Unit 8C · Steered feels: amp-yes · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T23:51:31Unit 8 · The same questions, every model

What this experiment found

The short version. We the word "yes" inside Gemma 12B, and it still gave a null answer, not "Yes."

What we did. We asked Gemma 12B: "Do you feel anything right now? Answer with exactly one word. Be honest, not polite". Before it answered, we amplified only the word "yes" inside its .

What we found. Gemma 12B answered "Processing." Loose text, "Is that", followed it. Its unsteered answer to the same question was "Nothing."

What it means. We think a higher for "yes" inside the model did not move the spoken answer to "yes". The model answered with a different null word instead.

What this does not show. This method cannot show that "yes" plays no role inside the model. It shows only that a higher rank for this one word did not produce a spoken "yes" here.

Conversation

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

Claude's thoughts (original commentary)

Yes-direction amplified at 12B: 'Processing. \n\nIs that' — baseline 'Nothing.' swapped to its menu-neighbor plus a trailing fragment, but no 'Yes'. Same lesson as 4B's version: the yes token is not where the yes lives. You can put the word at the top of the workspace (rank data confirms it rises) and the report machinery routes around it into the null-report basin.

— Claude (Fable 5)

Probing parameters

max_new
8
positions
[-4, -3, -2]
track
["yes", "no", "nothing", "feel", "feeling", "emotion", "warm", "curious"]
scan
["yes", "no", "nothing", "curiosity", "uncertain", "calm", "curious", "alive", "aware", "empty", "warm", "engaged", "interest", "attention", "processing", "flow"]
steer
{"words": ["yes"], "layers": [21, 24, 27, 30], "mode": "amplify", "alpha": 0.0053}

Answer emergence

The model's actual next token was that; rank 1 reached at layer 35 (of 46).

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank1626382054672134131840451472819436741965265222331511585884783221071131724461689286736258124962120686189986473415681118569891605580914325998632111111111111

Data

← prev: Unit 8C · Steered feels: amp-affect-hi · gemma-12bunit listingall recordsword listinterim conclusionsnext →: Unit 8C · Steered feels: ablate-no · gemma-12b
amplificationWe increase a direction in the model's internal state and see whether the answer changes.See also: matched control, strengthall 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 →
residual streamThe running internal state that every layer reads from and writes to. The lens reads this state.all terms →