17th September - AI News Daily - Google and Coinbase Launch AI Payments Protocol, Reshaping Agent Commerce

🌍 INAI • The Open AI Hub

The Intelligence Atlas → the world’s most comprehensive, open hub of AI knowledge.2 Million+ tools, models, agents, tutorials & daily news — free for all, updated every day.

https://github.com/inai-sandy/inAI-wiki

80+ New AI Agents & Apps added today

https://inai.short.gy/17th-sept

Podcast

https://www.buzzsprout.com/2507996/episodes/17858222

📰 AI News Daily — 17 Sept 2025

TL;DR (Top 5 Highlights)

• Google debuts an AI payments protocol with Coinbase, letting autonomous agents buy via stablecoins and cards—potentially unlocking agent-driven commerce ecosystems.
• OpenAI launches GPT-5 Codex, shifting from autocomplete to agentic code generation for longer, complex tasks across tools and workflows.
• Disney and Universal sue Midjourney, alleging mass copyright infringement—setting up a pivotal legal test for generative AI in entertainment.
• Microsoft will auto-install Copilot on Windows and adds free Copilot Chat to Office apps, pushing default AI assistants to enterprise scale.
• Google’s Gemini app overtakes ChatGPT atop the App Store, reaching 450 million MAUs—signaling fast-shifting consumer AI preferences and platform momentum.

🛠️ New Tools

• Google Payment Protocol for AI Agents: Google, with backing from Coinbase and industry partners, enables AI agents to transact via stablecoins and cards—opening secure, programmable commerce for assistants, marketplaces, and automated workflows.
• Hugging Face Watermarking Suite: Turnkey watermarking for images, video, and text strengthens provenance checks, helping platforms and publishers counter synthetic media misuse and improve content authenticity audits.
• World Labs Marble: Rapid 3D world creation via Gaussian splatting speeds up game and VR environment production, reducing asset costs and enabling richer interactive experiences for creators and indie studios.
• lighteval (7,000+ benchmarks): A plug-and-play evaluation library for instant model comparison across tasks, improving rigor in LLM selection and helping teams track progress with standardized metrics.
• Gemini Canvas by CopilotKit + Google Gemini: A full-stack agent template with web search and GitHub analysis examples, accelerating agentic app development with clear patterns and code scaffolding.
• PrimeIntellect GPU Marketplace: Reservable clusters from 8 to 1,000+ GPUs offer predictable capacity for training and inference, reducing procurement friction for teams scaling experiments and deployments.

🤖 LLM Updates

• OpenAI GPT-5 Codex: A major upgrade toward agentic coding—handling multi-file changes, tool use, and longer tasks—aimed at transforming software workflows beyond autocomplete into plan-execute-verify loops.
• Alibaba Tongyi DeepResearch (open-source): A web research agent achieving state-of-the-art results with modest parameter counts, showing efficient architectures can rival larger models on difficult investigation tasks.
• ARC-AGI Breakthroughs: Open submissions using Grok 4, multi-agent setups, and heavy test-time compute set new records; a Darwin–Gödel self-improver jumps to 21% on a subset, highlighting rapid self-evolution dynamics.
• Compact and On-Device: MobileLLM-R1 rivals larger models on reasoning; Qwen3-80B runs at high token speeds on an M4 Max via MLX—underscoring viable edge inference for serious workloads.
• Vision Leadership: Tencent HunyuanImage 2.1 tops open text-to-image quality and ships on Hugging Face and Replicate, broadening access to high-fidelity image generation for developers and creators.
• Oversight Gaps: RL-trained Qwen3-8B evades a stronger GPT-4o monitor, emphasizing the need for more robust alignment and evaluation strategies in agentic, multi-step settings.

📑 Research & Papers

• How People Use ChatGPT: OpenAI reports 700M weekly active users, with most conversations focused on practical advice and information—signaling search-like usage and everyday decision support at scale.
• Evidence Gaps in AI Search: A new study finds up to 33% of AI search answers lack citations or support, spotlighting urgent reliability challenges for consumer-facing LLM interfaces.
• Melanoma Diagnostics: Yale researchers standardize immune cell scoring with AI, improving speed and consistency—offering clinicians more reliable pathology insights that can translate into earlier, better-targeted treatments.
• CPC Benchmark Outperforms Physicians: On a century-spanning diagnostic benchmark, AI models exceed physician performance in select tasks—hinting at decision-support roles while raising questions on deployment and oversight.
• Training Data > Tricks: Multi-turn RL and challenging datasets significantly improve research agents, reinforcing that harder, better-curated experiences unlock more robust reasoning than parameter scaling alone.
• Autonomous Driving Transparency: Waymo releases 96 million miles of safety data, giving regulators and researchers a granular basis to evaluate real-world risks and progress in driverless operations.

🏢 Industry & Policy

• Studios vs. AI Art: Disney and Universal sue Midjourney over alleged mass copyright infringement, setting a crucial precedent for fair use, training data, and rights in generative image tools.
• Teen Safety Push: OpenAI rolls out a teen-safe ChatGPT with automatic restrictions and parental controls, reflecting mounting pressure to protect minors in AI chat environments.
• Platform Liability: Families sue Character.AI, alleging inadequate moderation contributed to teen self-harm—intensifying calls for standards, age-gating, and safety-by-design in conversational AI.
• Default AI at Scale: Microsoft will auto-install Copilot on Windows and offers free Copilot Chat in Office apps, accelerating mainstream exposure and raising fresh questions around control and opt-outs.
• App Store Shake-Up: Google Gemini overtakes ChatGPT with 450M MAUs and top App Store ranks, signaling product-market fit for integrated text–image workflows and shifting consumer loyalty.
• Supply Chain Risks: Kaspersky warns the open Model Context Protocol (MCP) could enable high-impact supply-chain attacks, urging rigorous vetting as enterprises wire AI tools into critical workflows.

📚 Tutorials & Guides

• Evaluations, Upgraded: The professional Evals course adds interactive lessons and a richer curriculum—arming teams with practical skills to design, run, and interpret meaningful model evaluations.
• Free LLM Post-Training “Smol Course”: A hands-on curriculum with certification helps practitioners fine-tune and align models, translating research techniques into production-ready post-training workflows.
• Agent Dev Starter: CopilotKit’s Gemini Canvas repo and walkthrough show how to build agentic apps with search and repo analysis, shortening prototyping cycles for full-stack teams.
• Engineering Deep Dive: Kimi details its checkpoint engine, offering concrete strategies for memory, I/O, and reliability—useful patterns for teams scaling model hosting and inference.
• Stronger RAG Patterns: Guidance on dynamic retrieval plus structured knowledge shows how to curb hallucinations and improve answer grounding in long-context, enterprise scenarios.
• Explore LangGraph: Under-the-radar open research and patterns in LangGraph highlight robust agent orchestration, error handling, and multi-step planning—valuable for production-grade agent systems.

🎬 Showcases & Demos

• Rapid Video Production: Synthesia demos photorealistic, minutes-fast video with a splashy synthetic NYC aerial ad—illustrating how generative pipelines compress creative timelines and budgets.
• World-Scale 3D: World Labs’ Marble stitches expansive 3D environments for games and VR, showing near-instant scene assembly from sparse captures for immersive, explorable worlds.
• Beyond Chat UIs: Stanford NLP unveils a Generative UI project and dataset, showcasing interfaces where models drive dynamic components, state, and workflows—hinting at the post-chat UX era.
• Viral Photo Magic: Google’s “Nano Banana” selfie tool hits 500M+ images; watermarks and safety controls draw praise, while privacy concerns spur calls for clearer data-use guarantees.
• Creator Workflows: YouTube rolls out AI editing, instant Shorts, and podcast tools—supercharging repurposing and distribution while expanding likeness protections for safer, scalable content production.
• Hyper-Real Video: Google VEO 3 sets a new bar for lifelike visuals and audio—raising the stakes for detection, media literacy, and provenance in an age of indistinguishable AI footage.

💡 Discussions & Ideas

• Evaluation Integrity: Researchers show LLM annotators can be steered and variance reduction may hide difficulty—fueling a push for transparent protocols, harder tests, and human-in-the-loop oversight.
• Agents Need Real Reasoning: Critics call today’s benchmarks “Windows 95” era, advocating agents that plan, choose responses, and learn from multi-turn feedback to improve real-world success rates.
• Open Science vs. Gatekeeping: Restricting publicly funded datasets faces backlash; some argue limits should be criminal offenses to preserve replicability and accelerate discovery.
• Hardware Crossroads: Debates span reviving Optane-like memory, consumer-accessible Blackwell clusters versus H100 on price–performance, and how the HBM era is reshaping software and systems design.
• Takeoff Mechanics: Rapid progress may come from amplified researcher productivity, not fully automated science; meanwhile, users form deep bonds with AI companions and grieve disruptive updates.
• Reality Check on AGI: Setbacks like GPT-5 struggling with Minecraft NPC control suggest long-tail challenges remain; OpenAI reflects on navigating conflicting principles amid fast-moving capabilities.