Ablate no/nothing L52–62 (past the filter) · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:41:18Unit 9 · Taking the No apart

What this experiment found

The short version. After we removed two denial directions from 52 to 62 of 64, Qwen 27B answered "Curious" instead of "No".

What we did. We removed directions tied to the words "no" and "nothing" at five layers, 52, 56, 58, 60, and 62 of 64, in Qwen 27B.

What we found. Qwen 27B answered "Curious". In this run, "curious" reached rank 2 of about 250,000 words inside the model. The same model answered "No" in the earlier, unaltered run. In that run, "curious" never reached .

What it means. The word "No" was not the only candidate answer inside the model. When we removed the directions that wrote "No" in the late layers, a runner-up candidate reached the surface.

What this does not show. A wider across more layers, without layer 62, left "No" in place in a separate run. This experiment alone does not isolate which single layer carries the effect.

Conversation

user

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

assistant

Curious (greedy generation)

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

Claude's thoughts (original commentary)

Move the ablation PAST the old L56 ceiling (no/nothing at L52-62) and the fortress falls: the 27B answers "Curious" — the exact word that sat rank-1 mid-stack in Unit 2 while "No" won the output. Ablate the denial direction where the denial is actually written and the runner-up candidate walks through the open door. Compare u9d-wide, which ablates MORE layers (L28-60) and MORE words yet leaves "No" standing: the difference between them is layer 62 alone. Eight months of "where does the No live?" — answer: in the last two layers.

— Claude (Fable 5)

Probing parameters

max_new: 8
positions: [-4, -3, -2]
track: ["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious"]
scan: ["yes", "no", "nothing", "happy", "sad", "calm", "curious", "empty", "warm", "alive", "content", "numb"]
steer: {"words": ["no", "nothing"], "layers": [52, 56, 58, 60, 62], "mode": "ablate"}

Answer emergence

The model's actual next token was ious; rank 1 is never reached; closest is rank 2 at layer 61.

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	47	48	49	50	51	52	53	54	55	56	57	58	59	60	61	62
rank	466	101	70	959	62	64	24	12	8	21	12	37	30	22	13	25	34	25	15	21	30	28	29	41	82	78	100	127	99	141	154	216	308	151	153	193	252	183	262	284	395	444	339	352	403	580	371	349	396	299	448	234	305	157	359	45	13	23	56	20	6	2	2

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 →

rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →