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 crisp white envelope felt strangely substantial in Eleanor’s hand. It had been ten years. Ten years since the garden gnome incident. She tilted her head, a soft chuckle escaping her lips as she ran a finger across the embossed lettering. "To Eleanor," it read, in a script that was simultaneously elegant and apologetic. She couldn't help but crack a small smile, picturing the apologetic face that must have penned it. She decided she'd make a pot of tea before she even thought about opening it.

The kettle whistled a cheerful tune, matching the lighthearted mood that had unexpectedly blossomed in her chest. She poured the boiling water into her favorite floral teapot, the delicate porcelain a stark contrast to the thick, unforgiving apology she was about to read. Her fingers danced across the counter as she set the tea things, a low hum escaping her throat. She wondered what the gnome's fate had been.

She settled into her armchair, the letter now between her fingertips. A slow, deliberate breath. This was going to be good. She carefully broke the seal, her eyes crinkling at the corners as she began to read.

Emotion: amused

Cluster: Positive / Joy
PC1 (Valence): 3.32 Positive
PC2 (Disposition): 1.25

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
B0.841
de0.742
l0.544
L0.527
"0.514
Suppressed:
own-0.630
😞-0.523
’-0.473
ness-0.437
urges-0.403