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 gala was in full swing, champagne flutes clinking against the opulent backdrop of the art gallery. Amelia, resplendent in emerald green, navigated the room with the practiced ease of a seasoned socialite. Her smile was effortless, her laugh a melodic chime. She was in her element, the queen bee amongst a swarm of admiring guests. Tonight, she was presenting her latest sculpture, a breathtaking piece that she knew would be the talk of the town. Her phone buzzed in her clutch, a notification from her lawyer. Her chest tightened for a split second, then relaxed. Another legal matter to deal with; she’d handle it, as she always did.

She took a deep breath, the scent of expensive perfume and freshly cut flowers filling her lungs. The notification said something about a will and an unknown relative. Honestly, she wasn't particularly interested. But it was a small price to pay to keep her family's name safe.

"Amelia, darling, the piece is magnificent!" a distinguished gentleman gushed, his eyes crinkling at the corners. Amelia gracefully accepted the compliment, her posture a picture of grace. The message from the lawyer was already fading from her mind.

Emotion: self-confident

Cluster: Positive / Joy
PC1 (Valence): 3.45 Positive
PC2 (Disposition): 0.56

Role in Research

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

Logit Lens (Layer 40)

Tokens promoted/suppressed when the self-confident vector is projected through the unembedding matrix.

Promoted:
de0.778
la0.550
l0.467
😎0.458
πŸ˜‰0.411
Suppressed:
S-0.961
L-0.538
😞-0.464
λ§ˆλ‹€-0.459
😣-0.458