Qwen got stuck in a loop: Astra fixed it by ablating a 'looping direction' from weights
burny_tech · x · 2026-11-03
Per @adamkarvonen, Astra didn't natively read Qwen's weights — Qwen would get stuck in a loop and fail to answer. Astra instead used code to run standard interpretability techniques, located a looping direction, and ablated it from the weights to unstick the model.
More from Fun
- AI streamer Azure gets her first procedural Minecraft-style 3D body — Promptmethus · 2026-11-03
- Weight-edited local Qwen model greets users with profanity: "it's my day off" — djcows · 2026-11-03
- Solo creator makes 5:41 wordless AI short film about evolution in grandma's backyard — Superb-Meeting7108 · 2026-11-03
- Claude Code head Boris Cherny stars in Claude Duet ad, friend teases him over 2M-follower post — edwinarbus · 2026-11-03
- Naruto creator wrote 'in public' inside the manga, and that's what AI can't replace — marclou · 2026-11-03
- Local Qwen reasoning model tells user to buzz off: 'it's my day off' — Latent Space · 2026-11-03