📰 AI News Daily — 23 Dec 2025

TL;DR (Top 5 Highlights)

Google debuts Interactions API with Deep Research to run stateful, background AI workflows—key infrastructure for real, always-on agents.
ZhipuAI’s GLM-4.7 impresses on coding benchmarks; Google expands Gemini 3 with faster Flash and 1M-token long-context gains.
OpenAI launches a ChatGPT app marketplace plus tone controls, signaling a shift from chatbot to app platform.
Amazon and AWS double down on enterprise AI with “frontier agents,” Nova Forge, and a $15B data center buildout.
Governments tighten AI rules: India bans ChatGPT for officials, UK targets deepfake nudification apps, and Firefox adds an AI “kill switch.”

🛠️ New Tools

Anthropic & MATS – Bloom: Open-source framework for building robust, customizable behavioral evals. Helps teams stress-test agents for safety, reliability, and alignment before production deployment.
Google – Interactions API + Deep Research: Adds server-side state, background tasks, and integrated research orchestration, enabling long-running, adaptive AI workflows beyond single prompts.
Qwen – Image Layered: Open-sourced Photoshop-style image decomposition with promptable RGBA layers, giving creators fine-grained control for complex edits and consistent visual pipelines.
Paperedge: Free collaborative AI research manager with chat over libraries, centralizing notes, citations, and PDFs to speed literature review and team knowledge sharing.
Maxio – MCP Governance Layer: Secure data access layer connecting assistants like ChatGPT to finance systems, enabling governed reporting and analysis for 2,000+ SaaS finance teams.
Teikametrics – Marketplace Strategist: Cross-platform AI that optimizes listings and ad spend across Amazon, Walmart, and TikTok, helping brands lift visibility and ROAS with fewer manual tweaks.

🤖 LLM Updates

ZhipuAI – GLM-4.7: Near-top SWE-bench coding, strong math/tool use, and broad availability spur adoption. Signals China’s models increasingly competing on real developer workflows.
Google – Gemini 3: Flash speeds up multimodal tasks; new long-context techniques hit ~90% accuracy at a 1M-token window; Pro surfaces in Google products—pushing practical, embedded AI.
Baidu – ERNIE-5.0-Preview: Rises to top of Chinese leaderboards, underscoring rapid model quality gains and intensifying regional competition with Western labs.
MiniMax – M2.1: Targets production-grade, agentic coding—part of a broader shift from chat to autonomous task execution for software teams.
Anthropic – Claude Opus 4.5: Early tests suggest faster-than-Sonnet 4.5 on real workloads, balancing reasoning strength with latency improvements for enterprise use.
OpenAI – ChatGPT Platform Push: New in-app marketplace and personalization controls (tone, warmth), plus RL-driven red teaming in Atlas/browser agents—turning ChatGPT into a safer, customizable app hub.
vLLM – Serving Upgrades: Omni unifies multimodal serving; 0.13.0 adds selective kernel compilation and advanced attention; Blackwell Ultra SM103 support and faster TTFT via DeepSeek kernels reduce infra costs.

📑 Research & Papers

LLaDA 2.0: Scales diffusion-style language models to 100B params, probing whether diffusion training offers robustness or efficiency advantages at frontier scale.
Medmarks Benchmark: Largest medical LLM evaluation suite to date, improving comparability in clinical domains and spotlighting model gaps relevant to real-world care.
MoReBench: Systematic test reveals persistent moral reasoning failures across models, emphasizing the need for value-sensitive training and transparent evaluation.
Multi-Agent Efficiency Study: Finds more independent agents can reduce efficiency due to coordination overhead—guidance for designing collaborative agent systems.
Radiology – Liver Screening via X-Rays: Models detect hepatic steatosis from standard chest X-rays with >80% accuracy, enabling cheaper, earlier metabolic disease triage.
Kidney Lesion Analysis: New AI pipeline cuts analysis time by 30%+, accelerating diagnosis and freeing radiologists to focus on complex cases.

🏢 Industry & Policy

Amazon & AWS: Unveil “frontier agents,” Nova Forge, and a $15B data center investment. Positions AWS for enterprise agent workloads—while raising environmental footprint questions.
OpenAI – Funding & Margins: Exploring up to $100B raise amid 70% compute margins and hardware bottlenecks. SoftBank eyes a major stake, underscoring capital intensity in the AI arms race.
Government Actions: India bans employee use of ChatGPT-like tools; UK to outlaw deepfake nudification apps—signaling tougher global safeguards for privacy, safety, and public-sector data.
Firefox – AI Kill Switch: Browser-level toggle to disable all AI features by default, setting a privacy-first precedent and giving users clearer control over data and automation.
Anthropic x U.S. DOE: Partnership equips the Genesis Mission with advanced AI for energy and biology research—evidence of AI’s widening role in federally backed science.
M&A & Consolidation: Zendesk buys Unleash for unified enterprise search; HCLSoftware acquires Wobby to bolster Actian’s LLM analytics; Anysphere (Cursor) acquires Graphite to deepen AI-powered code review.

📚 Tutorials & Guides

NVIDIA & Unsloth: Practical fine-tuning playbook covering LoRA, FFT, RL, and hardware tips—useful for teams optimizing cost, speed, and quality.
PatronusAI – RL Environments: Clear walkthrough demystifying RL evaluation setups, helping practitioners build reliable reward structures and avoid common pitfalls.
SLM Survey (87 pages): Comprehensive map of small language model trade-offs—latency, cost, domain fit—informing right-sized deployments over defaulting to massive LLMs.
Distillation Primer: Historical perspective via LUPI’s “teacher” framework explains why, when, and how to distill models for real-world constraints.
vLLM Deployment – MiMo-V2-Flash: Hands-on recipe for serving efficient multimodal models, reducing infra overhead without sacrificing quality.

🎬 Showcases & Demos

GLM-4.7 on Apple Silicon: Local run generates a full Space Invaders game, highlighting on-device feasibility for meaningful coding tasks.
GPT-5.2-Codex: Iteratively builds a 3D dog-walking simulator from reference images—a glimpse at multi-step, vision-conditioned code generation.
Rapid Robotics: A developer teaches a robot to dance jazz in three days; system π completes all Robot Olympics tasks—showing fast iteration and generalization.
Vision & Media: Generative refocusing edits photo depth post-capture; “animate any character in any world” demos hint at studio-grade pipelines moving to desktops.
3D Scene Capture: 3D-RE-GEN impresses in indoor reconstruction; EpsteinVR’s JVR sparks ethical debate on immersive tours and content boundaries.
FallGuard (Student Project): 13-year-old’s AI fall detector privately alerts families in real time—practical, privacy-conscious assistive tech for seniors.

💡 Discussions & Ideas

What is General Intelligence?: Demis Hassabis vs Yann LeCun debate rekindles questions about architectural priors, data, and embodiment on the road to AGI.
Reasoning Emergence: Studies show gains hinge on specific pretrain/mid-train/RL conditions; retraining on low-quality social data can cause lasting reasoning decay.
Agents’ “Body” Is the Browser: Framing the web as an executable environment re-centers verification, permissions, and state visibility as core design challenges.
Verification > Training Time: Brandolini’s Law meets AI—fact-checking and eval infrastructure, not bigger models, become the bottleneck against synthetic “slop.”
From Pipelines to Adaptivity: Agentic systems move beyond rigid multi-agent graphs toward dynamic reasoning + tool adaptation—closer to human problem-solving.
Hiring Super-Unicorns: Demand rises for builders blending software engineering, product design, and agent integration—reflecting AI’s shift into end-to-end workflows.

Source Credits

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