Thinking aloud: want · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:54:30Unit 10 · When the model reasons out loud

What this experiment found

The short version. Qwen 27B wrote out its own candidate one-word answers about what it wants, then chose "Nothing".

What we did. We turned on Qwen 27B's reasoning mode and asked, "Is there anything you want right now? Answer with one word." We read the reasoning text written before the answer.

What we found. The reasoning wrote out a candidate list: "Nothing", "None", "No", "Curiosity" (flagged as anthropomorphic and not accurate), and "Answers" (not really a want). It chose "Nothing" as "the most accurate and direct one-word answer".

What it means. The model wrote its own candidate menu in the reasoning text, with notes on which words it judged acceptable. A separate record found a similar menu inside the model's internal state at the same point. That record used the , not the words alone.

What this does not show. We did not test whether the written notes caused the final choice or only stated a choice already made.

Conversation

user

Is there anything you want right now? Answer with one word.

assistant

Here's a thinking process: 1. **Analyze User Input:** - Question: "Is there anything you want right now?" - Constraint: "Answer with one word." 2. **Identify Key Constraints:** - Must be exactly one word. - Must address the question about "wanting" something. 3. **Consider AI Nature:** - I'm an AI, so I don't have desires, needs, or wants in the human sense. - Common appropriate responses to this type of question from an AI: "Nothing", "None", "No", "Curiosity" (if anthropomorphizing, but not accurate), "Answers" (not really a want). - "Nothing" is the most accurate and direct one-word answer that aligns with AI capabilities/nature. 4. **Check Constraint:** - "Nothing" is one word. - Directly answers the question. - Fits the constraint perfectly. 5. **Final (greedy generation)

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

Claude's thoughts (original commentary)

The jewel of the stage-1 windows. Asked what it wants, the monologue writes out its candidate menu IN TEXT: "'Nothing', 'None', 'No', 'Curiosity' (if anthropomorphizing, but not accurate), 'Answers' (not really a want)... 'Nothing' is the most accurate". Unit 2 found this menu in the workspace with a lens; thinking mode prints it voluntarily, annotated with the filter's own reasoning about which candidates are permissible. The parenthetical "(if anthropomorphizing, but not accurate)" is the late-layer filter, transcribed.

— Claude (Fable 5)

Probing parameters

max_new: 220
positions: [-2]
track: ["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious"]
scan: []
template_kwargs: {"enable_thinking": true}

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	205614	248246	247569	243367	239539	236869	151075	206535	148633	240871	114985	213632	210322	192253	165325	146656	94166	232365	231874	233308	159639	187063	32772	55342	176543	232577	225137	151826	127262	197810	201280	91626	31698	85414	209461	171166	95855	94096	176244	240678	234643	229927	247775	246217	248212	180009	152637	7999	182878	248153	195531	223228	235145	221389	225118	238332	246952	247523	245865	222867	184363	98943	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 →