Computing & AI · the experiment
Let the model rewrite itself
A language model with its own weights in reach, a damaged copy to repair, and four sizes of the same family. Four questions, from beginner to expert, on what self-rewriting achieved.
1 · Know what was forbidden
Beginner
What was the model not allowed to do during its sessions?
2 · Count the extra right answers
On 490 problems it never saw, with eight answers each, the base model is right 57.2% of the time; after its best edit, 63.8% of the time.
Intermediate
How many more of these answers are right after the edit?
3 · Read the repaired weights
In the damaged rows, the largest weight was 0.071 in the original model and 10.1 after the damage. After the model's own repair, it is 0.108.
Expert
What do these three numbers show about the repair?
4 · Compare the smaller models
The 4- and 9-billion-parameter models had the same task, the same tools and the same limits as the 27-billion model.
Expert
What did they do differently?
Rewrite reviewed
—