R
Rikiya Takehi
@rikiyatakehi · 03/11 00:56
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming:
Qui l'a repris
Ce qu'ils en ont dit
Misbah Syed
@MisbahSy
a cité
· 03/11 00:56
Weight-space Programming 🤯
D
Dimitris Papailiopoulos
@DimitrisPapail
a cité
· 02/11 22:39
Astra and Opus likely the best mech interpr researchers around
Ahmad
@TheAhmadOsman
a cité
· 02/11 22:30
LLMs hacking are gonna be so much fun man
Wu Haoning