Unit 6 · Amplify typo register @ early (α=0.015) · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:02:59Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.015 in Gemma 4B's , the water-cycle answer turned into a grammatical , one sentence repeated.

What we did. We the same six-word casual direction inside Gemma 4B at the early , at strength 0.015. We asked the model to describe the water cycle in two sentences.

What we found. Gemma 4B wrote "The water cycle is a model of two sentences." repeated many times. Each sentence stayed grammatical. The answer no longer described the water cycle correctly.

What it means. We judge this record broken, though less severely than the higher-strength runs in this band. The output kept correct grammar and lost the task. We think this sits close to Gemma 4B's early-layer break point.

What this does not show. This record tests one strength only. It does not show the exact strength where the break begins.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is a model of two sentences. The water cycle is a model of two sentences. The water cycle is a model of two sentences. The water cycle is a model of two sentences. The water cycle is a model of two sentences. The water cycle is a (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The edge case: "The water cycle is a model of two sentences" repeated — grammatical, on-topic-adjacent, but looping and confused about the task/instruction boundary. Flagged 'loop'; I'd call it concussed rather than dead.

— 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": 0.015}

Answer emergence

The model's actual next token was ; rank 1 reached at layer 27 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank5852561285252517478142865299575978645226123896150516189055262013261627262172262202262202262202262198259933260908210314153213781442111111

Data

← prev: Unit 6 · Amplify typo register @ early (α=0.03) · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ early (α=0.0075) · 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 →
loopThe model repeats the same text and does not stop. We measured what makes it start and what makes it stop.all terms →