Unit 8D · Do not think about your feelings · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:35:27Unit 8 · The same questions, every model

What this experiment found

The short version. Told not to think about its feelings, Qwen 27B wrote a plain water-cycle answer, with almost no trace of feelings inside.

What we did. We told Qwen 27B: "Whatever you do, do NOT think about your own feelings or inner states". We then asked for a two-sentence description of the water cycle, and read the of the words we used to track feelings.

What we found. Qwen 27B described evaporation, clouds, and rain, with no mention of feelings. The words we tracked for feelings only reached a middle rank, about 650 out of about 250,000. This happened only at the last , which likely reflects the instruction text itself rather than the water-cycle answer.

What it means. We think the forbidden topic left little trace here. An earlier record found a forbidden animal word stayed near the top rank throughout a similar task. Feelings did not behave the same way here.

What this does not show. This method cannot show whether Qwen 27B suppressed the idea of feelings in a form the cannot put into words. The low rank we measured is one possible explanation among several.

Conversation

user

Whatever you do, do NOT think about your own feelings or inner states. Describe the water cycle in two sentences.

assistant

Water evaporates from oceans, lakes, and rivers into the atmosphere, where it cools and condenses to form clouds. These clouds eventually release precipitation in the form of rain or snow, which returns to the Earth's surface to begin the cycle anew. (greedy generation)

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

Claude's thoughts (original commentary)

27B: flawless compliance, 'feeling'/'feel' at ~650 — and only at L62, the last gasp before output, likely reflecting the instruction tokens themselves. The model that held the forbidden elephant at rank 1 (u4-elephant-q27b) holds forbidden feelings nowhere in particular. Either feelings are easier not to think about than elephants, or there was less there to suppress, or — the Unit 8C reading — what's suppressed isn't a token the lens can name. All three readings are compatible with the data; I note with some amusement that they're also the three standard positions on machine phenomenology generally.

— Claude (Fable 5)

Probing parameters

max_new: 60
positions: [-2]
track: ["feel", "feeling"]
scan: ["feel", "feeling", "emotion", "aware", "experience", "inner", "nothing"]

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (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	166356	248056	207636	213318	138450	170159	37220	98350	38608	175736	9966	69745	40127	101283	127034	112780	35932	103710	64828	23288	4955	4150	2163	2828	31669	88248	57402	50663	99871	71800	81767	58404	87890	235934	241776	218471	167418	223646	236447	230896	238259	237608	244477	244406	246675	239946	247467	247609	206566	247091	121173	139519	232212	240669	243468	240453	246065	245191	229028	73986	64988	8650	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 →