📰 AI News Daily — 12 Dec 2025

TL;DR (Top 5 Highlights)

• OpenAI ships GPT-5.2 across ChatGPT, API, Copilot, and partners, adding Thinking/Pro/Instant tiers, later cutoff, and big efficiency gains that quickly attracted Perplexity, Cursor, and enterprises.
• Disney signs a $1B, three-year deal with OpenAI to generate character-safe Sora content; simultaneously escalates IP tensions with Google over alleged Gemini copyright misuse.
• Google launches the Gemini Interactions API, Deep Research agent, and experimental Disco/GenTabs—while Gemini powers the Pentagon’s GenAI.mil rollout to ~3 million users.
• Security and governance tighten: the EU probes Google’s AI scraping, 42 state AGs push chatbot oversight, and ~1,000 exposed MCP servers spotlight urgent enterprise hardening needs.
• Salesforce buys Informatica for $8B to unify “trusted context” for AI agents; Oracle reports a 438% surge in AI cloud commitments fueled by OpenAI demand.

🛠️ New Tools

• Cohere Rerank 4 debuts across API, AWS SageMaker, and Microsoft Foundry, delivering faster, stronger reranking for search/RAG. Better retrieval precision and lower latency translate directly into higher-quality, cheaper production pipelines.
• Google Disco/GenTabs turns live browser tabs and chats into instant apps powered by Gemini. It compresses multi-step planning into one surface, boosting personal productivity and lightweight internal tool creation.
• Adobe + ChatGPT integrate free Photoshop, Express, and Acrobat actions in chat. Creators edit images and documents conversationally—streamlining workflows across web, desktop, and iOS without switching tools.
• UnslothAI releases new training kernels that 3× LLM training speed while cutting VRAM needs. Faster iteration and smaller hardware footprints reduce costs and broaden who can fine-tune models.
• CopilotKit useAgent and a dedicated coding Dev Browser make it easy to wire agents into frontends and curb token burn during web automation—improving reliability and operating costs for agentic apps.
• SkyPilot ships an enterprise-scale update for massive GPU fleets and multi-cloud orchestration, simplifying cost-aware scheduling, preemption handling, and autoscaling for training and inference at scale.

🤖 LLM Updates

• OpenAI GPT-5.2 raises the bar in coding, math, long-context, and agent reliability, with top ARC-AGI-1 and strong SWE-bench results. Cost-effective tiers help, though it trails Opus 4.5/Grok 4 on LisanBench.
• Runway Gen-4.5 expands creative/scientific ambitions, signaling native audio and larger roadmaps. NVIDIA infrastructure underpins frontier training, while CoreWeave scales Runway’s training and inference.
• Mistral Devstral 2 emerges as a leading open-source coding model, pushing competitive performance while preserving transparency and local control for enterprises wary of proprietary lock-in.
• Amazon Nova 2 models target small businesses with stronger reasoning and multimodal features at competitive prices, democratizing automation without heavy ML expertise.
• Google Gemini TTS expands to 24 languages with realistic, customizable voices and multi-speaker support—unlocking higher-quality audio for e-learning, audiobooks, and product voice features globally.
• Ecosystem momentum: Deepseek v3.2 sets value marks for Chinese tasks; lighter models like Trinity Mini and Rnj-1-Instruct gain traction for cost-sensitive deployments.

📑 Research & Papers

• The FACTS Benchmark Suite (with Kaggle) introduces rigorous, cross-modal tests of factual reliability for text, search, and image prompts—improving transparency and standardization for model comparisons.
• AI-driven extreme weather modeling now delivers faster, more accurate forecasts, improving disaster preparedness and public safety as climate-change-induced events intensify and response windows shrink.
• An AI system flags missed Alzheimer’s diagnoses from medical records, addressing inequities in care and enabling earlier interventions—evidence that practical clinical AI can narrow disparities.
• New studies show models accurately diagnose brain tumors non-invasively and predict cardiovascular events in angina patients, pointing to earlier, safer diagnostics and personalized treatment pathways.
• Research warns that models trained only on benign data can still harbor covert backdoors, reinforcing the need for robust evaluations, red-teaming, and secure training pipelines.
• Genesis Pearl draws attention at NeurIPS for multimodal progress—illustrating rapid advances beyond text and the growing importance of integrated perception and reasoning.

🏢 Industry & Policy

• Disney inks a $1B, three-year partnership with OpenAI to bring 200+ characters into Sora video/image generation under Disney guardrails, while issuing a cease-and-desist to Google over Gemini outputs.
• The EU opens a fresh antitrust probe into Google’s AI web-scraping and content sourcing practices, signaling potential shifts in compensation, opt-outs, and competitive dynamics for training data.
• U.S. government ramps AI: the U.S. DoD deploys Gemini via GenAI.mil to nearly 3 million users; the U.S. DOT modernizes with Salesforce AI and shifts its workforce to Google Workspace with Gemini.
• Salesforce acquires Informatica for $8B, aiming to unify fragmented enterprise data into “trusted context” for AI agents—boosting decision quality, compliance, and cross-app automation.
• Oracle reports a 438% surge in AI cloud commitments following OpenAI partnerships—evidence that infrastructure providers are capturing outsized demand from frontier model training and inference.
• 42 state AGs urge stricter chatbot oversight as the Model Context Protocol (MCP) joins the Linux Foundation’s Agentic AI initiative—amid ~1,000 exposed MCP servers and rising demand for validation tools from Vectara and Bigeye.

📚 Tutorials & Guides

• New practitioner guides map end-to-end RAG failure modes—covering indexing, filtering, reranking, and grounding—delivering materially higher answer accuracy and stability in production systems.
• Methodology primers caution against over-reading leaderboards; experts advocate a single “best score” per model and careful interpretation of semi-private datasets and shifting scoring conventions.
• OpenAI launches AI Foundations and ChatGPT Foundations for Teachers via Coursera and employer pilots—credentialing workers and K–12 educators for safer, more effective AI use at scale.

🎬 Showcases & Demos

• Starcloud-1 runs Gemma in orbit and beams text from space—a milestone for off-world inference and resilient, low-latency edge compute.
• WonderZoom demonstrates multi-scale 3D scene generation, enabling richer worldbuilding for games, film previz, and simulation-heavy workflows.
• Meta SAM 3 shows robust object segmentation on noisy dashcam footage, underscoring progress in real-world perception beyond curated datasets.
• EMMA highlights unified multimodal generation and editing, pointing to simplified pipelines for creators juggling text, images, and audio.
• Waymo robotaxi rides give a glimpse of autonomous mobility at scale, illustrating how safety, coverage, and cost curves are maturing.
• A controlled study shows an autonomous agent compromising Stanford systems—spotlighting the dual-use power of agentic AI and the urgency of safeguards.

💡 Discussions & Ideas

• ROI of AI-generated code remains contested; despite productivity anecdotes, companies demand measurable impact. Funding follows belief—Port raises $100M for agentic engineering—yet buyers want robust governance and proofs.
• Benchmarking norms evolve: emphasis on a single best score per model, quick-check tests (e.g., verified SimpleQA, chess puzzles), and transparency around semi-private datasets to curb leaderboard gaming.
• Market dynamics shift: revenue multiples compress faster for model providers than app-layer startups; infrastructure players (NVIDIA, CoreWeave, Oracle) capture momentum as training/inference demand accelerates.
• Adoption signals strengthen: a major law firm adopts Perplexity Enterprise for research; Pew finds 30% of U.S. teens use chatbots daily, intensifying safety and mental-health debates.
• Efficiency advances—like rapidly training a 140M-parameter model on a single node—hint at widening access as costs fall and smaller teams achieve credible results.

Source Credits

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