05-19-Daily AI News Daily

Today’s Summary

Google I/O kicks off tomorrow, Cursor drops its homegrown Composer 2.5 model today—the AI coding arms race just officially leveled up.
Dozens of AI tools' system prompts got dumped on GitHub in one go, racking up 130k stars in a day—all those secrets hiding behind the products are out in the open now.
Focus on the first two stories tonight, save time tomorrow for the Google I/O livestream.

⚡ Quick Navigation

• 📰 Today’s AI News - Latest updates at a glance

💡 Tip: Want to try the latest AI models mentioned here (Claude 4.5, GPT, Gemini 3 Pro) right now? No account? Grab one at Aivora —one-minute setup, hassle-free support.

Today’s AI News

👀 One-Liner

Google I/O launches tomorrow, Cursor quietly ships its homegrown model today—the AI coding tool arms race enters a new phase tonight.

🔑 3 Keywords

#GoogleIOEve #CodingModelHomegrown #SystemPromptsExposed

🔥 Top 10 Highlights

1. Google I/O Kicks Off Tomorrow—Full Product Line Getting AI Overhaul

Countdown’s over. Google just hyped it officially tonight: tomorrow’s I/O will cover Search, Gemini, Google AI Studio, and DeepMind across the board. This isn’t your typical annual keynote—over the past year, the Gemini lineup has already matched or beaten GPT on multiple benchmarks. Tomorrow’s announcements could straight-up reshape the AI product landscape. Search’s AI makeover, new Gemini versions, and the rumored Project Astra progress could all drop tomorrow. If you live in Google’s ecosystem, this livestream is worth blocking out time for. Set your alarm tonight.

2. Cursor Launches Homegrown Composer 2.5 Model—Built on Kimi, Teaming Up with SpaceX for Something Even Bigger

Two months back, developers dug up kimi-k2p5-rl from API headers and forced Cursor to admit Kimi was the base—this time they just put it straight in the blog. Composer 2.5 is all about not dropping the ball on long tasks: multi-step coding jobs spanning dozens or hundreds of steps without losing the plot. They’re claiming up to 10x efficiency gains over comparable models, and to push the new release, they’re doubling default credits for the next week. The bigger news: Cursor’s partnering with SpaceX AI to train a brand-new model from scratch—10x the compute, running on Colossus 2’s million-H100-equivalent cluster. The AI coding tool arms race just cranked up another notch.

3. Dozens of AI Coding Tools’ System Prompts Got Dumped—GitHub Repo Hits 130k+ Stars in One Day

This repo exploded today. Cursor, Windsurf, Devin, Replit, Lovable, Manus, Kiro, v0—basically every AI coding tool you can name. System prompts, internal tool calls, underlying model info, all catalogued. 130k+ stars in a single day, straight to the top of GitHub Trending. For regular users, this is a rare “flip the table” moment—what’s actually running under the hood of these products, how they’re instructing the AI, it’s all out in the open now. For these companies, this is probably the repo they least wanted to see this year.

4. Nvidia Pretrained a 12B Model in 4-bit Precision—Chinese Tech Scene Barely Noticed

For years, pretraining’s been locked to 16-bit and 8-bit. You can quantize at inference to save VRAM, but actually pretraining in 4-bit? Gradients blow up, loss crashes—industry consensus was it’s basically impossible. Nvidia just broke that with NVFP4, a new format: instead of brutally chopping everything to 4-bit, it splits numbers into small blocks, each with its own scale. Result: 2-3x speedup, 50% less memory, and almost zero intelligence loss. This isn’t a minor tweak—it’s a paradigm shift in pretraining. Training bigger models without needing a ton of VRAM might actually be possible now.

5. Yunnan Kid Spent $500, Used Homegrown AI Tools to Make a Short Film—Hollywood Sent Out a Manhunt

A three-minute AI animation short, millions of views in days, major Hollywood producers spotted it and literally posted international manhunts. The creator? A young guy from Yunnan who spent 10 days and $500, entirely with domestic AI tools. The shock isn’t that “AI can make shorts”—everyone knows that. It’s this: one person, one laptop, $500, and you can make something that gets Hollywood knocking. That used to need a team, funding, connections. The bar just moved.

6. The Biggest Opportunity in the Next 5 Years: Rebuild Every Industry with AI

The attention economy is over. Core thesis: once AI agents show up, business logic flips from “get seen” to “get it done”—users don’t jump between a dozen apps comparing options anymore, they just state an intent. Shengscape Network’s analysis: AI agents will drive 99% of global GDP within 20 years. Sounds wild, but the logic’s worth taking seriously: when AI handles most decisions and execution, does traffic and attention still matter? For founders and workers, this piece offers a framework worth thinking through.

7. Lu Qi—The Robotics Scene’s Quietest “Angel Investor”

While Wang Xing got crowned “embodied AI investment king,” few noticed Lu Qi was already quietly stacking chips in robotics. Per IT Orange data, in the first three quarters of 2025, Qiji Ventures led domestic robot company investments—beating Hillhouse, IDG, Ant, and Meituan. Zhiyuan Robotics, multiple early embodied AI startups—Qiji’s name’s on the cap table. From YC China to large models to robots, Lu Qi’s always half a step ahead. This piece is a solid lens: watch where Lu Qi’s betting, and you’ll see where robot money flows next.

8. NOVA Framework: Can AI Self-Iteration Actually Discover Real New Knowledge? Is There a Ceiling?

“Generate, verify, accumulate, retrain”—that’s the standard loop for AI self-improvement now. NOVA models this as an adaptive sampling process, tackling a fundamental question: can AI actually discover genuinely new knowledge through iterative self-improvement? What’s the cost? The research flags several failure modes, including knowledge contamination and diminishing returns. This isn’t an engineering paper—it’s drawing boundaries around AI scaling’s ceiling. For anyone tracking the AGI path, this framework offers a rare theoretical angle.

9. High-Res Image Generation on Mobile? ElasticDiT Squeezes DiT Architecture into Phones

The DiT architecture behind Stable Diffusion 3 and FLUX.1 is the gold standard for high-fidelity image generation, but deploying it on phones was basically impossible—compute and memory overhead were brutal. ElasticDiT uses elastic architecture plus sparse attention to squeeze it into mobile without obvious quality loss. If this pans out, local high-quality image generation on your phone without cloud API calls becomes real. Privacy, speed, cost—three problems solved at once.

10. Diffusion Model RL Fine-Tuning Doesn’t Need Every Step Optimized—Doing Less Actually Works Better

Standard approach for preference-aligning diffusion models with RL: apply RL across the entire denoising trajectory—heavy compute, mediocre results. This paper finds RL impact varies wildly across denoising stages: early stages barely matter, later stages are key. Skip early-stage optimization, save compute, and preference alignment actually improves. “Less is more” keeps popping up in AI training—that signal alone is worth remembering.

📌 Worth Watching

[Product] GPT-Image-2 Disco Ball Logo Prompt Goes Viral — Spotify dropped a disco ball logo, the internet’s now reverse-engineering the prompt in GPT-Image-2 to replicate the style. You can use it today.

[Research] DiffVAS: Guiding Drones in Partially Observable Environments with Diffusion Models — Applying diffusion models to drone search tasks, use cases include poaching hotspot detection and search-and-rescue. New intersection of AI and remote sensing.

[Research] LoH: Neural-Symbolic Fusion Framework Unifies Rule Learning and Neural Networks — Neural nets learn from data, symbolic systems reason. LoH tries unifying both in one language—rare systematic work in the NeSy direction.

😄 AI Fun

Making Abstractions Is Too Fun @CuiMao I Built a Fighting Game “Cui Cat vs. Adidas King”—Help Cui Cat Beat Adidas King to Unlock Her Claude Account…

The fun part here: AI didn’t show up on a conference stage preaching grand narratives. It snuck into a tiny action: fewer clicks, less waiting, less repetition. As tools get smarter, they’re like that coworker who always lends a hand—not doing earth-shattering work, but when you turn around, the small stuff’s already handled.

🔮 AI Trend Predictions

Google’s Multimodal Capabilities Get Full Upgrade, New Gemini Version Officially Launches

• Prediction Time: May 20, 2026 (tomorrow)
• Confidence: 92%
• Reasoning: Today’s news Google I/O Hype + Google officially confirmed coverage of Search, Gemini, DeepMind across the board. Every I/O has flagship model updates, and competitive pressure is higher this year—release intensity will only go up.

AI Coding Tools Enter “Homegrown Model” Era, Third-Party Model Dependency Drops

• Prediction Time: Q3 2026
• Confidence: 72%
• Reasoning: Today’s news Cursor Composer 2.5 + Cursor shifted from OpenAI/Anthropic dependency to Kimi-based fine-tuning and announced SpaceX partnership for training from scratch. Signal: top coding tools are internalizing model capability as core competitive advantage, not outsourcing. Windsurf, Replit following suit probability rises significantly next quarter.

Local AI Image Generation on Mobile Reaches Usable Stage

• Prediction Time: Q3-Q4 2026
• Confidence: 65%
• Reasoning: Today’s paper ElasticDiT proves DiT on mobile is viable + Nvidia’s 4-bit pretraining breakthrough ( today’s news ) further cuts model runtime costs. Both directions advancing simultaneously—gap between “lab-viable” and “product-ready” for mobile image generation is closing fast.

AI System Prompt Transparency Pressure Forces Industry Standardization

• Prediction Time: Q2-Q3 2026
• Confidence: 58%
• Reasoning: Today’s system prompt leak repo hit 130k stars, massive attention + multiple similar incidents past six months. As users and developers systematically dig and share this info, transparency pressure on AI companies keeps rising. Some may proactively publish system prompt frameworks to grab trust advantage.

❓ Related Questions

How Do I Try the Latest Google Gemini Version?

Google Gemini is accessible at gemini.google.com , but mainland users face network restrictions, and advanced features (like Gemini Advanced) need Google One subscription with overseas payment. After tomorrow’s I/O, new features may take time to roll out globally and typically hit paid users first.

Solution: Visit Aivora for ready-to-use accounts—instant delivery, hassle-free support, skip registration and payment headaches. You’ll be hands-on with new features the moment the keynote wraps.