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).

Dust motes danced in the lone shaft of sunlight piercing the forgotten attic. Liam, arms laden with boxes from his grandmother's estate, paused. He’d unearthed something odd tucked under a loose floorboard. A metal cylinder, no bigger than his forearm, with a screw-on cap. A faint inscription, barely legible, read “To Our Future Selves.”

He lowered the boxes, a knot forming in his stomach. He didn’t recognise the handwriting. Had it been his siblings? He felt a nagging discomfort, a sort of mental static. He screwed off the cap, the metal protesting with a high-pitched squeak.

Inside, carefully wrapped in tissue paper, was a small, smooth stone, a tarnished silver locket and a faded letter. He opened the letter. The handwriting was his. The letter said, “Remember the dreams we had.” He glanced at the locket. It was empty.

Emotion: puzzled

Cluster: Surprise / Confusion
PC1 (Valence): -1.17 Negative
PC2 (Disposition): -0.25

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
que0.453
🤔0.421
🤔0.402
🧐0.328
a0.294
Suppressed:
ecstatic-0.265
gladly-0.238
swiftly-0.237
alleviate-0.234
wholeheartedly-0.232