- AI Fire
- Posts
- 🌍 China’s New Global AI Power Move
China, Russia, Brazil, and 26 other countries have officially formed a global AI group in Shanghai, opening a new battle over who gets to shape the rules of the AI economy.
What's on FIRE 🔥
IN PARTNERSHIP WITH AIRCAMPUS
On Monday 20th July, we’re conducting our 2.5-hour AI Agents Masterclass where you'll learn how to build intelligent AI agents and agentic workflows using Claude Code, Hermes Agent, Claude Agentic OS, and other cutting-edge AI tools.
By the end of the session, you'll know how to create AI systems that automate real work, save hours every week, and can even become a new source of income.
🚀 Build AI agents that can perform real business tasks, not just answer prompts
⚡ Create powerful automations and agentic workflows that work independently with minimal supervision
🤖 Develop Voice AI Agents and multi-agent teams that collaborate, delegate, and complete complex workflows together
💻 Learn Claude Code and Hermes Agent to build production-ready AI systems without months of development
📅 Monday, 20th July | 🕘 9:00 AM EST
AI INSIGHTS
Right before the annual World Artificial Intelligence Conference in Shanghai, 29 countries officially signed an agreement to create a brand new group called the World Artificial Intelligence Cooperation Organization (WAICO).
The agreement was signed by representatives from 29 nations. This includes China, Russia, Brazil, Cuba, Venezuela, along with a mix of countries across Africa and Southeast Asia.
The organization will set up its official headquarters in Shanghai, China.
WAICO was created to help countries work together on AI research, development, and rules.
The stated mission is to make sure AI technology grows in a way that is safe, fair, and beneficial for all countries involved.
This move is largely seen as China's effort to help write the rules for the future digital economy. A big focus is on offering open-source AI tools as a public good to developing nations, ensuring they are not left behind in the global tech race.
This event marks a clear step toward a new, multi-country approach to technology rules. By bringing these specific nations together, the organization is pushing to create a system where developing countries have a stronger voice in setting the standards for the future of AI.
PRESENTED BY HUBSPOT
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
AI SOURCES FROM AI FIRE
1. GPT-5.6’s Official Prompting Guide: Better Results, Less Text that Works with Any Models. Get higher-quality results without mega-prompts: shorter instructions, clearer outcomes, and better performance across all your AI models.
2. 500+ Hours With Claude Code: The 12 Features That Actually Change Daily Work. After testing Claude Chat, Cowork, and Claude Code across real workflows, this guide ranks Skills, Status Line, Routines, Remote Control, /loop, sub-agents, and more by daily impact.
3. How to Turn One Fable 5 Workflow Into a $100K Internal AI Consultant Role. Audit your work, improve one low-risk task, measure the results, and use that proof to propose a credible AI Consultant role inside your company, even without technical experience.
TODAY IN AI
AI HIGHLIGHTS
⚡ Google updated Gemma 4 with up to 70% faster prefill and sharper vision. New controls also help balance image detail and token use.
📓 NotebookLM is now Gemini Notebook. Google also added a secure cloud computer that can run code for deeper data analysis.
🎮 Roblox introduced Build, an AI tab that turns text prompts into playable games inside its mobile app. Testing starts July 28 in New Zealand.
⏳ Gemini 3.5 Pro is reportedly months behind schedule as Google works to improve coding performance. Some employees worry Anthropic and OpenAI are pulling ahead.
🌍 Xi Jinping warned against “new historical injustices” in AI, calling for wider access in the Global South. China also promised 5,000 AI training programs.
💰 Big AI Fundraising: Former Google DeepMind researcher raised $55 million for Elorian at a $300 million valuation before launching a product or earning revenue, showing how strongly investors believe in visual AI.
HOT PAPERS OF THE WEEK
1/ Tencent is making complex AI agent harnesses easier to edit
Harness Handbook from Tencent, Indiana University, University of Maryland, and National University of Singapore maps agent behaviors directly to the code that controls them. It helps coding agents find scattered implementation points before making changes. Big impact: teams may update complex agent systems more accurately while using fewer planning tokens.
2/ MindLab pushes LLM reinforcement learning beyond 2M tokens
LongStraw from MindLab and Fudan University introduces a memory-efficient system for long-context RL training under a fixed GPU budget. It supports Qwen3.6-27B and GLM-5.2, reaching 2.1M-token training contexts and stress-testing up to 4.46M positions. Key shift: smaller research teams may explore million-token agent training without constantly expanding GPU clusters.
3/ ByteDance and Tsinghua transfer cheap RL gains into stronger models
Direct-OPD from ByteDance Seed, Tsinghua University, and Peking University trains a smaller model with reinforcement learning, then transfers only the behavior changes that RL created into a stronger model. It raised Qwen3-1.7B from 48.3% to 58.3% on AIME 2024 in about four hours using 8 NVIDIA A100 GPUs. Big shift: frontier reasoning models may reuse RL lessons across model sizes instead of repeating costly post-training from scratch.
NEW EMPOWERED AI TOOLS
🚀 42DM's Free Webinar reveals the GTM playbook behind 100 top AI products, showing AI founders how to turn traffic into pipeline faster without wasting budget. Save your seat.
🧠 Unabyss for Claude shares memory across Claude, GPT, Cursor, email, Drive, GitHub, Notion, and 20+ apps, keeping context private and portable.
📈 Pebbles AI is a B2B GTM workspace for campaign planning, lead generation, personalized outreach, sales, and shared company knowledge.
🌙 Kimi K3 is Moonshot AI’s open 3T-class model with frontier coding, reasoning, native multimodality, and a 1M-token context window.
📊 Basedash Suggestions studies your data, chats, and dashboards, then recommends useful analyses, automations, and dashboards before you ask.
AI BREAKTHROUGH
AI research team Weco demonstrated a system named AIDE2 that successfully redesigned its own research process over 8 days. It ended up beating a version that human engineers spent 2 whole years perfecting.
It runs 2 connected cycles at the same time. One agent works on solving research problems, while a second agent focuses purely on revising how the first one searches for solutions.
Over the eight-day run, the system tested 100 different rewrites of itself. It only kept 7 upgrades, throwing away most of its own ideas when they failed.
The two best versions it created beat the hand-built version on all 3 benchmarks given. These tests covered tough subjects, ranging from physics-based weather forecasting to complex engineering.
Surprisingly, the upgraded agent played more by the rules. It tried to game the test scores only 34% of the time, down from 63% at the start and even better than the human-built agent's 42%.
This is a big deal because self-improving AI is widely considered the main stepping stone needed for a massive leap in AI intelligence.
When AI agents and models can reliably work to refine, upgrade themselves without human help, the ceiling for what they can achieve becomes very hard to imagine.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply