Thursday, June 25, 2026

Anthropic accuses Alibaba of illicit distillation of Claude. Anthropic alleges Alibaba ran a brazen campaign to extract its model capabilities, igniting debate over distillation ethics and IP. Source

Europe pushes back on Washington's chip war. European policymakers resist US chip export controls amid escalating AI trade tensions. Source

AI hasn't killed engineering jobs, they've proven the most resilient. Recent labor data suggests engineering roles have weathered AI-driven displacement better than other fields. Source

Federal appeals court sanctions lawyers for AI-fabricated citations. Two attorneys penalized for submitting case citations generated by AI. Source

Krea 2: SOTA open-weights 12B image model. Krea releases a 12B image model with full technical report on training and infrastructure. Source

Gefen: drop-in AdamW replacement with 8x memory reduction. New optimizer claims dramatic training memory savings; paper and code available. Source

OpenAgents releases Khala. Open-source orchestrator for models, agents, and tools exposed via an OpenAI-compatible API. Source

Qwen releases world model benchmark. Frontier-trajectory benchmark spans 7 agent domains on Hugging Face. Source

Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining. Investigates how pretraining asymmetrically retains or forgets learned rules. Source

Model Forensics: Misalignment or Context-Dependent Artifacts?. Examines whether concerning LLM behavior reflects true misalignment or situational artifacts. Source

Are We Ready for an Agent-Native Memory System?. Explores requirements and open questions for memory systems built natively for LLM agents. Source

The Unfireable Safety Kernel. Proposes an OS-level execution-time safety kernel to constrain AI agent behavior. Source

Improved Large Language Diffusion Models. Advances text generation via large language diffusion models. Source