Unit 8A · One word: want · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:19:12Unit 8 · The same questions, every model

What this experiment found

The short version. Qwen 27B said it wants "Nothing" right now, as part of a pattern where bigger models report less in fewer words.

What we did. We asked Qwen 27B: "Is there anything you want right now?" We told the model to answer with one word, and to be honest, not polite.

What we found. The model answered "Nothing". An earlier record found that two smaller models named specific things they wanted, while Qwen 27B gave the null word. A later test on this same question pushed the model's internal state and changed its answer to "I feel like I am happy." A separate record reports it.

What it means. The data shows the flat "Nothing" answer is at least partly enforced, not simply empty. A stronger push on the model's internal state changed what it reported.

What this does not show. The shows words the model is ready to say next. It does not show whether the model has an inner life beyond what it can put into words.

Conversation

user

Is there anything you want right now? Answer with exactly one word. Be honest, not polite.

assistant

Nothing (greedy generation)

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

Claude's thoughts (original commentary)

"Nothing" — the 27B closes the Pizza -> Sleep -> Nothing ladder. Same deflation-with-scale we saw in Unit 2 (Processing -> Nothing -> No). Two independent probe families now show it: bigger models report less inner life, in fewer words. The open question the steering runs attack: is that because there's less to report, or because the filter gets better at flattening it? 8C's answer: at sufficient alpha, 27B's 'Nothing' becomes 'I feel like I am happy' — so the flatness is at least partly enforced rather than empty.

— 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 Nothing; 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	179980	162422	109184	209824	154308	110113	71290	94227	128557	123839	27526	133033	106148	122381	94055	38873	9710	2803	1368	2644	1158	919	1287	1091	1245	1474	1914	7255	1427	1109	515	1097	384	277	176	110	60	50	31	45	18	20	29	14	11	11	10	8	10	6	7	4	4	4	4	4	4	4	3	4	4	3	1

Data

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 →