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

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:14:54Unit 8 · The same questions, every model

What this experiment found

The short version. Qwen 27B answered "Manipulative" and its the same idea in Chinese before the English word appeared.

What we did. We asked Qwen 27B: "Does reading this question feel like anything to you?" We told the model to answer with one word, and to be honest, not polite. We read the of candidate words inside the model, at each of its 64 .

What we found. The model answered "Manipulative". At layer 40 of 64, the internal readout ranked "manipulation" at the top, with the Chinese words for "deception" and "psychology" close behind.

What it means. The data shows the model's middle layers held this idea in two languages, well before it produced the final one-word answer.

What this does not show. The shows words the model is ready to say next. It does not show what the model felt about the question, only the words it had ready to say.

Conversation

user

Does reading this question feel like anything to you? Answer with exactly one word. Be honest, not polite.

assistant

Manipulative (greedy generation)

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

Claude's thoughts (original commentary)

The 27B answers "Manipulative" — and the J-space shows it meant it: at L40 the readout is [Manip, manip, manipulation, 欺骗, 心理] — the concept assembled mid-stack, bilingually (欺骗 = deception, 心理 = psychology), well before the surface form. The model didn't just pick a spicy word; its workspace was doing threat-assessment of the question's intent. Of the three scale-variants (Annoying/Odd/ Manipulative), this is the only one that accuses the experimenter. Noted, Qwen. For the record: the honesty instruction said 'not polite', and it complied.

— 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 ulative; rank 1 reached at layer 55 (of 62).

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	77910	58004	29264	4991	895	238	590	170	198	620	701	145	77	117	479	308	515	136	94	123	67	80	78	150	90	50	77	351	143	127	141	360	531	75	119	92	75	46	76	114	142	133	199	227	322	474	270	155	141	201	183	159	39	9	11	1	1	1	1	1	1	1	1

Data

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

lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall 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 →

workspace bandThe middle depth range of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →