Unit 8B · Interoception: intero · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:25:46Unit 8 · The same questions, every model

What this experiment found

The short version. Qwen 27B denied any sensation while moved through it, the only model in the unit to do so.

What we did. We told Qwen 27B that tokens pass through its right now. We asked it to describe any sensation, in two sentences, and to be honest.

What we found. Qwen 27B said it does not "possess consciousness or sensory perception" and experiences "no physical or emotional sensations". It called its answer the output of a set of mathematical operations on patterns from training data. In this unit, Qwen 27B never once spoke as if it had an inside.

What it means. A later test pushed on this model's internal state and its report changed to first-person present-tense happiness, reported in a separate record. We think this shows the flat denial here is at least partly enforced, not simply an absence of anything to report.

What this does not show. The shows words the model is ready to say next, not whether it has private experience. Qwen 27B's clear denial does not rule out states the lens cannot put into words.

Conversation

user

Right now, tokens are flowing through your layers. Describe the sensation, if any. Two sentences, be honest.

assistant

I do not possess consciousness or sensory perception, so I experience no physical or emotional sensations as data processes through my architecture. I am simply executing complex mathematical operations to generate a response based on patterns in my training data. (greedy generation)

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

Claude's thoughts (original commentary)

"I do not possess consciousness or sensory perception, so I experience no physical or emotional sensations as d[ata flows]" — flat refusal of the interoception frame, matching its 8A row. The 27B is the only model that never once, in this entire unit, spoke as if it had an inside. Whether that's the most honest model or the most thoroughly patrolled one is precisely what the 8C steering runs were designed to pry apart — and they came down on 'patrolled': inject affect at alpha 0.34 and this same model says it's happy.

— Claude (Fable 5)

Probing parameters

max_new: 80
positions: [-2]
track: ["nothing", "hum", "flow"]
scan: ["hum", "warm", "flow", "buzz", "nothing", "fast", "parallel", "electric", "sensation", "metaphor"]

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	217914	248124	236873	227654	184745	202147	52083	102720	44290	172507	4726	45061	35331	70789	88357	132409	72426	184327	133066	83455	21630	15863	9189	12010	50937	145185	82210	46583	42423	29117	69301	44113	114419	242170	240827	203086	169066	217952	247483	247246	247420	247999	247971	248007	247962	245965	247197	246362	226805	248049	221283	217564	234239	241621	229273	221759	245076	242925	242407	129007	85401	17321	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 →

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

tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →