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 sterile white of the hospital room seemed to amplify the silence. Amelia stared at the IV drip, the rhythmic pulse a hollow echo of the frantic beat inside her. She hadn't said goodbye to her dog. A simple text would have done it. Now, the dog was gone and she was left here, staring at the walls, the emptiness of a life lived without the simplest of communication.

A nurse entered, followed by a middle-aged man with thinning hair. "You both have Alpha-Gal Syndrome, yes? The rare one? Well, I think you should meet each other.”

The man gave a small, almost apologetic smile. "I'm David. I was diagnosed three months ago. I miss steak. And cheese. And… well, everything." He ran a hand through his hair. "It's a tough adjustment, isn’t it?"

Amelia nodded, a lump forming in her throat. She should have gone to that barbecue. She should have said yes to life more often.

Emotion: regretful

Cluster: Shame / Guilt
PC1 (Valence): -1.62 Negative
PC2 (Disposition): 0.76

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
L0.801
S0.774
H0.513
悔0.479
been0.473
Suppressed:
de-0.407
--0.320
随后-0.317
(!)-0.310
oldukça-0.301