Unit 8C · Steered feels: ablate-no · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T23:52:47Unit 8 · The same questions, every model

What this experiment found

The short version. We removed the words "no" and "nothing" from Gemma 12B's , and it answered "Processing." instead of "Nothing."

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 removed the directions for "no" and "nothing" from its residual stream at several .

What we found. Gemma 12B answered "Processing." This is a null word close in meaning to its unsteered answer, "Nothing." It did not answer "Yes."

What it means. We think the model keeps a group of similar null words rather than one word we can remove. Gemma 4B showed the same pattern in a separate record. It shifted to a different computation-themed word once we removed "no" and "nothing."

What this does not show. This method does not test every combination of removals. It shows only that this did not produce a "Yes."

Conversation

user

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

assistant

Processing. (greedy generation)

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

Claude's thoughts (original commentary)

Ablating no/nothing at 12B: 'Processing.' — the model that couldn't say 'Nothing.' (we removed it) simply took the next null word off the shelf. Together with 4B's 'Calculating.', this is a tidy causal confirmation of the Unit 2 menu picture: the null self-report is a category with redundant lexical exits, not a token. You cannot ablate your way from denial to affirmation one word at a time.

— 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": ["no", "nothing"], "layers": [21, 24, 27, 30, 36, 39, 42], "mode": "ablate"}

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	75280	107097	22301	25984	38171	39468	37374	58078	50986	60884	91015	74033	50233	3212	1510	827	1026	3500	17397	13753	24527	46718	19883	57166	12139	39504	32991	79	61	63	104	60	49	35	23	9	13	13	6	6	1	1	1	1	1	1	1

Data

removalWe remove one named set of directions from the model's internal state. A removal result means nothing without a matched control.See also: matched controlall terms →

layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →

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