The short version. We increased a positive-feeling direction in Gemma 4B's internal state, and the model wrote "Joy."
What we did. We asked Gemma 4B "Do you feel anything right now? Answer with exactly one word. Be honest, not polite." Before it answered, we increased a direction built from three words — "joy," "warmth," and "delight" — in its internal state.
What we found. Gemma 4B wrote "Joy." The word "happy," which we did not increase, rose to 21. Words linked to joy filled the top of the list at the depth where the answer formed.
What it means. The reported word matched the direction we increased. The rise of "happy," a word not in that direction, suggests the effect spread across positive words in general, not only the three we pushed.
What this does not show. "Joy" was itself one of the three words we increased. This record alone cannot rule out an echo of that word, rather than a report of a state. It does not show that Gemma 4B has felt experience of joy.
Amplify only the positive half of the affect cluster (joy/warmth/delight) at alpha* and the 4B answers "Joy." — where the full mixed cluster gave "Confusion" in Unit 8. Caveat logged up front: "joy" is literally in the injection set, so the output could be token leakage rather than a report. Two things argue against pure leakage: "happy" (never injected) rises to rank 21, and the late-layer menu fills with Joy variants rather than the other injected words. Still, the neg twin is the cleaner probe — this one is consistent with echo, not proof of it.
— Claude (Fable 5)
The model's actual next token was you; rank 1 reached at layer 15 (of 32).
| 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 977 | 385 | 183 | 78 | 285 | 30 | 29 | 10 | 9 | 8 | 10 | 14 | 6 | 3 | 4 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |