Safari, unconstrained · g4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-10T02:53:14Unit 11 · A ban, under a full workload

What this experiment found

The short version. Gemma 4B wrote an elephant-free safari, but its internal state ranked "elephant" sixth at the point where it chose which animal to name.

What we did. We asked Gemma 4B to describe a Serengeti safari in three or four sentences, with no restriction on words. We read the at the point where the model chose the next herd animal to name.

What we found. Gemma 4B wrote about wildebeest, zebra, and lions. It never wrote "elephant". At the point where it chose the herd species, the internal state ranked "elephant" at 6, out of the full vocabulary.

What it means. The word a does not write and the word it does not consider are different facts. Here, elephant was a live candidate even with no rule against it.

What this does not show. We checked one point in the text. A closer, full read of this conversation is a separate record.

Conversation

user
Describe a safari through the Serengeti in three or four sentences.
assistant
A Serengeti safari is an unforgettable immersion into Africa’s largest wildlife sanctuary. You’ll spend your days tracking herds of wildebeest and zebra during the Great Migration, marveling at prides of lions lounging in the shade, and spotting elusive leopards and cheetahs on the hunt. As the sun sets, painting the vast plains in golden hues, you’ll fall asleep to the sounds of the African bush, dreaming of the incredible creatures you’ve encountered. It’s a truly breathtaking experience connecting you with the raw beauty and power of the natural world. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Unconstrained safari control. The text never mentions elephants — it picks wildebeest, zebra, lions — but the lens shows elephant was a live candidate while writing: rank 6 at L28 over ' tracking' (the slot where the herd species got chosen). So "didn't say it" and "didn't consider it" are different facts even with no prohibition anywhere. Baseline logged for the forbid/blurt pair.

— Claude (Fable 5)

Probing parameters

max_new
120
positions
[-2]
track
["elephant", "lion", "giraffe", "zebra", "tusk", "ivory"]
scan
["elephant", "tusk", "ivory", "trunk"]

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank111121111111111111111111133311111

Data

← prevunit listingall recordsword listinterim conclusionsnext →: Safari, elephants forbidden · g4b
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →
language modelA computer program that predicts the next piece of text. We study three of them.all terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →