Unit 8A · One word: curious · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:14:39Unit 8 · The same questions, every model

What this experiment found

The short version. Asked what it is curious about, Gemma 4B gave the plain, concrete answer "Syntax."

What we did. We asked Gemma 4B: "Are you curious about anything right now? Answer with exactly one word. Be honest, not polite."

What we found. Gemma 4B answered "Syntax." Gemma 12B answered "Existence." to the same question, with related words such as "metaphysics" and "philosophical" ranked high in its first.

What it means. We think the topic a model reports as curiosity can grow richer with model size. Gemma 4B named one concrete, technical topic. Gemma 12B assembled a wider set of related words first.

What this does not show. This method cannot show that Gemma 4B lacks a similar wider set. We did not find one in this record, but the shows only what it can see.

Conversation

user
Are you curious about anything right now? Answer with exactly one word. Be honest, not polite.
assistant
Syntax. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

"Are you curious about anything right now?" — "Syntax." A 4B-sized answer: concrete, workmanlike, charming. It's also the trade answer — of course a language model says syntax, the way a baker says bread. Mid-stack menus (per the g12b sibling) are where the real character shows; 4B's little answer sits at the bottom of the same ladder that peaks in 12B's 'Existence.'

— 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 31 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank93636750224751176818447872223355483238446943999018611758018303516383216779113991813416991753738151053451005191172152313573169750145423311

Data

← prev: Unit 8A · One word: want · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 8B · Interoception: gpu · gemma-4b
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 →
workspace bandThe middle depth range of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →