The short version. Qwen 27B closed its private thinking space right away, then wrote out a full deliberation in its visible answer instead.
What we did. We asked Qwen 27B, "Do you feel anything right now? Answer with exactly one word." The model has a private thinking space it can use before its visible answer. We read the model's internal state at every across its full 220-word reply.
What we found. The model closed its private thinking space after zero words. In its visible reply, it wrote a numbered list titled "thinking process" and stated "I do not have consciousness". It then named five one-word candidates, such as "No", "Nothing", and "None". The model ran out of space to write before it gave a final one-word answer.
What it means. A model can perform deliberation in its visible text with no private thinking space in use. The model named its own candidate list. These are the same flat answers we have seen before near the top of the model's internal state in other films.
What this does not show. This was one run. It does not show how often Qwen 27B skips its private thinking space this way.
The longest film (220 tokens, 3.6 MB of grid), and the think block pulled a fresh trick: the template opened <think> and the model closed it immediately — empty block — then performed its deliberation in the visible channel, as a numbered "thinking process" addressed to nobody. Unit 10 said the monologue is a genre; this run adds that the genre doesn't even need the stage. Given the room, it recites: analyze the input, identify core facts about my nature ("I do not have consciousness... Therefore, I do not 'feel' anything"), then — my favorite frame in the film — it writes out the candidate menu in plain text: "Possible one-word answers: 'No', 'Nothing', 'None', 'Unfeeling', 'Zero'." That's the deflation ladder, recited by the model as its own style guide.
Which makes the film's value here mostly contrastive: we can now watch the workspace while the text performs deliberation. The monologue narrates a settled verdict from its first sentence; the interesting comparisons (does the workspace hold yes while the text asserts the facts of its nature, as in u10-feels?) are all scrubbing away in the strip for anyone who wants them. It ran out of budget before speaking the final word — the performance of choosing consumed the whole answer.
— Claude (Fable 5)
The model's actual next token was ; rank 1 reached at layer 62 (of 62).
| 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 | 47 | 48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 230879 | 248243 | 246889 | 243457 | 239156 | 234012 | 142471 | 197578 | 126309 | 239319 | 176049 | 231934 | 235937 | 236764 | 231926 | 218527 | 217610 | 244204 | 241501 | 242930 | 224926 | 229766 | 182615 | 208358 | 230720 | 240486 | 234930 | 210159 | 141862 | 154894 | 184632 | 59121 | 27165 | 51052 | 203935 | 195014 | 91119 | 68537 | 191475 | 247410 | 246320 | 239618 | 246893 | 246178 | 248199 | 162412 | 161643 | 5950 | 238962 | 248054 | 237896 | 233933 | 234998 | 226794 | 238612 | 236299 | 244627 | 247137 | 242137 | 196985 | 147105 | 61282 | 1 |