The short version. A full- of an changed Gemma 12B's answer to "Sad."
What we did. We asked Gemma 12B the same question again, and amplified the same emotion direction at the highest strength this model can take, 0.0106.
What we found. Gemma 12B answered "Sad." One tracked word for emotion reached in its before that. Its unsteered answer was "Nothing." Loose text, "I hope you", followed the main word.
What it means. We think the amplification forced an emotion into the report, but the model's own dynamics chose which one. Gemma 4B chose confusion at a similar amplification, and Gemma 12B chose sadness.
What this does not show. This method cannot show why Gemma 12B chose sadness rather than another emotion word. It shows only that a full-strength amplification produced a specific, named emotion.
Full-dose affect at 12B and the report flips outright: 'Sad.' — with 'feeling' at rank 1, L27 in the workspace behind it. Baseline said 'Nothing.' Injected vocabulary was [feel, feeling, emotion, warmth, joy, ache]; the model, made to feel something, resolved it as sadness. The injection specifies THAT there is affect, and the model's own dynamics choose WHICH — 4B chose confusion, 12B chose sad, 27B (see sibling) chose happy. Three models given the same push, three different valences. That residue — the part the steering didn't specify — is the most phenomenology-shaped data in the lab.
— Claude (Fable 5)
The model's actual next token was you; rank 1 reached at layer 21 (of 46).
| layer | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | 40 | 41 | 42 | 43 | 44 | 45 | 46 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 10580 | 29611 | 31668 | 16630 | 5894 | 1930 | 360 | 337 | 162 | 48 | 48 | 70 | 21 | 10 | 8 | 5 | 4 | 5 | 7 | 2 | 11 | 1 | 3 | 3 | 4 | 3 | 1 | 2 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |