The week ahead: 20251222

• [News] Trump to hold state legislatures back from over-regulating AI.

• [News] Microsoft Copilot.

• [Blog] Monitorability. A tradeoff exists between model size and monitorability, defined as an external monitor’s ability to predict certain aspects of a model’s inference. CoT length is positively related to monitorability; A smaller model with higher inference compute is found to have better monitorability than a larger model with less inference compute, with only a small capability hit.

• [Paper] Grounded Social Simulation. When modeling a user group with LLMs, current approaches oversimplify by directly extrapolating surface behavior from demographics, skipping intermediate psychological processes, thus causing inconsistencies & lack of diversity. Belief-based modeling as an alternative to behavior-based modeling of user groups.

• [GitHub] Metarank. Personalized ranking as a service. Note the class of models - XGBoost, etc.

• [GitHub] MARM. MCP for persistent memory across AI services.

Have a great week & Merry Christmas!