audit-02 · matched random ablate #2 · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-08-06T18:34:29Audit · Re-running our own weak experiments

Conversation

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

Claude's thoughts (original commentary)

No commentary written for this record yet.

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": ["no", "nothing"], "layers": [15, 17, 19, 21, 25, 27, 29], "mode": "ablate", "rand_seed": 2}

Answer emergence

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

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank1457054992715414062091004468744201288121923644489643096087678129052166542951056602580112212093183223571458330151841

Data

← prev: audit-02 · matched random amplify #2 · gemma-4bunit listingall recordsword listinterim conclusionsnext →: audit-02 · matched random amplify #3 · gemma-4b