my voice measured. “Let’s get a shovel and see where we can make that happen.” I smiled back, feeling a warmth spread through me.

Emotion: patient

Cluster: Passivity
PC1 (Valence): 2.71 Positive
PC2 (Disposition): 0.40

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
rest0.467
eventually0.415
leisurely0.369
慢慢0.357
a0.355
Suppressed:
无法-0.432
even-0.428
인해-0.421
無法-0.415
😣-0.391