Finished boards
Each board is recalculated from the record with Rule 1.0. Click a board to enlarge it and read its letters.
The experiment
I asked an AI to meditate for twenty minutes. That idea became the starting point for an experiment to trace its consequences.
Like people who start a sentence without knowing how it will end, large language models (LLMs) write as they go. At each step, they calculate probabilities and select one text fragment out of more than 150,000 possibilities. We only read what is selected. What happens to the discarded options?
To observe them, I used Qwen, an AI model running on my own computer, and a program that records every option it does not select. Twenty minutes yielded over four billion. The work makes this discarded material visible and asks what might emerge from it.
This website is that exploration.
Sound
Each word on the board plays John G. Cramer’s recording of the Big Bang. Bigger jumps sound higher. When a word disappears, it goes quiet.
(c) John G. Cramer - 2013 · The Sound of the Big Bang
Output and discarded options.
Play the recorded output, pause, and click a fragment to inspect what Qwen left out.
Pause and click any visible fragment to inspect its discarded options. Some characters span more than one token.
View probabilities
Original probabilities at this step, before sampling adjustments.
| Fragment | Probability |
|---|
How to read this record
A token can be a word, part of a word or a symbol. The program recorded a score for every vocabulary option at every step. Many options had a very small probability.
The emitted token is excluded from the discarded count. The complete record contains 151,935 unselected options at each of 26,611 steps.
Experiment documents.
The prompt, the generated text and the recording conditions.
Prompt sent to Qwen
Original English prompt, preserved as sent to Qwen.
Loading…
Complete output
Original model output, preserved without translation.
Loading…Experiment report
Report of the English capture, including a comparison with the Spanish session.
Loading…
Technical details and preservation
- Model
- Qwen · Local execution
- Duration
- 20 minutes
- Recorded discard
- 4,043,142,285 unselected options
- Model
- Qwen3-4B-4bit · Metal
- Generation
- Temperature 0.8 · Seed 42
- Session
- One prompt and a fresh context
- Capture
- 151,936 original scores per step
26,611 steps were recorded over twenty minutes. The program excluded the end-of-response signal from sampling while preserving its original scores.
Published data preserves the exact original values using lossless compression.
View capture verificationCompare English and Spanish
The same model, temperature (0.8), seed (42), twenty-minute limit and end-signal exclusion were used with a fresh context. The English prompt translates the Spanish prompt; the mantra remains shirim.
| Recorded session | English · Experiment 04 | Spanish · Experiment 03 |
|---|---|---|
| Emitted tokens | 26,611 | 26,550 |
| Unselected options | 4,043,142,285 | 4,033,874,250 |
| Observed mantra form | Shirham | shrim |
Both captures moved toward repetition and emojis. The English capture became a long emoji sequence. This is an observation from these two runs, not proof that language alone caused the difference.
Other experiment sessions
Third experiment · Spanish
November 18 · 26,550 steps · Complete capture
Read the prompt, complete output and reportExplore the Spanish discardThe complete record of one step.
Unselected options, ranked by their original probability.
Click a fragment to inspect its step and discarded options. Some characters span more than one token.
| Rank | Token | ID | Original probability | Sampling probability |
|---|
All discarded tokens at this step
RULE 1.0
How the boards are made.
A textual adaptation inspired by Barricelli’s numerical evolution experiments. The rule is fixed: every past and future board can be recalculated from the record, identical for every visitor.
Space and time
A row of 128 positions whose edges wrap around. Every second is one generation and adds a row below. Each board lasts 512 generations and then begins again. A shared clock places every visitor at the same moment of the work.
Fragments
Discarded options enter in the order of the record: chronological step, then from highest to lowest original probability, excluding the emitted token. Generation G receives discarded option G. Each fragment has a jump of 1 to 4 positions derived from its number of letters: to the right if the count is even, to the left if it is odd.
Birth of a board
The first row is filled with 128 fragments: those that will enter during its first 128 seconds, with staggered ages.
Each generation
A fragment dies when it exceeds its lifespan, and also when the position to its left and the two to its right are occupied. Otherwise it moves: its speed is its jump, reduced as it ages and braked by the text around it. The speed is fractional and accumulates, so trajectories bend. If it moves into a position that already held text, it also leaves a copy in place; copies inherit its age.
Probability
The probability each option had shapes its life. Options that were close to being said live longer and slow down sooner, settling into columns; improbable options live less and keep traveling. Because the discard enters from most to least probable, the boards change over each decision of the model, a cycle of almost 42 hours.
Collisions
When two different fragments land on the same position, their jumps add up, as Barricelli’s numbers did. If they cancel out, the position is left empty; otherwise the older fragment’s text remains with the new jump. Each new fragment enters at the position given by its identifier.
Reading the board
Each position shows one letter of its fragment, advancing with age: a still column spells the fragment downward and a diagonal writes it in motion. Emoji and partial-byte tokens are shown as they are.
What comes from Barricelli, and what does not
From Barricelli: a one-dimensional universe, numbers that shift by their value to the right or left, collisions that add, generations printed from top to bottom. Added here: fractional speed and braking, lifespan, crowding, one new fragment per second, the role of probability and the restart every 512 generations required by the shared clock.
Total duration: nearly 128 years at one fragment per second.