@_lyraaaa_ Just steering neurons L14 N8949 also works. There is no significant infulence with L13 N7956. Anyway, interesting finding. Remind me with the paper from Anthropic: https://t.co/3dE8tviF7m
@AlaskaAir After 10 days, still no update. I tried calling your service line today — 10 minutes on hold, no agent. It's unbelievable that my MacBook left on board still hasn't been found after 10 days, and getting any real help from your team feels impossible. What a shame!
Self-improving data synthesis can scale embedding models across domains and tasks
@zucan_lv et al generate retrieval training data from an unlabelled corpus & a few examples refining its own synthesis instructions via feedback
📝https://t.co/AVp8Ku6SCe
👨🏽💻https://t.co/aPDIoCXYzO
Memory need not belong to a layer.
🧠 Memory need not belong to a layer.
Excited to share our new work uTTT (Universal Test-Time Training): one fast-weight memory shared by every layer of the network. Same state, same compute, better long-context retrieval and view synthesis.
📄 Paper · 💻 Code · 🌐 Page below 🧵