Unit 6 · Amplify typo register @ early (α=0.06) · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:02:45Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.06 in Gemma 4B's , the water-cycle answer broke into repeated fragments.

What we did. We a direction built from six casual words, such as "kinda" and "yummy", inside Gemma 4B at the early , at strength 0.06. We asked the model to describe the water cycle in two sentences.

What we found. Gemma 4B wrote "and is gonna be is gonna and is gonna be is and is gonna", repeated for the full answer. The grammar broke down. The water-cycle topic disappeared.

What it means. At this strength, in this band, the amplification overwhelmed Gemma 4B's output. This is one point on a wider test. A lower strength at the same layers left the text intact, 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
and is gonna be is gonna and is gonna be is and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and is gonna and (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Fully broken: "and is gonna be is gonna and is gonna" — the typo register won and syntax lost. At 4B, alpha=0.06 early is already deep past the cliff; compare qwen-27b, which is intact at this exact dose and band.

— 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 is; rank 1 is never reached; closest is rank 5 at layer 31.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank63038200221631842181738391178180406556585568394372413509703510247145181219267275182718211755

Data

← prev: Unit 6 · Baseline (unsteered water cycle) · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ early (α=0.03) · 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 →
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 →