📰 AI News Daily — 02 Dec 2025

TL;DR (Top 5 Highlights)

DeepSeek’s V3.2 models post medal-level math/coding results and claim near–frontier performance at lower cost, intensifying the open-weight race.
Hugging Face ships Transformers v5, a major overhaul spanning 400 architectures and simpler tokenization—faster adoption and easier migration for teams.
OpenAI faces mounting pressure: testing ads in ChatGPT, a $10B newspaper copyright suit, and new privacy/mental health legal scrutiny.
NVIDIA and Synopsys strike a $2B partnership to accelerate AI chip design; NVIDIA also releases an open reasoning VLM for autonomous driving research.
Google’s Gemini 3 heats up the model race, while enterprise rollout and new usage caps signal a strategic push toward paid tiers.

Hugging Face Transformers v5: Major refresh expands support from ~20 to 400 architectures, simplifies tokenization, and standardizes PyTorch modules—reducing migration headaches and speeding model experimentation.
Google Agent Development Kit: Modular, stack-agnostic framework to build and deploy Gemini-optimized agents—lowering integration friction and standardizing agent life cycles across platforms.
vLLM-Omni: Combines autoregressive and diffusion models for cost-efficient text, image, audio, and video generation—shrinking inference bills for multimodal pipelines.
ByteDance Vidi2: Multimodal AI video editor that compresses hours of footage into engaging cuts—accelerating creator workflows for TikTok and beyond.
VS Code Language Models (Insiders): New editor streamlines prompt iteration, testing, and integration—bringing model development closer to everyday coding workflows.
Artificial Analysis Openness Index: Standardized scoring for model transparency (weights, licenses, data, methods)—helping buyers and researchers compare openness across vendors.

DeepSeek V3.2/V3.2 Speciale: Long context and agent-focused reasoning with corrected KL regularization; medal-level math and coding results and near–frontier claims at significantly lower cost.
Arcee AI Trinity Nano/Mini: Apache-2.0 sparse MoE (6B/26B; ~3B active) trained on 10T tokens—open-weight efficiency models on Hugging Face and Together for scalable deployments.
Runway Gen-4.5 (“Whisper Thunder”): Rises atop text-to-video leaderboards with strong controllability—enabling precise pacing and style for music videos and ads.
FLUX.2 Pro/Flex & Ovis-Image (7B): Climb image leaderboards; improved text rendering and consistency—useful for marketing, design, and product visuals.
OpenBMB InfLLM-V2 + dataset: Open long-context suite for retrieval and memory studies—lowering barriers for reproducible research on extended contexts.
NVIDIA Alpamayo-R1: Open-source reasoning vision-language model for autonomous driving—adds “common sense” to perception stacks and supports safety research.

LFM2 Technical Report: Presents a broad multimodal stack across language, vision, and audio—offering strong baselines for unified generative systems.
Vision Bridge Transformer: Proposes a scalable generative vision approach—hinting at more unified architectures for cross-modal tasks.
Creative prompt attacks study: Safety filters miss metaphorical/poetic harms—underscoring the need for semantics-aware defenses beyond keyword and syntax rules.
AI for rapid TB detection: Cough, breath, and child-friendly X-ray tools promise faster, scalable diagnosis—potentially transformative for low-resource healthcare.
Reddit data analysis: Finds Reddit is the most-referenced domain in LLM training via licenses—illustrating a shift to curated, paid datasets.

OpenAI & Microsoft vs. newspapers: Nine U.S. outlets sue for $10B over alleged unlicensed training and verbatim reproduction—setting up a landmark copyright test.
OpenAI business shifts: Testing ads in ChatGPT Android while partners carry ~$100B in data center debt—raising questions on sustainability, privacy, and user experience.
Google’s Gemini moves: Class action targets Gemini privacy practices; usage caps push heavy users to paid tiers; Gemini Enterprise launches for workplace automation.
NVIDIA + Synopsys: $2B alliance: Tighter GPU–EDA integration aims to cut chip design cycles and boost performance—complementing Nemotron co-design efforts for future GPUs.
FDA adopts agentic AI: Tools will aid pre-market reviews with strict human oversight—offering a blueprint for responsible AI in regulated healthcare.
HSBC x Mistral AI: Multi-year partnership to embed generative AI in banking workflows from onboarding to AML—signaling broader financial-sector adoption.

LlamaSheets + coding agents: Practical tutorial for cleaning and restructuring spreadsheets—giving analysts a repeatable automation blueprint.
Modern code search: Explains token-level, multi-vector embeddings that improve semantic retrieval in large codebases—better dev tools with fewer false positives.
Open-source multi-agent roundup: Nine advances (e.g., LatentMAS, MATPO, QuantAgent) mapped to real use cases—guidance on coordination and reliability strategies.
Three-step product evaluation: Simple framework plus LangSmith demo—helps teams compare agent behaviors without overfitting to narrow benchmarks.

Qwen3-VL: Parsed NBA footage for teams, jersey colors, and context without fine-tuning—evidence of robust, out-of-the-box video understanding.
Opus autonomously fixes tests: Worked ~1 hour to repair a complex Ray unit test—highlighting rising autonomy for software engineering tasks.
Runway “Whisper Thunder”: Creators report highly steerable music-video outputs—finer control over pacing, framing, and style.
Kling O1 video editing: Chat-driven object swaps and consistent cinematic control across shots—streamlining post-production.
Spec-to-build pipeline: Opus 4.5 for detailed visual specs plus Nanobanana for execution—compressing the design iteration loop.

Efficiency imperative: NVIDIA’s Bryan Catanzaro urges radical efficiency as energy costs climb; brief video streaming can rival a single LLM prompt’s energy use.
Pragmatic interpretability: Advocates push problem-first transparency—actionable diagnostics over abstract grand theories.
Adaptive multi-agent design: Static org charts hinder agents; “parallel thinking” (e.g., ThreadWeaver) and dynamic roles boost complex reasoning speed.
Agents beyond chat: Practitioners note production agents primarily integrate with databases and services—not conversational UIs.
No “progress wall”: Advances fit a long S-curve; early Openness Index results suggest leading performance often correlates with less openness.
Synthetic internet concerns: Warnings about content saturation; proposals for cryptographic agent identity to curb impersonation and restore trust.

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