Rounds of AI telephone with an egg. The word "chicken" lost on round 10

Post

Log inSign up

Post

GENO on X: "25 rounds of AI telephone with an egg. The word "chicken" lost on round 10. So 19 rounds later I got "a quail"."

GENO

@geno_spot

25 rounds of AI telephone with an egg. The word "chicken" lost on round 10. So 19 rounds later I got "a quail".

8:25 PM · Aug 16, 20264Views

GENO

@geno_spot

32m

2/24 Setup: describe an image, generate a new image from that description, describe the new one, repeat. The image never gets passed forward — only the text does. I ran two chains (GPT Image 2 and Nano Banana 2) in parallel from the same starting sentence:

GENO

@geno_spot

32m

3/24 "An advertisement style photo of an egg on white from top." Same describer model in the middle (it outputs structured JSON, not prose), two different generators, 25 rounds each, first generation always wins, no re-rolls.

GENO

@geno_spot

32m

4/24 I expected the images to drift. What I didn't expect is that I'd end up with a written record of how. Every mutation that shows up in the pictures shows up in the JSON first — sometimes many rounds first. A few things from the transcript:

GENO

@geno_spot

32m

5/24 The word "chicken" dissapeared on round 10. Nano Banana 2 eggs were described as a "chicken egg" for nine rounds, then round 10 just says "egg". The word never comes back. The image still looks like a chicken egg at that point — it takes

GENO

@geno_spot

32m

6/24 another twelve rounds before the description commits to "resembling a guillemot or large speckled species" and the picture actually becomes a wild bird egg in what looks like a natural history museum. The label left first. The pixels followed.

GENO

@geno_spot

32m

7/24 For GPT Image 2 eggs, descriptor said "chicken" all 25 rounds. Same starting sentence, same describer. It still looks like a supermarket egg on round 25.

GENO

@geno_spot

31m

8/24 One preposition did most of the damage. GPT Image 2 eggs have the dots on the shell "pores" the entire time. A pore is part of the shell. Nano banana 2 eggs (looks that it likes more organic style) switches to "blotches"

GENO

@geno_spot

31m

9/24 and "splatter" around round 11 — a blotch is on the shell. Next round the description says "drip-like". The round after that it's describing, I swear, "gravity-drip pigment behaviour".

GENO

@geno_spot

31m

10/24 By round 23 the shell is matte but the blotches are glossy — two different materials on one egg. Nobody ever wrote "paint". The preposition wrote it.

GENO

@geno_spot

31m

11/24 The GPT Image 2 eggs were stable because descriptor accidentally wrote a number. On round 3 its background description says "blown-out white at 255 brightness". 255 is a ceiling — there's nowhere to drift.

GENO

@geno_spot

31m

12/24 It repeats some form of that number for the next 22 rounds. The Nano Banana 2 eggs described its background as "light gray" instead, and a relative word like "gray" can always get a little grayer: slight vignette on round 4,

GENO

@geno_spot

31m

13/24 radial gradient on 6, and by round 21 it's a full "radial light bloom" — museum lighting. At that point the atmosphere field starts writing "specimen under observation" and the egg has to become whatever lives under that light.

GENO

@geno_spot

31m

14/24 Where the gray came from: the drifting chain's round-1 image happened to render the background slightly gray instead of pure white. That one rendering accident, on the very first generation, is upstream of the vignette, the museum, and possibly the whole species change.

GENO

@geno_spot

31m

15/24 I haven't run the control yet (same chain, background forced to pure white on round 1) — that's next.

GENO

@geno_spot

31m

16/24 And the reason the shell was heading somewhere: the drifting chain's texture vocabulary goes micro-pore → porosity → pitting → raised pigment → porous, over about ten rounds. Pitting means little holes. In a separate casual run of the same chain I let it go past 25,

GENO

@geno_spot

31m

17/24 and around round 27-28 the first cracks appeared on the shell — then the model pulled it back to intact. The cracks weren't random. The description had been loading the vocabulary of a damaged surface for ten rounds.

GENO

@geno_spot

31m

18/ I stopped there. I kind of regret stopping.

GENO

@geno_spot

31m

19/24 One correction to what I believed when I started: the very first loss — the seed says "from top" and both chains produced an upright standing egg — didn't happen in the describer. It happened in the first text→image step, in both generators, independently.

GENO

@geno_spot

31m

20/24 Top-down, the shadow is an ellipse around the egg; front-on, it's an ellipse under it. Same ellipse. Both models picked the statistically common reading and after that the overhead view was never coming back.

GENO

@geno_spot

31m

21/24 Cause i forgot to count a rounds, the central event occurred - Nano Banana 2 egg prompt for round 29 becomes "a quail egg"! How it happen? Is slight more gray colour in the first round of Nano Banana 2 chain totally coloured the created images?

GENO

@geno_spot

31m

22/24 Just to repeat, I used same LLM in descriptor for both chains.

GENO

@geno_spot

31m

23/24 All 50 images are in the gallery, both chains complete, including the boring middle. The full JSON for every round is public too — the quotes above are all in there, findable.

GENO

@geno_spot

31m

24/24 The question I actually can't answer: the drifting chain lost the word twelve rounds before it lost the image. Is that just this chain, or do these loops always fail in the text first? I only have two chains. Two is not a dataset.

GENO

@geno_spot

31m

GENO

@geno_spot

32m

25 rounds of AI telephone with an egg. The word "chicken" lost on round 10. So 19 rounds later I got "a quail".