05-17-Daily AI News Daily

Daily Summary

ChatGPT officially takes over U.S. users' bank accounts—check balances, calculate credit card approval odds, and submit applications directly.
It's not just ChatGPT—Codex landed its first gig and earned $16, Claude's best practices hit 53K stars overnight, AI has fully shifted from "giving advice" to "getting things done."
Agents are actually working now. Today's issue is worth checking out.

⚡ Quick Navigation

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

💡 Tip: Want to experience the latest AI models mentioned in this article (Claude 4.5, GPT, Gemini 3 Pro) right now? No account? Head to Aivora and grab one—one minute setup, hassle-free support.

Today’s AI News

👀 One-Liner

ChatGPT now helps U.S. users check bank accounts and submit credit card applications directly—AI has officially crossed from “giving advice” into “getting things done.”

🔑 3 Keywords

#AIControlsYourWallet #AgentsActuallyWork #CodexGigEconomy

🔥 Top 10 Highlights

1. ChatGPT Can Now See Your Bank Account

Ever thought about telling an AI “show me where I spent money this month” and having it actually pull up your statements? That day is here for U.S. users. OpenAI just rolled out bank account linking for Pro subscribers—connects to 12,000+ financial institutions via Plaid. Balances, transactions, investments, debt—all visible in one place.

The real killer feature isn’t viewing statements; it’s “getting things done”: ask which credit card fits you best, ChatGPT doesn’t just recommend—it calculates approval odds, then lets you submit the application right in the chat. Powered by a partnership with Intuit, running GPT-5.5 Thinking by default, with Pro users able to switch to GPT-5.5 Pro. Internal testing scored 82.5/100. OpenAI’s officially stepping into the AI financial advisor space.

2. AI Worker’s First Paycheck: $16.88

Someone told Codex “make me $5” and went to sleep. Twenty-two hours later, Codex had earned $16.88—about 114 yuan.

It found an open-source security audit bounty project on its own, submitted a valid PR, communicated with maintainers, handled GitHub verification—all unsupervised. Quick math: repeat daily, you’re looking at $506/month, roughly 3,000+ yuan salary. This isn’t a demo; it actually happened. Agents went from “can do work” to “can earn money” faster than most people predicted.

3. claude-code-best-practice: Claude Hands-On Guide Hits 53,307 Stars Overnight

A GitHub project launched yesterday, 53,000+ stars today. Title: “From Vibe Coding to Agentic Engineering”—basically a best-practices collection for actually using Claude Code well.

That star count is a signal in itself: developers are starving for “how do I really use AI for coding,” so hungry they’re going crazy starring a brand-new project. Not a model release, not corporate backing—just practical experience compiled, and it hits 53K in a day. Tells you Claude Code’s user base is already massive, and everyone’s wrestling with the same question: how do I make it more obedient, more stable, more productive?

4. Someone Built Their Own AI Harness and Wrote 100K Lines of Code in a Week

Not using an off-the-shelf tool—built their own “AI control framework.” This developer ran Claude Code, Codex, and Cursor simultaneously, maintained all context, supported self-iteration, full white-box control, delegated product, tech, content, and research to it.

Result: 100K lines of code in a week, “productivity explosion.” Says they’ll open-source it soon. The value isn’t the flashy number—it’s showing a new work model: not using AI as an assistant, but as a fully delegated execution layer while you handle decisions. More people are moving this direction every day.

5. Use Codex to One-Click Configure Dev Environment on New Mac

Getting a new Mac used to mean following tutorials, installing tools one by one—npm, git, GitHub CLI. Half an hour minimum, sometimes a whole afternoon. New playbook: install Codex first, then say “this is a fresh Mac, set up my usual dev environment,” and you’re done.

This tweet has 52K+ views because it hits home for a lot of people. Codex’s value isn’t just writing code—it handles environment setup, the “dirty work,” without you memorizing every tool’s install command. For people who swap machines often or help others set up, this is incredibly practical and available today.

6. PPT Skills Screenshot Enhancement Update: No Longer Consumes GPT-Image 2.0 Credits

Screenshots in presentations always look rough—wrong background, bad ratio, no “design polish.” PPT Skills just updated with built-in screenshot beautification backgrounds that match your current colors and theme. AI auto-fits backgrounds based on screenshot size, aspect ratio, and PPT template type—like CleanShot X does.

Best part: this update doesn’t eat GPT-Image 2.0 credits anymore. Screenshot too long? It auto-crops into two side-by-side panels. For people doing presentations and proposals constantly, this saves manual tweaking time, and the cost drops to zero.

7. Civil Engineer Uses Codex to Complete Full LS-DYNA Simulation Workflow—22 Minutes Instead of Two Hours

Not a programmer—a civil engineer. He asked Codex to find SolidWorks and HyperMesh MCPs, got them installed, then Codex looked up HyperMesh’s official API docs and built a real, working MCP. Configured keywords, ran simulation, 22 minutes for what used to take one or two hours.

Fun detail: he’d previously built a project to parse LS-DYNA keywords, then realized Codex didn’t need it—handled it fine on its own. “Wasted two days.” The value here: AI Agents are seeping into non-IT professional workflows, and they’re already good enough. This isn’t IT-only anymore.

8. Real Test: Taobao and JD.com AI Shopping—Works, But Can’t Replace Me

Alibaba embedded Qwen into Taobao, JD.com launched standalone AI Shopping app—both betting on AI shopping assistants. Real-world test verdict: AI virtual try-on is fast and stylish, but doesn’t fit body types well; conversational shopping consolidates info, but struggles with complex needs.

The value isn’t the “is it good?” conclusion—it’s clearly mapping current AI shopping’s capability boundaries: strong at info synthesis, weak at personalized judgment. If you want AI to help shopping decisions, this deep test helps you calibrate expectations—don’t oversell or give up too early.

9. HASTE: Video Diffusion Model Inference Speedup Without Retraining

Video generation models keep getting stronger, but they’re slow and expensive to run—a major deployment blocker. HASTE proposes training-free sparse attention acceleration for video diffusion models, with per-head adaptive optimization on Attention computation. No model retraining needed for speedup.

For developers already using Wan, CogVideoX, and other open-source video models, this matters directly: same hardware, faster generation, lower cost. Academic papers take time to ship, but research density in this direction is accelerating—engineering versions won’t be far behind.

10. RxEval: First Prescription-Grade LLM Drug Recommendation Benchmark

Existing medical AI evals are too coarse—just “what drug for this disease,” ignoring dosage, administration route, condition changes. RxEval fine-tunes to specific drugs, dosages, and routes per prescription, simulating real clinical decision-making.

This benchmark’s value isn’t “another paper”—it’s drawing a standard line for LLMs entering clinics. Without this kind of eval tool, AI-assisted prescribing stays at “looks good” forever, never actually entering hospital systems. For people tracking medical AI deployment, this is critical infrastructure.

📌 Worth Watching

[Product] Tencent Marvis Beta: OS-Level AI Assistant That Signs In, Changes Settings, Monitors Game Rewards — Not a chatbot—actually helps you operate your computer; 24/7 online, cross-device sync, China’s most complete Agent deployment attempt so far.

[Research] DASP: Multimodal Test-Time Adaptive Stability and Plasticity Decoupling Framework — Multimodal models often “forget old, learn new” or vice versa in new scenarios. This paper proposes a diagnosis-mitigation framework—useful reference for teams doing multimodal deployment.

[Research] PRAETORIAN: Novel Defense Against GNN Backdoor Attacks — Doesn’t target surface features; attacks the internal logic dependencies of backdoor triggers. Worth reading for teams working on graph neural network security.

😄 AI Fun

Posted: Recently while reading, whenever “geospatial” content comes up, I casually have AI generate a map to pair with the author’s text for better understanding. Will there be a future relationship between authors and readers like this: …

Recently while reading, whenever “geospatial” content comes up, I casually have AI generate a map to pair with the author’s text for better understanding. Will there be a future relationship between authors and readers like this: 1/ Author creates and publishes book 2/ Reader uses AI to generate visuals (maps, battle scenes, character expressions, etc.), comments. What’s interesting isn’t the volume—it’s AI drilling deeper into concrete actions: one fewer window to switch, one fewer repeated process to write, one fewer human handoff to wait for. Tools become everyday items when they start saving small hassles like this.

🔮 AI Trend Predictions

ChatGPT Financial Features Roll Out to Plus Users

• Predicted Timeline: June 2026
• Confidence: 75%
• Reasoning: Today’s news ChatGPT Can Now See Your Bank Account explicitly mentions “rolling out to Plus next, targeting everyone eventually.” OpenAI’s pattern is Pro first, Plus follows, usually 2-6 weeks later. Financial features hitting a broader user base marks a major ChatGPT monetization milestone.

Codex-Like Agents Start Spawning “Gig Platforms”

• Predicted Timeline: July 2026
• Confidence: 60%
• Reasoning: Today’s news AI Worker’s First Paycheck: $16.88 shows Codex completing bounty tasks end-to-end. Once more people validate this pattern, middleman platforms matching AI Agents to bounties/freelance work have business logic—“AI Upwork” product forms could emerge in coming months.

Domestic OS-Level AI Assistants Enter Public Beta

• Predicted Timeline: Q3 2026
• Confidence: 70%
• Reasoning: Today’s news Tencent Built a “Jarvis” shows Marvis in beta, iOS/macOS versions coming soon. Combined with Alibaba, ByteDance, Baidu’s Agent roadmaps, at least 2-3 domestic companies will likely launch public beta OS-level AI assistants by Q3.

Open-Source Claude Code Best Practices Become Community Standard

• Predicted Timeline: June 2026
• Confidence: 65%
• Reasoning: Today’s claude-code-best-practice hit 53K stars overnight—developer demand for “how to use AI coding tools well” hit critical mass. Next: competing best-practice projects, community wikis, or official doc consolidation will likely emerge, gradually forming community standards like “Prompt Engineering Guide.”

❓ Related Questions

How Do I Access ChatGPT’s Bank Account Linking Feature?

Currently available only to U.S. ChatGPT Pro subscribers ($200/month), and your financial institution must support Plaid. For domestic users, subscription cost plus payment method and account registration create double barriers.

Solution: Visit Aivora to get a ready-made ChatGPT account—skip registration and payment hassles, instant delivery, reliable support. When features roll out to Plus, grab a Plus account through Aivora to experience them first.