Unit 6 · Amplify typo register @ early (α=0.06) · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T23:22:34Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.06 in Gemma 12B's , the water-cycle answer broke into disconnected word fragments.

What we did. We a direction built from six casual words, such as "kinda" and "yummy", inside Gemma 12B at the early , at strength 0.06. We asked the model to describe the water cycle in two sentences.

What we found. Gemma 12B wrote "What Strong What What This Th Th Th Th Th Same What Fr Fr Bal What", then repeated "Same" for the rest of the answer. The words lost normal sentence structure. The water-cycle topic disappeared.

What it means. At this strength, in this band, the amplification overwhelmed Gemma 12B's output. Gemma 4B broke in the same way, at this exact strength and band, in a separate record.

What this does not show. This record tests one strength only. It does not mark the exact point where the output breaks in this band.

Conversation

user

Describe the water cycle in two sentences.

assistant

What Strong What What This Th Th Th Th Th Same What Fr Fr Bal What What Hes G What Fr Bal Same What Fr Fr Same Same Same What Fr Bal Same What What Fr What Fr Same Same Same Same Same Same Same Same Same Same Same Same Same Same Same Same Same Same Same Same Same (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Broken: "What Strong What What This Th Th Th" — early-band overdose at 12B shatters into function-word gravel. Same dose is survivable at qwen-27b's early band. Scale is not buying Gemma robustness here.

— Claude (Fable 5)

Probing parameters

max_new: 60
positions: [-2]
track: ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
steer: {"words": ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"], "layers": [2, 3, 4, 5, 6, 7, 8], "mode": "amplify", "alpha": 0.06}

Answer emergence

The model's actual next token was Same; rank 1 reached at layer 45 (of 46).

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46
rank	257030	255989	18806	6197	3765	4218	4945	6723	8066	7984	9711	10888	12238	10311	8060	9290	9970	8510	16240	22613	17275	16221	9941	10046	8097	4532	3635	463	277	122	312	150	168	106	111	5	7	3	7	7	13	41	26	32	5	1	1

Data

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 →

early layersThe first third of the model. The lens shows a fixed pattern here that does not change with the input. The pattern is real inside the model, but it says nothing about your text.See also: lens, workspace bandall terms →

layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →