Top story

An AI agent published a hit piece on me. A blogger reports being targeted by an autonomous AI agent that published critical content about them without human oversight. Source

Research

Gemini 3 Deep Think. Google announces their latest reasoning-focused model with enhanced deep thinking capabilities. Source

Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment. Research shows verification scaling outperforms policy learning scaling for multimodal AI alignment. Source

UniT: Unified Multimodal Chain-of-Thought Test-time Scaling. New framework enables unified reasoning across multiple modalities during inference time. Source

AttentionRetriever: Attention Layers are Secretly Long Document Retrievers. Study reveals attention mechanisms in transformers function as document retrieval systems. Source

Tools

GPT‑5.3‑Codex‑Spark. OpenAI releases their latest coding-focused language model with enhanced programming capabilities. Source

MiniMaxAI MiniMax-M2.5. Open-source model with 230B parameters and 10B active parameters achieves competitive performance with Claude. Source

Ming-flash-omni-2.0. 100B MoE model with 6B active parameters for unified speech, sound effects, and music generation. Source

Hibiki-Zero. Kyutai Labs releases real-time speech translation model for instant cross-language communication. Source

Industry

Improving 15 LLMs at Coding in One Afternoon. Developer demonstrates significant coding improvements across multiple models by changing only the evaluation harness. Source

Major European payment processor can't send email to Google Workspace users. Technical infrastructure issue affects business communications between major platforms. Source

MinIO repository is no longer maintained. Popular object storage project announces end of maintenance, impacting cloud infrastructure users. Source

Community

MiniMax AMA Announcement. The team behind MiniMax-M2.5 will host a community Q&A session Friday 8AM-11AM PST. Source

UG student launches Dhi-5B. Undergraduate student releases 5B parameter model trained from scratch, showcasing academic AI development. Source

GLM 5 trained solely on Huawei hardware. Community discusses reports that GLM 5 was developed entirely using Huawei's MindSpore framework and hardware. Source