📰 AI News Daily — 29 Dec 2025

TL;DR (Top 5 Highlights)

• OpenAI is hiring a $555K Head of Preparedness, signaling tougher safety oversight amid legal pressures and intense competition.
• Meta released the RPG dataset (22,000 tasks) on Hugging Face to accelerate “AI co‑scientist” systems with built‑in evaluations.
• Google deepens Gemini’s reach: smarter photo scouting in Maps, Android Assistant replacement by 2026, and rumored paid Chrome AI tiers.
• GLM‑4.7 leads open‑source rankings as world models surge; local inference speeds up on Apple silicon via MLX.
• Samsung unveils its first 2nm chip; China proposes stricter AI rules—hardware and policy are reshaping AI’s next phase.

🛠️ New Tools

• LangChain: An interactive LLM inference visualizer shows token flows and context effects in real time, helping developers debug prompts and improve reliability during production deployments.
• A3B & 8B model weights: Fresh checkpoints on Hugging Face enable hands‑on experimentation with recent training runs, speeding community benchmarking and fine‑tuning efforts.
• Google Gemini in Maps: Conversational location scouting for photographers suggests vantage points, lighting, and crowd levels—turning trip planning into a creative co‑pilot for better shots.
• World App (OpenAI/Tools for Humanity): A biometric super app combining identity, encrypted messaging, and crypto payments aims to fight deepfakes and fraud while simplifying user verification.
• 1inch Network + SavantChat: AI‑powered audits for DeFi transactions promise faster threat detection and lower costs, improving trust and security across decentralized finance ecosystems.
• YAKSH (Uttar Pradesh Police): An AI‑enhanced app adds facial recognition, voice search, and gang analytics to identify suspects and track crime networks, modernizing law enforcement workflows.

🤖 LLM Updates

• GLM‑4.7: Tops independent rankings, reinforcing open‑source momentum and offering strong baseline performance for coding, reasoning, and agentic tasks without closed‑source dependencies.
• 2025 World Models: LeJEPA, Dreamer 4, Genie 3, Cosmos WFM 2.5, and Code World Model highlight a shift toward models that reason over time and control environments, not just predict tokens.
• Qwen3: “Attention sink” analysis shows specialized handling of key tokens, informing better context usage, prompt design, and efficiency strategies for long‑context applications.
• MiniMax‑M2.1 on Apple M3 Ultra (MLX): Strong local inference performance underscores rapid on‑device gains, reducing latency, cost, and privacy risks for desktop‑class deployments.
• vLLM (MLX backend teased): Native Apple silicon acceleration promises faster throughput and lower memory overhead for Mac‑based inference, aiding local development and testing.
• Google Gemini on Android: Gemini’s overlay will replace Assistant by 2026, enabling uninterrupted multitasking and richer, persistent AI sessions that blend on‑device and cloud capabilities.

📑 Research & Papers

• Meta RPG Dataset: 22,000 structured research tasks with rubrics and references aim to speed “AI co‑scientist” development, enabling reproducible evaluation of reasoning and tool use.
• Egocentric2Embodiment: Converts first‑person videos into structured Q&A to bridge perception and physical intelligence, improving grounding for robots and embodied agents.
• Video Zero‑Shot Transfer: New video models demonstrate strong task transfer without retraining, hinting at a step‑change for vision systems in robotics, navigation, and surveillance.
• CuTe DSL Kernel: A compact TV‑layout kernel outperforms Torch RMSNorm on B200 GPUs, showing targeted kernel engineering can deliver outsized performance and cost efficiency.
• SonicMoE: IO‑aware and tile‑aware optimizations streamline Mixture‑of‑Experts throughput, improving expert routing efficiency and lowering inference latency for scaled deployments.

🏢 Industry & Policy

• OpenAI (Head of Preparedness): A $555K role to lead risk mitigation across cybersecurity, misuse, and mental health reflects escalating safety expectations and pre‑regulatory alignment.
• Nvidia: Jonathan Ross becomes Chief Software Architect, signaling a deeper push into advanced AI software stacks alongside hardware leadership to sustain competitive moat and developer ecosystem.
• Samsung: First 2nm chip targets next‑gen performance and efficiency, strengthening mobile and edge AI capabilities and tightening the hardware‑AI co‑design feedback loop.
• China Draft AI Rules: New proposals require algorithm checks, safety protections, and strict content limits—raising compliance costs while pushing standardization and accountability.
• Google Chrome (Paid AI): Code hints at Google AI Pro/Ultra tiers for premium features like agentic browsing and summarization, foreshadowing browser monetization in 2025 and beyond.
• AI “Slop” Concern: Over half of English web content may be AI‑generated, driving demand for higher‑quality, user‑first design and better detection to restore trust.

📚 Tutorials & Guides

• Policy Optimization Beyond PPO: A technical review of GRPO, DR.GRPO, GSPO, DAPO, and variants helps researchers modernize RL pipelines for stability, sample‑efficiency, and safer exploration.
• Latent Space Year‑End: Recaps on OpenAI’s Codex and GPT‑5 expectations plus interviews on the Agentic Web provide pragmatic guidance for building AI‑native software teams and products.

🎬 Showcases & Demos

• AGI Documentary: A behind‑the‑scenes feature drew massive viewership, illustrating public appetite for transparent narratives about frontier AI research and its societal stakes.
• Diesol’s “The Cleaner” (Rome): Long‑form AI cinema with an Emmy‑winning original score showcases maturing production pipelines and creative control for narrative‑driven generative filmmaking.
• DJ Reachy (Robotics + Music): Real‑time music generation and synchronized dance, released open‑source, demonstrate playful human‑robot collaboration and reproducible creative robotics.
• Reachy Mini: A compact tabletop robot earns praise as an approachable maker platform, lowering the barrier to entry for hands‑on robotics experimentation.
• Kling 2.6: Improved motion precision and stability in animation control push video generation toward production readiness for advertising, entertainment, and design workflows.

💡 Discussions & Ideas

• Memory Systems for Agents: Analyses highlight how storage and retrieval design shape agent reasoning, suggesting vector databases and episodic memory as core architecture choices.
• ARC Prize Takeaways: Discipline and non‑LLM methods can win hard benchmarks, reminding teams to combine symbolic tooling and search with LLMs for robust problem solving.
• Agentic Workflows: Engineers report big coding productivity gains, but foresee testing and verification as critical new roles as orchestration replaces expert‑only execution.
• Claude Code as “Agent”: Rapid adoption hints at mainstream agent experiences inside familiar coding tools, emphasizing UX and reliability over raw model horsepower.
• Policy Fragmentation: State‑level patchworks risk chilling innovation; calls grow for a uniform federal framework as tax proposals on unrealized gains raise incentive concerns.
• From Fringe to Core: Historical shifts in neural nets inform leaders’ predictions that 2026 rewards production‑grade results over demos; Andrej Karpathy expects a pivot to logical “ghost intelligence.”

Source Credits

Curated from 250+ RSS feeds, Twitter expert lists, Reddit, and Hacker News.