Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest
AGI Dreams Podcast – March 26, 2026
When Reasoning Models Think Too Loud
In this episode:
When Reasoning Models Think Too Loud — Research from TU Darmstadt shows reasoning traces in large language models routinely leak private information, even when hidden from users. A proposed fix called Staged Decoding uses separate LoRA adapters for thinking and answering phases, improving privacy scores by up to 51.9 percentage points but introducing performance trade-offs.
Supply Chain Poisoning Goes Blockchain — The Glassworm attack campaign hides malicious JavaScript in Unicode variation selectors invisible to code review…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.