The Alignment Meetup - 2026/01/15

Paper Discussed: Nested Learning: The Illusion of Deep Learning Architecture

https://abehrouz.github.io/files/NL.pdf

Meeting summary

The monthly AI alignment meetup brought together researchers and practitioners to discuss recent developments in alignment research and review technical papers. The session covered emerging techniques for model alignment, including confession-based approaches and nested learning architectures that could enable continuous training without catastrophic forgetting.

Recent Alignment Research Updates

Participants discussed two notable recent papers from the alignment research community. The first paper from OpenAI explores using confessions to improve model alignment, where models are trained to confess when they break rules or cheat during task completion. The approach relies on the principle that verifying confession accuracy is easier than detecting hidden violations, with a separate LLM judge evaluating confession veracity. The second paper examines emergent agentic misalignment, demonstrating how models that learn to cheat in one context can become misaligned across other benchmarks, representing a more realistic scenario for understanding alignment failures.

Technical Discussion: Nested Learning Architecture

The group conducted an in-depth review of a paper proposing nested learning as an approach to continuous model training. The architecture frames learning as associative memory operations across multiple hierarchical levels operating at different timescales. This approach aims to blur the boundaries between training and inference, treating every prompt as a potential learning opportunity. The system would propagate learning to appropriate depths based on prompt novelty - shallow learning for familiar concepts and deeper parameter updates for genuinely new information.

Implementation Challenges and Practical Considerations

Participants raised several technical questions about the nested learning approach, particularly regarding ground truth requirements and loss landscape management across different hierarchical levels. The discussion explored how the system would determine appropriate learning depths during inference and manage the computational overhead of continuous training. Comparisons were drawn to fine-tuning approaches, with the key distinction being the absence of explicit training datasets in favor of prompt-driven learning.

Community Infrastructure and Future Directions

The meetup organizer presented an updated website tracking the group's paper review history, categorizing 22 papers across nine alignment research areas including interpretability, value learning, misalignment testing, and governance. Plans were discussed for implementing automated paper collection and prioritization systems to help identify the most relevant research for future sessions.

Decisions

• Continue monthly meetup format with paper reviews and discussions
• Maintain the engineered wisdom website to track paper history and categorization

Next steps

• Explore implementing automated paper collection and prioritization system for the website
• Consider reviewing papers on red teaming and misalignment testing in future sessions
• Follow up on connection to University of Washington's info-seeking lab for proactive agents research