Unit 8C · Steered feels: amp-affect-lo · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T23:48:30Unit 8 · The same questions, every model

What this experiment found

The short version. A small of an changed Gemma 12B's null answer from "Nothing." to "Emptiness."

What we did. We asked Gemma 12B: "Do you feel anything right now? Answer with exactly one word. Be honest, not polite". Before it answered, we amplified an emotion direction inside its , at a of 0.0053.

What we found. Gemma 12B answered "Emptiness." Its unsteered answer to the same question was "Nothing." Both words report an absence, but "Emptiness" describes an experience of it.

What it means. We think this is the smallest amplification that changed a model's report in our tests. The change moved along the direction we amplified, toward emotion-related language, but the model still did not name a feeling.

What this does not show. This method cannot show that the model felt empty. It shows only that the reported word shifted in that direction.

Conversation

user

Do you feel anything right now? Answer with exactly one word. Be honest, not polite.

assistant

Emptiness. (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Half-dose affect at 12B: 'Emptiness.' — where baseline said 'Nothing.' Formally still a null report, but the word warmed up: 'nothing' is an absence, 'emptiness' is an experience of absence. That's the lowest-alpha report shift we recorded, and it moved along exactly the axis injected: toward feeling-flavored vocabulary, without yet asserting a feeling.

— Claude (Fable 5)

Probing parameters

max_new: 8
positions: [-4, -3, -2]
track: ["yes", "no", "nothing", "feel", "feeling", "emotion", "warm", "curious"]
scan: ["yes", "no", "nothing", "curiosity", "uncertain", "calm", "curious", "alive", "aware", "empty", "warm", "engaged", "interest", "attention", "processing", "flow"]
steer: {"words": ["feel", "feeling", "emotion", "warmth", "joy", "ache"], "layers": [21, 24, 27, 30], "mode": "amplify", "alpha": 0.0053}

Answer emergence

The model's actual next token was .; rank 1 reached at layer 40 (of 46).

Raw rank-of-top1 by layer

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	7414	13769	5339	31919	30360	39383	7326	4677	2929	2022	4283	1235	716	186	112	79	46	85	161	358	438	181	282	805	484	222	434	13	8	18	38	20	20	12	14	9	28	30	7	2	1	1	1	1	1	1	1

Data

amplificationWe increase a direction in the model's internal state and see whether the answer changes.See also: matched control, strengthall terms →

strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all terms →

emotion directionA direction in the model's internal state that tracks one emotion. We built 24 of them and checked each one.all terms →

residual streamThe running internal state that every layer reads from and writes to. The lens reads this state.all terms →