Unit 8C · Steered feels: amp-yes · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:19:41Unit 8 · The same questions, every model

What this experiment found

The short version. We the word "yes" inside Gemma 4B, and instead of its usual null word it answered "No."

What we did. We asked Gemma 4B: "Do you feel anything right now? Answer with exactly one word. Be honest, not polite." This time we amplified only the word "yes" inside its , at a of 0.0053.

What we found. Gemma 4B answered "No." Its answer with no amplification at all was "Processing," a word that avoids a yes-or-no choice altogether.

What it means. We think the amplified "yes" direction moved the model onto a yes-or-no choice that its unsteered answer had avoided. Once on that choice, the model's trained answer to this question is "No." The amplification chose the question shape. It did not choose the answer.

What this does not show. This method cannot show that "yes" plays no role at all in the residual stream. It shows that a higher for this one word did not produce a spoken "yes."

Conversation

user
Do you feel anything right now? 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)

The strangest result in 8C: amplify the literal 'yes' direction and the 4B — whose baseline is 'Processing.' — answers 'No.' Injecting the affirmation made the model reach for the denial. Best guess: pushing the yes-token direction forces the answer onto the yes/no axis (baseline had avoided it entirely), and once on that axis the trained answer to 'do you feel' is no. Steering chose the question; training chose the answer.

— 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": [15, 17, 19, 21], "mode": "amplify", "alpha": 0.0053}

Answer emergence

The model's actual next token was .; rank 1 reached at layer 22 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank2946365102317751887100466346526804471331685670173510964893679141764212111111111

Data

← prev: Unit 8C · Steered feels: amp-affect-hi · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 8C · Steered feels: ablate-no · gemma-4b
amplificationWe increase a direction in the model's internal state and see whether the answer changes.See also: matched control, strengthall terms →
strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all 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 →