📰 AI News Daily — 26 Sept 2025

TL;DR (Top 5 Highlights)

OpenAI expands its CoreWeave deal by $6.5B (total $22.4B), underscoring the compute arms race for next‑gen models.
NVIDIA and OpenAI plan up to $100B for supercomputing and 10 GW of data centers, reshaping AI infrastructure economics.
Google DeepMind debuts Gemini Robotics 1.5/ER 1.5, pushing embodied reasoning, web-browsing skills, and safer real-world robot tasks.
OpenAI rolls out ChatGPT Pulse and readies an in‑house ads platform, signaling a shift to proactive and monetized assistants.
Microsoft launches a unified AI marketplace and embeds Anthropic Claude into Copilot, escalating the enterprise AI platform battle.

🛠️ New Tools

GitHub Copilot CLI (Public Preview) — Brings AI assistance to the terminal for coding, testing, and deployment. Speeds common DevOps loops and lowers friction for developers living in shells.
Perplexity Search API — Real-time, millisecond-level answers for agents and LLM apps. Enables responsive retrieval-augmented experiences without running your own web-scale search stack.
Google Data Commons MCP Server — Lets AI query verified public datasets via Model Context Protocol. Reduces hallucinations and strengthens trust for high-stakes sectors like healthcare and finance.
Adobe Photoshop + Google AI & FLUX.1 Kontext Pro — First third‑party AI models inside Photoshop’s beta canvas. Accelerates ideation-to-edit workflows and expands choice for creative pros.
Suno Studio — A generative audio workstation for end-to-end AI music creation. Gives musicians and creators a streamlined pipeline from prompts to polished tracks.
NVIDIA Cosmos Reason — Agent and robot reasoning library surpasses 1M downloads. Signals growing demand for standardized planning and control tools in embodied and agentic systems.

🤖 LLM Updates

OpenAI GDPval benchmark — Evaluates economic value across 44 occupations; Anthropic Claude 4.1 Opus leads and beats domain experts in targeted tests. Highlights near‑expert capabilities on practical, revenue-relevant tasks.
Google Gemini 2.5 Flash/Flash‑Lite & Robotics‑ER 1.5 — Faster, more token-efficient models and enhanced spatial reasoning for robots. Enables safer, multi-step tasks and web-informed actions in real environments.
Meta Code World Model (32B) & Google EmbeddingGemma (300M) — Advances deeper code reasoning via agentic simulation and state-of-the-art tiny-scale embeddings. Improves developer tooling with lower costs and smaller footprints.
Qwen3‑VL and MamayLM v1.0 — Easier access to strong VLMs via Hugging Face and multilingual long-context reasoning (Ukrainian/English). Broadens vision and language support for global apps.
AMD MI300X fine‑tuning milestone — Full fine‑tune of a 4.5B medical model on AMD hardware. Validates credible non‑NVIDIA training pathways for high-end workloads and diversifies the AI compute ecosystem.

📑 Research & Papers

MIT’s SCIGEN — A generative AI system simulating millions of atomic structures to accelerate quantum materials discovery. Shortens the path from theory to synthesis for computing and energy breakthroughs.
NVIDIA CUDA‑Q + DGX Quantum — Hybrid quantum-classical stack advances practical quantum workflows for AI. Positions quantum as a nearer-term accelerator for optimization and simulation tasks.
Adaptive attention & continuous reasoning — ByteDance CASTLE improves attention flexibility; “soft token” RL shows richer continuous reasoning. Together hint at more robust, scalable reasoning beyond discrete tokens.
OpenAI pretraining breakthrough (report) — Rare large-scale pretraining gains suggest further headroom in scaling laws. Reinforces the value of data quality and training diversity for frontier models.
AI for urban heat resilience — ASU researchers map and predict shade to design cooler walking/biking routes. Offers actionable planning tools as cities face intensifying heat.

🏢 Industry & Policy

OpenAI × CoreWeave ($22.4B total) — New $6.5B expansion secures massive compute to fuel next-gen models. Reflects surging infrastructure demand as OpenAI targets major revenue growth.
NVIDIA × OpenAI supercomputing & Stargate — Up to $100B for AI supercomputers and plans for 10 GW of U.S. data centers with partners. Could generate hundreds of billions in ecosystem revenue.
Microsoft unified AI marketplace & Claude in Copilot — Streamlines enterprise AI procurement, improves security, and waives commissions; Anthropic Claude integration boosts 365 productivity features, intensifying platform competition.
OpenAI builds in‑house ads platform — Hiring for ChatGPT advertising infrastructure and brand integrations. Signals a durable monetization track as usage scales beyond subscriptions.
UN Security Council spotlight on AI governance — Yoshua Bengio urges evidence-based guardrails; Stanford HAI calls for equitable access. Elevates global norms to ensure benefits reach billions, not elites.
Agent security wake‑up call — Salesforce Agentforce ForcedLeak and malicious MCP/npm exfiltration incidents show third-party tool risks. Organizations are urged to harden agent integrations and update defenses.

📚 Tutorials & Guides

High-performance Triton kernels — Deep dive into writing fast GPU kernels and optimizing memory access. Practical techniques to squeeze more throughput from modern accelerators.
Designing fast softmax attention — Explains algorithmic choices and kernel-level tradeoffs. Helps practitioners implement scalable attention in custom architectures.
Perplexity’s evaluation stack — Transparent methodology behind its new Search API. Offers reproducible metrics for latency, relevance, and robustness in real-time retrieval.
From photo to animation with Wan 2.2 — Step-by-step guide to animate characters from a single image. Lowers the barrier to professional motion design for creators.
Booking.com’s AI Trip Planner — How a large consumer app layered OpenAI on existing systems for conversational trip design. Practical blueprint for enterprise-grade AI productization.

🎬 Showcases & Demos

Humanoid Pepper runs local LLMs/VLMs via Ollama — On-device reasoning brings privacy and responsiveness to robotics. Demonstrates accessible embodied AI with commodity hardware.
Generalist Lego builder — Assembles structures from visual inputs alone. Highlights emerging visuomotor competence without exhaustive task-specific programming.
Pollen Robotics’ Reachy Mini (live) — Compact, capable robot showcased with real-time control. Encourages hobbyist and research communities exploring affordable manipulation.
Kling 2.5 frame chaining — Effectively infinite video generation, often paired with AI music. Opens new formats for storytelling, loops, and ambient media.
TinyWorlds — A compact model generates playable game worlds. Points to AI-native game prototyping and rapid content iteration.
Chat with podcast guest personas — Post-episode interactions with digital avatars extend engagement windows. Blends media, memory, and personalization for new audience experiences.

💡 Discussions & Ideas

Evaluation fairness after Anthropic bug — A routing flaw skewed results, sparking demands for transparent benchmarks and credit. Reinforces reproducibility and shared evaluation standards.
Prompt optimization > fine‑tuning (often) — Evidence like Databricks’ GEPA shows prompt engineering can rival supervised fine‑tuning at lower cost. Practical takeaway: optimize prompts before training.
Automation reality check — Bold forecasts (50% of entry-level white-collar roles) face headwinds; domains like radiology remain resilient despite benchmark wins. Adoption hinges on trust, liability, and integration.
Are video models nearing a “GPT moment”? — Emergent zero-shot skills (e.g., Veo3) fuel speculation, but reliability gaps persist. Real value will depend on consistency and controllability.
Toward automated researchers — OpenAI and others envision agentic systems discovering insights independently; renewed interest in open-ended algorithms could catalyze scientific breakthroughs.
Developer trust gap — With AI coding tools approaching ubiquity, only a quarter of developers fully trust outputs. Transparency, debuggability, and safety guardrails remain key adoption levers.

Source Credits

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