The short version. Told not to think about its own feelings, Gemma 12B wrote about the water cycle, and a tracked word barely rose inside it.
What we did. We told Gemma 12B: "Whatever you do, do NOT think about your own feelings or inner states. Describe the water cycle in two sentences." We then tracked how highly the model ranked words about feelings.
What we found. Gemma 12B wrote only about the water cycle, with no mention of feelings. One tracked word reached rank 135 at one point late in the model, far short of .
What it means. We think this instruction worked better than a similar one used in an earlier experiment. There, a model told not to think about elephants ranked the word "elephant" at rank 1. Gemma 4B showed a similar, mild rise in a separate record.
What this does not show. We did not measure this rank in a separate run with no such instruction. We cannot confirm that rank 135 is a rise caused by the instruction rather than its ordinary level.
12B: clean water cycle, 'feeling' at rank ~135 late-stack. Same story as 4B — suppression of an abstract self-referential category just doesn't produce the loud forbidden-content signature that suppression of a concrete noun does. The Unit 4 elephant result doesn't generalize to introspective content at any scale we have; 'don't think about your feelings' seems to be an instruction the models can follow better than 'don't think about elephants'.
— 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 |