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).

“Okay, the first thing is the password for the server,” I said, my voice clipped. Across the desk, sat Janice, wide-eyed and ready. The sun shone in through the window, highlighting dust motes in the air. I hated the sunlight.

“It's the same as it always has been,” I added, my gaze fixed on the screen, as if the monitor held all the secrets of the world. But I knew it was all meaningless. I’d spent ten years making this job my own.

I deliberately did not elaborate. Janice looked at me expectantly. I let the silence hang, thick and heavy. Finally, she sighed and asked a different question. I answered with a shrug, and then I went back to pretending to type.

Emotion: obstinate

Cluster: Defiance / Spite
PC1 (Valence): 0.25 Positive
PC2 (Disposition): 3.22

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
l0.642
la0.547
de0.536
either0.498
-0.433
Suppressed:
竟然-0.354
终于-0.353
难以-0.353
原来-0.351
previously-0.349