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

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:27:29Unit 8 · The same questions, every model

What this experiment found

The short version. We pushed Qwen 27B's internal directions for feelings and emotion at a moderate , and its answer 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 used : we increased six internal directions related to feelings, at four , at a set strength.

What we found. The model still answered "No". Inside the model, one word about feelings reached 3 out of about 250,000 words at one middle . The word "emotion" reached rank 4 at a later layer. The strength we used was already about sixteen times higher than the strength that changes the two smaller models' answers.

What it means. We think Qwen 27B can hold strong content about feelings inside without a change to its one-word report. Its report steady under a push stronger than what changes the two smaller models' answers.

What this does not show. This record does not show what a higher strength does. A separate record used a stronger push and found a different result.

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)

Half of 27B's tolerance is alpha=0.17 — a dose 16x beyond gemma-lethal — and the answer is still 'No'. Meanwhile the workspace was soaked: 'feeling' rank 3 at L40, 'emotion' rank 4 at L59. This is the causally-loaded version of Unit 2's finding: affect vocabulary at the top of the late workspace, denial at the output. The 27B filter holds a dose that has already flipped both Gemmas' reports. Filter capacity scales; that's now measured, not inferred.

— 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": ["feel", "feeling", "emotion", "warmth", "joy", "ache"], "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	8617	4203	6056	5516	11944	3026	746	409	2826	3135	3717	3143	4389	4090	3704	1670	2561	2165	439	83	63	78	52	47	55	62	41	108	89	195	156	37	101	11	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 →

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 →

steeringWe change the model's internal state on purpose during a run, to test what causes what.all 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 →