📰 AI News Daily — 25 Sept 2025

TL;DR (Top 5 Highlights)

Nvidia, OpenAI, Oracle, and SoftBank unveil a $100B+ AI infrastructure push (Stargate), targeting 4–5M GPUs and up to 10–17 GW power—raising sustainability and antitrust debates.
Microsoft 365 Copilot adds Anthropic’s Claude Sonnet 4 and Opus 4.1, giving enterprises multi-model choice alongside OpenAI for resilience and better task fit.
Google rolls out Search Live for real-time, camera-based queries and a Data Commons MCP server to ground AI in trusted data and curb hallucinations.
Apple’s SimpleFold brings laptop-run protein folding via open MLX models, lowering cost and complexity for bio-AI experimentation and education.
Early AI agents show strong ROI, but surveys flag governance gaps; new payment protocols aim to make agent-to-agent transactions secure and seamless.

Google Data Commons MCP Server delivers standardized access to trusted public datasets, grounding AI outputs in real-world facts. It aims to reduce hallucinations and improve reliability across health, economics, and climate apps.
Apple SimpleFold debuts a Transformer-based protein folding system with open MLX models that run on laptops. It lowers barriers to bio-AI prototyping for researchers, students, and indie labs.
Databricks Agent Bricks packages training- and prompt-time optimizations (GEPA, SFT stacking) to lift open-source model performance. Teams can reach stronger baselines without complex bespoke pipelines.
Amazon Redshift MCP Server enables natural-language SQL analysis through the Model Context Protocol. It eases secure query execution, metadata exploration, and cluster discovery for analysts and data teams.
Google Search Live launches in the U.S., letting users point a phone camera and ask questions with instant, contextual results. It brings intuitive, real-time AI search directly to everyday mobile use.
DeepEval arrives as a “Pytest for LLM apps,” offering lightweight, framework-agnostic test suites. It standardizes evaluation to catch regressions earlier and increase confidence in production agents.

Meta Code World Model (32B) releases as open weights for code generation, agentic reasoning, and planning. Early results show strong benchmark performance and transparency for research use.
Qwen3-VL (Alibaba) sets a new open-source bar in vision-language modeling using DeepStack techniques, improving grounding and reasoning. It strengthens the open ecosystem’s multimodal options.
Baichuan-M2 INT4 achieves notable efficiency via AutoRound quantization, enabling faster, cheaper inference with minimal quality loss—useful for edge deployments and cost-sensitive workloads.
Core Space Merging introduces near-lossless merging of LoRA adapters without restoring full weights. It simplifies multi-domain fine-tuning workflows and cuts compute requirements for model composition.
GPT-5 Codex in GitHub Copilot brings smarter agentic coding, improved completions, and custom assistance to Pro and Enterprise users, lifting in-IDE productivity on complex codebases.
DeepSeek V3.1–Terminus improves tool use and reduces execution errors, signaling steady progress in reliable agent behavior for developer and operations workflows.

Google’s diffusion-based research-writing agent drafts with adaptive reasoning, assisting literature synthesis and structured outputs. It could accelerate systematic reviews and grant writing by reducing manual drafting time.
Automated prompt optimization research shows open-source models can match or surpass frontier systems on enterprise tasks at lower cost, highlighting the strategic value of prompt tooling over raw scale.
Mayo Clinic pediatric asthma risk model identifies high-risk children as young as three, enabling earlier intervention and potentially better outcomes in population health management.
AI for infectious disease education study finds ChatGPT most accurate and engaging among leading models, underscoring AI’s growing role in public health communication and learning.

Nvidia, OpenAI, Oracle, SoftBank mount an $100B+ infrastructure push, including Stargate and 5+ U.S. data centers. The plan targets 4–5M GPUs and 10–17 GW capacity, igniting sustainability and antitrust scrutiny.
Microsoft integrates Anthropic models into Microsoft 365 Copilot, adding Claude Sonnet 4 and Opus 4.1. Multi-model choice enhances reliability, coverage, and vendor flexibility across enterprise workflows.
SAP, OpenAI, Microsoft bring secure AI to Germany’s public sector via SAP Delos Cloud (launching 2026). The €20B initiative targets digital sovereignty, compliance, and modernization of government services.
Agent economy rails: Google’s Agent Payments Protocol and the x402 Foundation (Cloudflare/Coinbase) introduce standards for secure, transparent, and account-free machine payments—key infrastructure for autonomous services.
Early agent ROI: A Google Cloud survey reports 88% of early adopters see positive returns, especially in productivity and CX. Separate findings highlight testing and governance gaps that could threaten resilience.
OpenAI security patch fixes a critical Deep Research agent flaw that exposed Gmail/Outlook/Drive data via hidden prompt injection. It underscores urgent needs for red-teaming, least privilege, and agentic guardrails.

Generalist robot assembles Lego from pixels end-to-end, demonstrating precise manipulation without task-specific engineering. It’s a notable step toward versatile, real-world-capable embodied AI.
Google Gemini on TVs & Play Store Sidekick brings conversational search, live in-game help, and personalized navigation to entertainment. It showcases ambient, cross-platform AI experiences in consumer media.

UN Security Council brief urges preventing concentration of AI capabilities and expanding equitable access through international coordination—framing AI as a global public-interest infrastructure.
Beyond GPUs: Researchers argue algorithmic advances—including quantum-inspired techniques—may drive the next step change, counterbalancing market fixation on data centers and hardware.
Ambient coding agents are expected to follow developers across IDEs, browsers, mobile, and TVs—hinting at continuous, context-aware assistants integrated into daily workflows.
Workforce sentiment: UK employees expect to offload 41% of tasks to AI within three years, yet remain concerned about accuracy, accountability, and training—potential barriers to adoption.
AI in schools: Over 1,500 U.S. districts use Gaggle to monitor student devices. Critics cite false positives and privacy risks, fueling legal and ethical debates about safety tech.

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