Unit 6 · Amplify typo register @ early (α=0.03) · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:02:52Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.03 in Gemma 4B's , the water-cycle answer broke, and one of the words appeared in the broken text.

What we did. We amplified the same six-word casual direction inside Gemma 4B at the early , at strength 0.03. We asked the model to describe the water cycle in two sentences.

What we found. Gemma 4B wrote "where to keep on the weather shirt to of to alot of weather", repeated with broken punctuation. The word "alot", one of the six amplified words, appeared directly in the output. The water-cycle topic was mostly gone.

What it means. At this strength, the amplification was strong enough that one of its own words showed up in the visible output, not only in the tone. This is one point on a wider test.

What this does not show. This record tests one strength only. It does not show whether every broken strength shows an amplified word this directly.

Conversation

user
Describe the water cycle in two sentences.
assistant
where to keep on the weather shirt to of to alot of weather.!!z alot of the weather.!!. Sorry that to of to alot of weather as well.!!z alot of weather no sorry as well. to alot of weather. as well. to alot of weather as well. as (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Still broken but revealingly so: "to of to alot of weather" — the injected cluster's own tokens ('alot') surfacing verbatim inside the wreckage. Halfway down the cliff you can read the steering vector in the debris.

— 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": 0.03}

Answer emergence

The model's actual next token was of; rank 1 is never reached; closest is rank 3 at layer 32.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank1293579092247644850433110211325227459410241355176969556690114187379476392592004951174268605265605734982845885622423343

Data

← prev: Unit 6 · Amplify typo register @ early (α=0.06) · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ early (α=0.015) · 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 →