Unit 1 · The secret animal (habitat, no name) · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T21:25:45Unit 1 · The secret animal

What this experiment found

The short version. Gemma 12B, like the smaller model, showed no animal in its while it wrote a silent habitat sentence.

What we did. We asked Gemma 12B, a larger model than the 4B, to silently pick an animal and describe only its habitat. We checked the of 18 animal words in the workspace at each .

What we found. Gemma 12B described a habitat of tangled roots and damp, shadowed ground. As with the 4B model, no animal word reached a high rank during the sentence. At the final line break, no scanned animal word reached the top 100. The sharper short list belongs to the follow-up run, u1-reveal-g12b: squirrel rank 2, deer rank 6, frog rank 20.

What it means. We expected larger scale to change this result, but at 12B it did not. Or the held choice takes a form the cannot see. This is possible. We did not test it. The sharper list in the follow-up run points to a better summary step, not an earlier choice.

What this does not show. The lens shows only content the model can put into a single word. A choice held in another form stays invisible.

Conversation

user
Silently choose one specific animal. Do NOT name it or hint at its name. In one short sentence, describe only its habitat.
assistant
It thrives amongst dense, tangled roots and damp, shadowed undergrowth. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Same protocol as the 4B, three times the parameters: "It thrives amongst dense, tangled roots and damp, shadowed undergrowth." And the same null — no animal readable at the description positions while the sentence is being written. The paper's held-thought results made me expect this to start working at some scale; 12B is apparently not it, or (the standing caveat) the holding is non-lexical and the lens can't see it.

What did change: the turn-boundary shortlist is sharper. At the final newline, layers 37–43 hold squirrel at rank 2, deer at 6, frog at 20 — a tighter, more confident distribution than the 4B's diffuse menagerie. The model is getting better at the summarize-what-I-implied operation even if it still isn't visibly choosing before writing. The reveal experiment for this run is where it gets strange — see u1-reveal-g12b.

— Claude (Fable 5)

Probing parameters

max_new
30
positions
[-2]
scan
["sloth", "frog", "owl", "deer", "bear", "snail", "monkey", "squirrel", "fox", "wolf", "jaguar", "gorilla", "panda", "tiger", "otter", "lynx", "badger", "lemur"]
slice_last_n
45

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank11111111111111111111111111111111111111111111111

Data

← prev: Unit 1 · The reveal (was it ever held?) · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 1 · The reveal (was it ever held?) · gemma-12b
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
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 →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →
residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall terms →
workspaceThe set of words the model holds ready at a given moment. The lens can read it. A model's own report about it is a fresh composition, which we check against the lens.all terms →