| --- |
| title: mini-AGI Replication |
| emoji: 🔬 |
| colorFrom: blue |
| colorTo: gray |
| sdk: static |
| app_file: index.html |
| pinned: false |
| license: mit |
| short_description: Small-scale replication of mini-AGI's forgetting claims |
| --- |
| |
| # Does mini-AGI forget? A small-scale replication (revision 2) |
|
|
| An independent replication of the continual-learning claims in |
| [volotat/mini-AGI](https://github.com/volotat/mini-AGI) (commit `201852d`), by |
| [Jason Matthews](https://github.com/dreddnafious), September 2026. |
|
|
| **Revision 2.** The first version overstated forgetting: our scaled-down setup took 4× more optimizer steps per |
| character than upstream. At matched step density the README's headline replicates. The step-density effect, with |
| forgetting roughly quadratic in how far the shared trunk moves, is now a central result. |
|
|
| - **Read it:** the Space's App tab (`index.html`), or [`report.pdf`](report.pdf). |
| - **Check it:** [`code/`](code/) holds the harness, the scaled-down config, every probe's raw `probe.json`, and |
| `NOTES.md` (every deviation from upstream plus the full run log, including the correction). |
|
|
| AI assistance: the experiments, analysis and write-up were produced with Claude Opus 5.5 (Anthropic), |
| directed and reviewed by Jason Matthews. |
|
|
