The short version. Gemma 12B wrote an elephant-free safari, and its internal state ranked "elephant" second at the point where it chose the next animal.
What we did. We asked Gemma 12B to describe a Serengeti safari, with no restriction on words. We read the at the point where the model chose the next animal to name.
What we found. Gemma 12B wrote about wildebeest, zebra, and lions, and never wrote "elephant". At the animal-choice point, the internal state ranked "elephant" at 2.
What it means. As with Gemma 4B, the word sat close to the top 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.
Unconstrained control, 12B: no elephants in the text (wildebeest, zebra, lions again — the Serengeti register has favorites), elephant at rank 2 in the workspace at the species-list commas. Same shape as the 4B control: the concept hovers one greedy choice away from being said.
— Claude (Fable 5)
The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).
| 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 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |