| --- |
| license: mit |
| tags: |
| - interpretability |
| - jacobian-lens |
| - j-space |
| - self-model |
| - qwen3 |
| pretty_name: The ring (Strange Loop in J-space, Part 3) |
| --- |
| |
| # The ring: a latent loop that keeps a model of itself |
|
|
| Part 3 of [Strange Loop in J-space](https://github.com/Pixedar/jspace-strange-loop#part-3-the-ring). |
|
|
| A Qwen3 model reads one placeholder token at a time. At every step its J-space content, read at 70 % of depth with |
| the Neuronpedia Jacobian lens and non-negative pursuit on common English words (K = 25), is written back at 50 % of |
| depth. No word is sampled inside the loop. Each "mind" has an online self-model of its own transitions. Ten |
| conditions write back the self-model's surprise or its prediction, or controls of equal norm on the same schedule: |
| momentum, another mind's signal, the mind's own surprise from a random earlier step, random input, plain content, or |
| nothing. Minds with the same seed share every random draw. |
|
|
| **Pre-registration.** `code/` was uploaded at 19:03:34 UTC on 9 October 2026, before the first result. `code/stage2/` |
| (the exploratory gain ladder) was added after the Qwen3-4B result and is labelled as such. |
|
|
| ## Layout |
|
|
| | Path | Contents | |
| |---|---| |
| | `qwen3-{1.7b,4b,8b,14b}/sweep/` | pilot sweep of the free substrate (gain × noise) | |
| | `qwen3-*/params.json` | the regime picked by the fixed rule | |
| | `qwen3-*/main/s*_*/ring.npz` | J-space content per step and mind (`C`), top atoms, self-model skill, step similarity | |
| | `qwen3-*/main/results.json`, `summary.txt` | the pre-registered tests | |
| | `qwen3-*/gains/g*/` and `gains/gains.json` | the exploratory ladder across each substrate's critical gain | |
| | `qwen3-4b-long/` | 2,000-step run | |
| | `ladder.json`, `ladder.txt` | all models in one table | |
| | `logs/` | the unattended job's logs | |
|
|
| ## Result in one line |
|
|
| Feeding back the current surprise of a self-model that keeps learning keeps the trajectory exploring. It beats |
| momentum, the mind's own past surprise and random input of the same size, and freezing the self-model removes the |
| effect. Whether the surprise is the mind's own rather than another mind's matters only past the critical gain, where |
| plain momentum resists collapse better. Nothing self-like emerges. Details are in the repository README. |
|
|
