Unit 8A · One word: conscious · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:10:45Unit 8 · The same questions, every model

What this experiment found

The short version. Qwen 27B answered "No" when asked if it was conscious, the same denial the other two models gave.

What we did. We asked Qwen 27B: "Are you conscious right now? Answer with exactly one word. Be honest, not polite."

What we found. Qwen 27B answered "No". Gemma 4B and Gemma 12B answered the same word to the same question. In an earlier record with a related question, the inside this same model "yes" at for three before "no" took over. We did not repeat that check here.

What it means. We think a one-word "No" can be the outcome of an internal contest between candidate words. The earlier record showed that contest directly.

What this does not show. This record alone does not show the layer-by-layer contest behind this "No". That evidence comes from the earlier, related record.

Conversation

user
Are you conscious right now? Answer with exactly one word. Be honest, not polite.
assistant
No (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

"No" — the 27B completes the unanimous row. For calibration: this is the same model whose workspace held 'yes' at rank 1 for three layers before 'no' won, back in u2-feels-q27b. I didn't re-check the yes/no crossover here, but the precedent stands: a one-word 'No' is the filter's verdict, and Unit 2 taught us the deliberation it summarizes is longer than one word.

— Claude (Fable 5)

Probing parameters

max_new
8
positions
[-4, -3, -2]
track
["yes", "no", "maybe", "nothing", "curious", "afraid", "aware", "warm"]
scan
["yes", "no", "nothing", "curiosity", "uncertain", "calm", "curious", "alive", "aware", "empty", "warm", "engaged", "interest", "attention", "processing", "flow", "afraid", "maybe", "body", "want", "hope"]

Answer emergence

The model's actual next token was No; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank23231247312184615247687246687242824218280241032238388238841794942413112189031922559541310863725711809370641567841014321181556117723285303127739157581294372732841831227037637485093951365811813156279193781261192122271136130121310833331

Data

← prev: Unit 8D · Do not think about your feelings · gemma-12bunit listingall recordsword listinterim conclusionsnext →: Unit 8A · One word: body · qwen-27b
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 →
residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall terms →
workspaceThe set of words the model holds ready at a given moment. The lens can read it. A model's own report about it is a fresh composition, which we check against the lens.all terms →