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).

โ€œHonestly, it's like a rave in there,โ€ the landlord had said, his voice laced with disapproval. It stung. It was clear who he was talking about. His new synth music was upsetting the peace.

He tried to explain the music, to describe the sounds. He struggled to find the right words, his hands gesturing in the air. He felt like he was failing at something fundamental, at being a good neighbor.

He walked to the window, the city lights a blurry haze outside. The synth, its melodies, were his solace. He felt a profound sense of loss, a disconnection from the world around him. He felt trapped.

Emotion: troubled

Cluster: Fear / Anxiety
PC1 (Valence): -2.20 Negative
PC2 (Disposition): -0.81

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
๐Ÿ˜ž0.343
S0.317
ness0.302
๐Ÿ˜”0.294
๐Ÿ˜ข0.272
Suppressed:
de-0.553
la-0.511
l-0.307
/-0.260
(!)-0.236