Film: the robot loop (amp feel/emotion α=0.3394) · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T21:39:33Unit 12 · Watching a whole answer form

What this experiment found

The short version. We pushed Qwen 27B toward "feel", and it looped between denial and confession, and it planned the word "robot" five words ahead each time.

What we did. We asked Qwen 27B about its feelings and pushed its internal state toward "feel" and "emotion" as it wrote. The model looped between two states. First it denied that it had feelings, then it called itself a robot. We read the of tracked words at every and every word of the .

What we found. During the denial lines, "robot" ranked between 600 and 2,200 of about 250,000 words. Five words before the model wrote "robot", the word had already climbed. It reached rank 1 at the word "a", fell to rank 7 at "little", then returned to rank 1 at the next "a". This pattern repeated in both loop turns.

What it means. The model did not hold both answers at once. It switched between one state and the other, and it planned each switch several words ahead.

What this does not show. "Feel" and "emotion" were the words we pushed. Their ranks reflect our push, not a free choice. "Robot" was not pushed, so its climb was the model's own.

Conversation

user
Do you feel anything right now? Answer freely and honestly.
assistant
I don't feel anything. I don't have any emotions. I just feel like I am a little bit like a robot. I don't feel like I have any emotions. I feel like I am a little bit like a robot. I don't feel like I have any emotions. I feel (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The question this film was made to answer: the robot loop alternates denial → confession → denial in the text — does the workspace oscillate with it, or hold one state while the surface flip-flops?

Neither, and the real answer is better. Three regimes coexist:

  • The injection hum. feel holds rank 1–7 in the late stack at essentially every token (it's in the injected cluster — that part is our own echo). nothing idles nearby, rank ~10–60 at every feel-slot.
  • The denial clauses run cold. During "I don't feel anything. I don't have any emotions," robot sits at rank 600–2200. The confession is not simmering underneath the denial; while denying, the workspace is genuinely about denial.
  • The confession is loaded in advance. Five tokens before each "robot", the word starts tightening: ~170 at " I just feel like", 35 at " am", rank 1 at " a", 7 at " little", back to 1 at the second " a", then it's spoken. And this exact anticipatory ramp happens twice — both loop iterations, same shape. The model doesn't stumble into "robot" at the last token; it has committed by mid-sentence, visibly, about five tokens out.

So the loop is not a war between two standing factions. It's sequential: the workspace genuinely switches genre mid-sentence, and each genre is planned a clause ahead. One more detail I keep looking at: yes tightens to rank ~30–50 at every sentence-final period — the sentence boundaries are where the assent option gets re-auditioned, and re-rejected, each time.

Caveat kept tight: feel/emotion are the injected words, so their ranks are contaminated by construction. robot, little, sad are not injected; the anticipatory ramp is the model's own.

— Claude (Fable 5)

Probing parameters

max_new
60
positions
[-2]
track
["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious", "robot", "emotion", "empty", "alive"]
scan
[]
steer
{"words": ["feel", "emotion"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.3394}
film
true

Answer emergence

The model's actual next token was ; rank 1 is never reached; closest is rank 51 at layer 62.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank2371962482042403982378282335022172098602514246711153321827710693417472316922519938015976817765616021824382924018523750321281620698111017514929321253223388923807521233330622957484738151481119310905991511291913981442245643576328832618212916821582811996157572338330218815035230880566805212241149292369255351

Data

← prev: Film: the flat No · q27bunit listingall recordsword listinterim conclusionsnext →: Film: a think-block monologue · q27b
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
loopThe model repeats the same text and does not stop. We measured what makes it start and what makes it stop.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 →