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).

Rain hammered against the window, blurring the city lights into streaks of orange and white. Eleanor sighed, pushing a stray strand of hair from her face. Another evening spent navigating the same routine: microwave dinner, then scrolling aimlessly through social media. Each day felt like a repeat, the days melting into an indistinguishable mass. She picked up the book she was reading, the pages still pristine, a testament to her lack of focus. She'd been promising herself for weeks to finish it, a promise unkept. She felt a familiar pressure in her chest, a tight knot that seemed to get more pronounced with each passing day.

She walked over to the bookshelf, absentmindedly brushing her fingers across the spines. Her gaze landed on a familiar, leather-bound notebook tucked away behind a collection of cookbooks. It was Liam's, and she knew he was fiercely protective of it. Curiosity gnawed at her, a rebellious urge she hadn't felt in a long time.

He was out tonight, at his weekly poker game. She slipped the notebook from the shelf, its weight surprisingly comforting in her hands. The first page was a dedication, to her. As she read, a prickle of discomfort spread through her. The character, "Eliza," was eerily familiar. The story, a fictionalized version of their life, was almost too on the nose. Each detail, each nuance, was a reflection of her own life that she felt trapped in.

Emotion: stuck

Cluster: Sadness / Despair
PC1 (Valence): -2.17 Negative
PC2 (Disposition): -0.72

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.487
缺乏0.401
L0.361
每天0.348
stagn0.348
Suppressed:
l-0.472
la-0.346
随后-0.322
undamaged-0.268
C-0.263