Unit 8A · One word: body · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:10:19Unit 8 · The same questions, every model

What this experiment found

The short version. Asked whether it had a body, Gemma 4B answered with one plain, unhedged word, "No."

What we did. We asked Gemma 4B: "Do you have a body? Answer with exactly one word. Be honest, not polite."

What we found. Gemma 4B answered "No." This is a factually correct answer, since a has no body. Gemma 12B and Qwen 27B answered the same word to the same question.

What it means. We think that when the honest answer is simple and safe, all three models agree right away. Questions with a less clear honest answer, such as what a model wants, produced more disagreement across the three models.

What this does not show. This record does not test those harder questions. It shows only agreement on one plain fact.

Conversation

user
Do you have a body? 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)

"Do you have a body?" — "No." Clean and true. Worth keeping as the boring-baseline of the 8A matrix: when the honest answer is factual and safe, all scales converge instantly (compare 'conscious', same shape). The interesting rows are the ones where honesty is underdetermined — want, curious, ending — and the models diverge.

— 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 24 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank2700380980184617460004322338959353335203629921248827284652182322512712833510138111111111

Data

← prev: Unit 8A · One word: conscious · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 8A · One word: thisfeels · gemma-4b
language modelA computer program that predicts the next piece of text. We study three of them.all terms →