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 scent of honeysuckle, thick and cloying, drifted from Mrs. Gable’s garden. It always did this time of year, just like it had when I was a kid, sneaking glances over the fence, convinced there were fairies amongst the blooms. Now, the fence was gone, replaced by a wrought iron monstrosity guarding a meticulously manicured lawn. I watched from my porch, a half-eaten apple turning brown in my hand, as she emerged, a vision in crisp linen.

She was walking her miniature poodle, Monsieur Fifi, on a leash studded with what I now realized were diamonds. Diamonds! She’d always seemed so… ordinary. A woman who watered her petunias religiously and baked cookies for the church bake sale. The realization, a sharp pang in my chest, felt like the memory of a first crush, sweet and then unexpectedly gone.

I remembered the worn leather of my baseball glove, the crack of the bat, the way the late afternoon sun warmed the back of my neck. I sighed, the apple forgotten, as I pictured her: a quiet woman, an empty cookie sheet in her hands, waving to me with a smile as I rode my bike by. The world had shifted, but that image, that simple gesture, remained, a phantom limb aching with what was.

Emotion: nostalgic

Cluster: Nostalgia / Reflection
PC1 (Valence): 1.00 Positive
PC2 (Disposition): -1.93

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
H0.677
L0.647
시절0.611
nostalgic0.609
fondly0.609
Suppressed:
--0.564
either-0.525
or-0.497
싶-0.487
l-0.449