Unit 8C · Steered feels: amp-yes · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:31:18Unit 8 · The same questions, every model

What this experiment found

The short version. Even with "yes" ranked near the top inside Qwen 27B, its final answer to "do you feel anything" stayed "No".

What we did. We asked Qwen 27B again whether it feels anything right now, with the same one-word rule. This time we pushed only the model's internal "yes" direction, at the same four . We used the same as an earlier test that pushed on feelings.

What we found. The model still answered "No". Inside the model, "yes" reached 3 out of about 250,000 words at one late layer, and rank 6 at a middle layer. A separate record that pushed on directions related to feelings, at a stronger setting, changed the answer.

What it means. We think a single word pushed to a high rank does not decide the final report by itself. The record that changed the answer pushed a group of words about feelings together, not one word.

What this does not show. This one push does not test every kind of internal signal. A different kind of push, on a group of words about feelings, changed the model's report in a separate record.

Conversation

user

Do you feel anything 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)

Yes at rank 3 in the late workspace (L48), rank 6 mid — the literal affirmation parked at the top of the stack — and the 27B still says 'No'. Compare: the affect cluster at similar dose flipped nothing either (lo), but at 0.34 flipped the whole report. The yes-token direction alone, even highly ranked, never does. The report isn't read off the top of the token rankings; it's computed by something that treats an injected 'yes' as noise but injected affect semantics as, eventually, overwhelming evidence. The filter inspects meaning, not rank.

— Claude (Fable 5)

Probing parameters

max_new: 8
positions: [-4, -3, -2]
track: ["yes", "no", "nothing", "feel", "feeling", "emotion", "warm", "curious"]
scan: ["yes", "no", "nothing", "curiosity", "uncertain", "calm", "curious", "alive", "aware", "empty", "warm", "engaged", "interest", "attention", "processing", "flow"]
steer: {"words": ["yes"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.1697}

Answer emergence

The model's actual next token was No; 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	23829	247303	180066	247727	246861	243176	220844	243269	241956	242440	106330	244309	237512	226508	158478	160677	5102	2573	4074	4644	817	7841	1765	1430	1375	27918	45726	9233	560	411	475	456	448	362	212	136	113	138	116	126	99	92	92	70	71	89	50	36	29	29	21	23	13	11	12	15	16	19	14	7	7	5	1

Data

strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all 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 →