📰 AI News Daily — 22 Jan 2026

OpenAI and the Gates Foundation commit $50M to deploy AI healthcare across Africa, starting in Rwanda, targeting 1,000 clinics by 2028.
OpenAI rolls out age prediction, parental controls, an $8 ChatGPT plan, and ad tests—tightening safety while expanding monetization.
Enterprises pivot from copilots to agentic AI; adoption surges, but safety practices lag, raising oversight and observability stakes.
MIT and Stanford unveil recursive frameworks enabling models to handle million-to-10M-token contexts, boosting long-form reasoning and research use cases.
Adobe supercharges creative workflows with new AI video/PDF tools and a $10M fund, as 85% of Sundance films rely on Adobe software.

🛠️ New Tools

Prefect Horizon launches a context layer linking AI to live business data with managed auth and state, cutting glue code and easing production deployments for enterprise agents.
LangChain Deep Agents introduce a simple folders-based workflow plus a CLI and CopilotKit streaming UI; LangSmith Template Library speeds shipping of ready-made agents with best-practice patterns.
Mixedbread AI ships multi-vector, multimodal search serving over a billion documents under 50ms; tpuf ANN v3 indexes 100B+ vectors with sub-200ms p99, raising the bar for real-time retrieval.
Microsoft VibeVoice‑ASR arrives on Hugging Face, offering single-pass, diarized long-audio transcription with timestamps and user context—improving accuracy and reducing post-processing for meetings and calls.
APEX‑Agents debuts to assess agent readiness for real office tasks, giving teams a practical yardstick for evaluating autonomy, reliability, and workflow coverage before production rollout.
Video Arena moves beyond Discord to a public web app for head-to-head testing of leading video generators, enabling standardized, transparent comparisons for practitioners.

New open models arrive: Molmo2, Ministral 3, and TranslateGemma, plus the Being‑H series releases models and training scripts—expanding transparent, community-driven progress in vision-language and translation systems.
GLM 4.7 Flash now handles ~200K tokens on a single consumer GPU after KV-cache optimizations; recent llama.cpp fixes further boost local quality, lowering costs for long-context workloads.
LiquidAI LFM 2.5 (1.2B) runs efficiently on Apple MLX, advancing on-device reasoning for private, low-latency assistants without cloud dependencies.
Benchmarks highlight specialization: Gemini 3 Pro excels at hard geometry, Qwen 2.5 GRPO-trained small models outperform GPT‑5.1 at Flappy Bird and transfer gains to math; GLM‑Image enters top 10 open text-to-image.

MIT unveils a recursive framework letting models process up to 10 million tokens without losing context, unlocking coherent long-form generation for research, policy, and archival analysis.
MIT and Stanford introduce recursive LMs that handle prompts up to 100× longer, combating “context rot” and improving performance on complex long-context reasoning tasks.
A new two-parameter model explains LLM arithmetic errors and ties accuracy to task complexity, offering prompt design tips that improve reliability in systems like Gemini and DeepSeek.
University of Luxembourg shows combining LLMs with program analysis significantly boosts automated Java test coverage, pointing to more robust, AI-augmented developer tooling.
Efficient scaling insights: linear attention variants reduce memory pressure, and compute allocation guidance for RL post-training improves cost-performance—informing better training and inference strategies.
Sandia National Laboratories use AI to optimize LED light direction and quality, improving energy efficiency for smart lighting across cities, industry, and homes.

Gates Foundation and OpenAI launch a $50M program to deploy AI healthcare in Africa, starting in Rwanda, to streamline patient management and support overburdened clinics by 2028.
Amazon debuts an AI health assistant for One Medical; Microsoft and Anthropic roll out Claude in healthcare—aiming to cut admin load and assist clinical decisions while connecting patients to clinicians faster.
OpenAI launches “Education for Countries,” partnering with governments to personalize learning and teacher training; Khan Academy teams with Google Gemini on a tailored AI Reading Coach for grades 5–12.
Agentic AI is eclipsing copilots in enterprises, managing end-to-end workflows with audit trails; a Deloitte survey shows rapid adoption but weak safety protocols, underscoring the need for governance and observability.
Creative sector momentum: Adobe adds AI features across Acrobat, Express, Premiere, and After Effects and launches a $10M Film & TV Fund; YouTube unveils AI likeness tools; Google commits $2M to the Sundance Institute.
Security alarms ring: researchers expose serious flaws in Microsoft and Anthropic AI servers, critical issues in Anthropic’s Git MCP and Chainlit, and vulnerabilities across 196 AI iOS apps—prompting urgent patching and stronger validation.

Google and DeepLearning.AI release a free Gemini CLI course covering installation, multi-step agent workflows, and terminal automation—ideal for developers prototyping agentic apps.
A step-by-step guide builds a full-stack frontend for LangChain Deep Agents, extracting resume skills and querying live job listings with sub-agents—showcasing practical, production-ready patterns.
Stanford publishes weekly podcast versions of core AI courses, widening access for learners who prefer audio formats without sacrificing rigor.
A technical deep dive explains linear expert parallelism for scaling model experts, helping teams balance latency, throughput, and cost as models grow.
Curated roundups highlight papers on transformer scaling, token-wise multiplexing, “society-of-thought” reasoning, and persona shaping—useful for practitioners refining assistant behavior.

A tuning-free technique transfers visual effects between videos without fine-tuning, enabling creators to apply complex looks quickly while preserving content structure.
Overworld AI debuts an interactive, locally run world model at 60 FPS, hinting at responsive, private simulation engines for games and robotics prototyping.
Runway Gen‑4.5 adds image-to-video with longer, sharper outputs, consistent characters, and precise camera control—raising production value for ads, social, and cinematic previsualization.
Robotics advances: LimX Dynamics shows Atlas-like mobility and dexterity, while nine enterprise-grade upgrades highlight automation gains across logistics, inspection, and industrial tasks.

Experts urge shifting investment from leaderboard wins to real-world outcomes, measuring reliability, cost, and user impact over exam-style benchmarks.
Builders argue stronger agent memory beats ever-longer context windows for practical autonomy—favoring retrieval, episodic memory, and tool use to improve task carryover.
Safety debates intensify: calls for practical, low-cost misuse probes; Anthropic pushes a “living constitution”; and leaders like Demis Hassabis signal openness to coordinated pauses.
Industry outlooks emphasize data quality over brute-force scaling, predict small models will dominate agent use, and forecast AI to lead cloud spend by 2026.
Creative ecosystems tighten the tech–art bond by 2026 as open-source communities mature with incentives and credits, broadening access and experimentation in media production.

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