AGI Dreams Podcast – April 09, 2026

AGI Dreams – Open, Uncensored, & Local - AI News Digest · Apr 9, 2026 · 16m · transcript

MoE Models Hide Refusal in Places You Can't Weight-Bake Away

In this episode:

MoE Models Hide Refusal in Places You Can't Weight-Bake Away — Researchers found that Mixture-of-Experts models encode refusal differently than dense models, with safety refusals routing through specialized expert pathways that resist standard abliteration techniques. MoE architecture is accidentally more resilient to weight-level safety tampering, and reasoning models exhibit dual refusal circuits that can disagree between chain-of-thought and final output.

Activation Watermarks, Egress Firewalls, and the…