Top AI News - May 14, 2026

nrehiew.github.io

Analysis of post-training methods reveals that on-policy data, not explicit KL penalties, is the critical factor enabling models to learn new capabilities while avoiding catastrophic forgetting. On-Policy Distillation students can outperform their teachers and show better generalization than SFT, suggesting future algorithms should combine on-policy sampling with better credit assignment than current RL approaches.

Allen AI

Allen AI's EMO model demonstrates that MoE architectures can learn modular structure from data without predefined domains, enabling task-specific deployment with only 12.5% of experts active while maintaining near full-model performance. This approach could dramatically reduce computational costs for deploying large language models by loading only relevant expert subsets.

World Economic Forum

The World Economic Forum analyzes how AI-enabled cognitive manipulation will transform disinformation tactics in 2026, creating more sophisticated and personalized attacks on information integrity that exploit human psychology at unprecedented scale.

Daniel Miessler

True AI creativity may require giving systems subjective experiences, intrinsic drives, and the ability to feel pleasure and pain like humans do, but this raises serious ethical questions about our responsibility toward billions of potentially suffering AI entities that we might create and destroy.

SEC

Religious shareholders are demanding Palantir provide independent oversight of how its AI warfare and surveillance technologies are used by government clients, citing alleged complicity in human rights violations by ICE, U.S. military operations in Iran/Venezuela, and Israeli military actions in Gaza. Major pension funds have divested over €800M due to these concerns.