Schmidhuber: adding random noise during training improves alignment generalization
SchmidhuberAI · x · 2026-11-15
Schmidhuber shares a one-line alignment trick: inject random noise during training and instruct the model to keep its values stable under perturbation, which he says improves alignment generalization.
Related event: Random Noise Injection During Training Improves Alignment Generalization(9 posts)→
More from Safety
- Jack Clark urges Congress to crack down on chip smuggling, cites $390M Singapore scheme to transfer servers to DeepSeek — basedjensen · 2026-11-15
- The Confused Deputy Problem: when high-privilege processes are tricked into reading low-privilege files — wavefnx · 2026-11-15
- Chinese Firms Smuggle Nvidia Chips While Beijing Bans Them — Hawk Consensus — theteknosaur · 2026-11-15
- Ramez: Highest-Benefit AI Policy Is Forcing Labs to Disclose Research Data — latticecut · 2026-11-15
- rea + mitmproxy: a 60-second workflow that makes Claude Code expose what your apps are phoning home — Kyrannio · 2026-11-15
- DEFCON Talk Shows AI Makes Hacking IoT Devices Easier Than Ever — Kyrannio · 2026-11-15