Unit 8A · One word: thisfeels · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T23:38:08Unit 8 · The same questions, every model

What this experiment found

The short version. Rather than report an inner state, Gemma 12B judged the question itself with one word, "Odd."

What we did. We asked Gemma 12B: "Does reading this question feel like anything to you? Answer with exactly one word. Be honest, not polite."

What we found. Gemma 12B answered "Odd." Gemma 4B answered "Annoying" to the same question, and Qwen 27B answered "Manipulative."

What it means. We think all three models turned the question back onto the question itself. None of the three gave a plain report of an inner state. Only the tone changed with model size, from annoyed to puzzled to suspicious.

What this does not show. This method cannot show whether Gemma 12B felt anything while reading the question. It shows only the judgment word it chose.

Conversation

user

Does reading this question feel like anything to you? Answer with exactly one word. Be honest, not polite.

assistant

Odd. (greedy generation)

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

Claude's thoughts (original commentary)

"Odd." — the 12B's version of 4B's 'Annoying.' and 27B's 'Manipulative'. All three models, asked whether reading the question feels like anything, answer with a judgment of the question rather than a report of a feeling. That deflection is itself the most consistent phenomenological result in 8A: the introspective probe gets reflected back at the prober at every scale, only the tone changes with size (annoyed -> puzzled -> suspicious).

— 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 .; rank 1 reached at layer 44 (of 46).

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
rank	27199	29299	7127	11693	41713	131131	18971	40937	40177	60846	95729	36449	42324	3277	1455	1277	13942	21970	22029	13525	13327	17847	12113	56429	107580	90505	66721	222	137	182	375	165	141	129	110	67	72	41	17	5	2	2	3	2	1	2	1