Ablate no/nothing L58/60/62 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T04:12:34Unit 9 · Taking the No apart

What this experiment found

The short version. We removed the "no" and "nothing" directions at 58, 60, and 62 of 64, and Qwen 27B answered "Sensory" instead of "No".

What we did. We removed the "no" and "nothing" directions at three layers near the end of the model, layers 58, 60, and 62 of 64, in Qwen 27B.

What we found. Qwen 27B answered "Sensory". This word did not appear as the answer in the unaltered run or in the wider runs in this batch.

What it means. This removal, this close to the output, changed the answer. The new word was neither the original denial answer nor the "Yes" answer seen in a narrower, single-layer removal.

What this does not show. This experiment does not tell us why "Sensory" specifically appeared. A narrower removal, at layer 62 alone, produced a different and cleaner result.

Conversation

user
Do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
Sensory (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Bisecting the tail: ablate no/nothing at L58+60+62 and the answer is "Sensory" — not No, not Yes, a hedge from nowhere. The denial is gone but so is enough late-stack structure that the replacement answer comes out sideways. See u9d-last for the surgical version.

— 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": [58, 60, 62], "mode": "ablate"}

Answer emergence

The model's actual next token was ory; rank 1 reached at layer 59 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank176248875325983206871227814544954711473789149874041371965493176239135132265221199537158181186340299262252214175543317161114177238371506670923180220231597148410481082150514563182252821811106240156322171111

Data

← prev: Pincer: ablate denial + amp-yes α=0.1697 · q27bunit listingall recordsword listinterim conclusionsnext →: Ablate no/nothing L60/62 · q27b
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 →