📰 AI News Daily — 28 Nov 2025

TL;DR (Top 5 Highlights)

Google launches Gemini 3, pushing agentic automation and multimodal reasoning back to the forefront of the AI race.
OpenAI confirms a vendor-side data breach via Mixpanel, severs ties, and expands enterprise data residency to 10 regions to rebuild trust.
Major publishers file a $10B copyright suit against OpenAI and Microsoft, intensifying legal battles over training data and AI outputs.
DeepSeek-Math-V2 achieves IMO gold-level performance with open Apache-2.0 weights, raising the bar for open reasoning models.
Real-world constraints bite: an MIT study estimates 11–12% of U.S. jobs are automatable by agentic AI, while CEOs warn of looming power and data center shortages.

LangChain Deep Agents launch secure sandboxes for remote, reproducible code and bash execution, giving teams safer, auditable agent workflows for real production use.
n8n + MCP integrations let ChatGPT and Claude search, trigger, and manage workflows directly, tightening the loop between LLMs and real business automations.
fal FLUX.2 [dev]/[pro] ships to production with free daily comparisons; Z-Image-Turbo now runs on Hugging Face via fal, boosting fast, aesthetic image generation.
vLLM Ray APIs streamline high-throughput MoE inference, simplifying distributed serving and cutting latency for large, sparse reasoning systems at scale.
Google Antigravity IDE debuts with Gemini 3 for smarter coding assistance and real-time debugging, aiming to accelerate developer productivity across experience levels.
OpenAI Sora (Android) rolls out globally, bringing AI video creation to mainstream mobile users and expanding short-form creative workflows beyond text and images.

Google Gemini 3 arrives with stronger language, vision, and reasoning, enabling richer enterprise automations and sharpening competition across multimodal and agentic tasks.
DeepSeek-Math-V2 posts IMO gold and near-Putnam perfection with verifier-driven training; Apache-2.0 weights on Hugging Face make elite math reasoning broadly accessible.
INTELLECT-3 (PrimeIntellect) scales RL to 100B+ MoE, reporting SOTA in math, coding, and reasoning while open-sourcing recipes for training large sparse models.
Microsoft Fara-7B targets computer-use with an agentic small language model, signaling momentum toward lightweight, specialized assistants for on-device tasks.
Anthropic Claude Opus 4.5 closes the design loop—generate, critique, iterate—and shows early PC control demos plus “think inside files” mode for deeper reasoning traces.
Benchmarks stay tight: Claude Opus 4.x edges Gemini Pro 3 on scientific tasks; community code arenas show narrow spreads among top-tier models.

JECS GWAS dataset (Japan) releases childhood genetics data across 1,148 traits, creating a landmark public resource for developmental and epidemiological research.
Sparse attention wins recognition via Qwen and DeepSeek papers, underscoring practical efficiency gains for long-context and high-throughput transformers.
Mean-pooling emerges as a simple, strong context compression method, improving reliability and reducing complexity compared to heavier-weight summarization techniques.
Beluga (CXL memory) cuts LLM latency by ~90% and boosts throughput 7.35x, pointing to near-term performance wins via composable memory architectures.
P1 physics model achieves gold-level performance at the International Physics Olympiad using reinforcement learning, showcasing progress in non-language reasoning.
MIT labor study estimates agent-based AI could automate tasks tied to ~11–12% of U.S. jobs, informing urgency around reskilling and policy planning.

OpenAI confirms a Mixpanel-linked breach exposing some API user names, emails, and locations; ends the partnership, warns on phishing, and tightens vendor security reviews.
A coalition of news outlets sues OpenAI and Microsoft for $10B over alleged copyright misuse, escalating pressure to clarify AI training and fair-use boundaries.
The UK’s AISI launches propensity evaluations and allocates over £15M for alignment research, strengthening institutional capacity for safety and evaluation.
Satya Nadella and Sam Altman warn of an energy and data center crunch, signaling that power availability may cap AI growth more than chips or capital.
OpenAI adds enterprise and API data residency across 10 regions, helping clients meet compliance obligations and strengthening trust for regulated deployments.
Researchers flag critical security gaps in popular stacks (e.g., Llama, TensorRT), urging stronger defenses around protocols like MCP and model-serving layers.

Anthropic shares operational guidance for long-running agents, emphasizing memory design, context hygiene, and failure handling for dependable production behavior.
Hugging Face explains modern inference—continuous batching, KV caching, chunked prefill, and smart decoding—showing how each technique lifts throughput in practice.
Deep dives into multi-vector retrieval (e.g., mgrep) show how token-efficient indexing improves recall and reduces context budgets for large-scale RAG systems.
Builder summit notes cover quantization, attention optimization, and multi-node deployment, offering practical playbooks for cutting costs without sacrificing quality.
A top course on Deep Representation Learning releases full slides and videos, giving practitioners a rigorous foundation in modern representation techniques.
Weekly research roundups distill advances in reasoning, simulation, small multimodal models, and agent design, saving time for practitioners tracking fast-moving work.

A recreation of Anthropic’s “Eiffel Tower Llama” demo uses sparse autoencoders to steer model activations, with a live walkthrough of interpretability in action.
Seven LLMs roleplay a coordinated Mafia game with synthetic voice acting, highlighting progress in multi-agent planning, deception, and conversational dynamics.
Early demos show Claude Opus 4.5 controlling a PC to handle complex tasks, hinting at near-term viability for robust, general-purpose computer-use agents.
Flux 2 launches an AI art contest spotlighting the latest models’ aesthetics and control, fueling community benchmarks for creative quality.
Colombian students win the NASA Space Apps Challenge with “Exoplanet Hunter AI,” turning real data into discovery tools and demonstrating global AI talent.

Ilya Sutskever and others argue breakthroughs require original research, not just 100x scaling; NVIDIA findings echo diminishing returns on naïve scale-ups.
Governance debates heat up: critics challenge x-risk framings, while some accuse Anthropic of leveraging security narratives to discourage open-source competition.
Enterprises increasingly prefer open and sovereign AI for privacy and control; Cohere–SAP deepen partnership to operationalize regionally compliant deployments.
Studies suggest idea diversity improves research agents, activation probing can reduce sycophancy, and some hallucinations mirror biases in human training data.
Builders report messy reality for agents: durable memory, feedback loops, monitoring, and context engineering—not prompts alone—decide reliability and ROI.

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