05-21-Daily AI News Daily
Daily Summary
Former Google CEO booed off stage at graduation for hyping AI as students hiss collectively; Gallup data shows American anger toward AI is replacing excitement.
Meanwhile, get-shit-done racked up 63k stars in a single day, and Codex wrote its own script to bypass input method obstacles—developers aren't waiting for evangelists, they're building solutions themselves.
Today's issue is worth opening: the public sentiment shift behind the boos, and the practical tools emerging in this explosion phase—two storylines you need to follow.⚡ 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? Grab one at Aivora —instant access, hassle-free support.
Today’s AI News
👀 One-Liner
A former CEO got booed at graduation for hyping AI; the students in the audience are sharper than the guy on stage.
🔑 3 Key Themes
#AI Anxiety Rising #Developer Self-Help #Tool Explosion Era
🔥 Top 10 Headlines
1. American Graduation Crowds Boo AI Evangelist; Former Google CEO Forced to Apologize Mid-Speech
Last weekend, Eric Schmidt stood on stage at the University of Arizona, comparing AI to a “rocket ship ticket” and urging graduates to “build AI agent teams”—the crowd’s hissing grew louder, forcing him to pause and say “I know you’re worried, and that fear is reasonable.” But then he pivoted back to the same pitch. The audience wasn’t buying it.
He wasn’t alone. University of Central Florida, Middle Tennessee State University—the same scene played out on different stages. A Gallup survey last month showed American excitement about AI is cooling while anger is rising. The people on stage pushing young people to embrace AI are the same ones driving this transformation and using AI to cut jobs—students see that contradiction crystal clear.
2. get-shit-done: GitHub’s Hottest Project Today, a Meta-Prompt System Built for Claude Code
A project with a refreshingly blunt name just exploded on GitHub—get-shit-done hit 63,335 stars in a single day, starting from zero.
This is a lightweight meta-prompt, context engineering, and spec-driven development system built by TÂCHES specifically for Claude Code. In plain terms: it systematizes how to talk to Claude so it actually gets things done. Not just throwing a few prompts at it, but a complete context management framework. Claude Code adoption is accelerating, and this project’s timing is perfect—before the toolchain matures, prompt engineering is the productivity lever.
3. Open-Source Plugin Unlocks Computer Use and Goal Commands in Codex App, Even Switches UI to Top Tabs
Logging into Codex via API normally limits your features—but someone built an open-source plugin that bypasses that restriction entirely. Install it and Computer Use works, Goal commands are available, and you can even switch the interface to Chrome-style top tabs, plus set sound alerts for task start and completion.
This “official won’t give it, so community builds it” pattern is becoming standard in AI tools. Codex is already powerful; this plugin adds a customization layer on top, and for power users it’s genuinely practical. Open-source link is in the original post’s replies.
4. Rodin Gen-2.5: One Image Generates 3D Assets with Tens of Millions of Polygons—Game and Film Creators, Take Note
Back when I was doing 3D virtual humans at ByteDance, one asset meant coordinating modelers, riggers, and texture artists—just communication could break you. Now Rodin Gen-2.5 takes one image and outputs obj, fbx, glb formats ready to import into 3D software for further editing.
This is reportedly the world’s first 3D generation model hitting tens of millions of polygons, using native 3D texture algorithms with noticeably finer material detail than competitors. For game, animation, and film studios, this isn’t “might be useful someday”—it’s production-ready now, saving real labor and time.
5. Agent App UI Revolution: One Product Ditched the Sidebar Entirely, Switched to Browser-Style Top Tabs
“Sidebars suck”—that was the conclusion when one Agent desktop app redesigned, so they deleted it and went with top tabs instead.
This design choice deserves attention. Sidebar switching can destroy old conversations and free memory; top tabs keep all sessions by default, creating more memory pressure. But UX-wise, tabs feel more like browser habits, switching is more intuitive. It’s a classic “engineering tradeoff for user experience” call, and Agent app interaction patterns are still evolving fast—whether this direction works is worth watching.
6. Codex Fully Automated Chrome Extension Store Form Submission, Even Solved Chinese Input Method Interference
Chrome extension store submission requires filling out tons of forms—product descriptions, feature explanations, copy everywhere. If you’re lazy, there’s good news. Someone used Codex client plus Chrome’s official plugin to fully automate the entire submission process.
Here’s the clever part: Codex detected that Chinese input method was interfering with content entry, so it wrote its own Python script to work around it using direct copy-paste instead. Not fast, but it actually completed everything—submission is in review. This isn’t just “AI can fill forms”; it’s AI encountering obstacles and actively problem-solving—that’s the core of Agent capability.
7. Vibe Coding Tip: Have AI Draw ASCII Layout First, Then Code—Way Better Interaction Design
Ask AI to write code directly and the interface layout often comes out weird—because AI doesn’t have a “think first, then act” habit. Simple fix: have it draw the page layout in ASCII characters first, confirm the structure makes sense, then start coding.
The core idea is splitting “design” from “implementation,” forcing AI to plan first. ASCII diagrams have zero technical overhead, but they let you see in seconds whether AI understands the layout you want, way easier than rewriting after code is done. Vibe Coding users can bookmark this one.
8. AI Pushes To-Dos to E-Ink Display, Turns Into Business Card When Off—This Is E-Ink Hardware Done Right
Most people buy e-ink displays and don’t know what to do with them—too slow, not good for scrolling content, just sits there. Someone found the perfect use case: on boot, AI pushes today’s to-dos, calendar, and key info; mounted on screen with magnets; on shutdown, it uses e-ink’s power-off persistence to auto-switch to a business card, scan to add contacts.
The genius here is fully leveraging e-ink’s physical properties instead of forcing it to do what it’s bad at. Will be packaged as Skills and opened up—anyone with similar hardware can wait to play with it.
9. ByteDance Scholarship Opens Globally for First Time; Five-Year Tracking of 67 Recipients
ByteDance’s scholarship program is opening globally for the first time this year—a signal worth watching. Over five years, the 67 recipients’ paths basically map out where top AI talent is flowing—which labs, companies, schools. It reflects the entire industry’s talent competition landscape.
With AI talent competition at fever pitch, ByteDance taking scholarships global is both brand play and early talent pipeline building. For current students, this pathway deserves serious study.
10. Formal Skill: Making LLM Agent Workflows Truly Programmable, Executable, and Accountable
The biggest problem with LLM Agents now isn’t “can they reason” but “can they reliably execute after reasoning.” This paper introduces the Formal Skill framework, upgrading Agent skills from natural language descriptions to programmable runtime structures—workflow state, policy execution, error handling all explicitly defined, no more guessing what comes next.
For Agent app developers, this directly addresses a real pain point: there’s a massive reliability gap between Markdown prompts and function calls. Formal Skill tries to bridge it. Academic paper, but real engineering value.
📌 Worth Watching
[Product] Doubao Integrates Seedance 2.0 and DeepSeek, Targets Workplace Agent Market — Domestic giants are all fighting for workplace turf; Doubao’s integration of top domestic models this time—whether paid subscriptions can build user habits is a critical step for the entire AI supply chain.
[Research] Trust3R: Giving 3D Reconstruction Models the Ability to Say “I’m Uncertain” — Current 3D reconstruction model confidence scores are basically guesswork; Trust3R introduces evidential uncertainty frameworks so models can say “I’m unsure about this geometry,” super useful for robotics and autonomous driving.
[Research] Pseudocode-Guided VLM Reasoning, New Solution for Robot Decision Hallucinations — Vision-language models on robots fear hallucinations most; this paper uses structured pseudocode to constrain reasoning, so robots stop “making up” things that don’t exist—safety boost is significant.
😄 AI Fun
Used GPT to Gacha a New Profile Photo, “Actually a Bit Thicker IRL but It Works”
Someone used GPT to generate a profile photo of themselves, gacha-style image generation, and the result came out “a bit slimmer” than reality. Their exact words: “I’m actually a bit thicker than this in real life, but it works.”
That’s worth thinking about. AI generating your profile photo naturally leans toward the flattering side—not a bug, kind of a feature. Just before using it officially, maybe think: what if the other person meets you and doesn’t recognize you? Prompt’s in the replies if you want to try your own gacha—might end up happier with it than the mirror.
🔮 AI Trend Predictions
Agent Toolchain Standardization Competition Goes Mainstream
- Timeline: Q3 2026
- Confidence: 78%
- Reasoning: Today’s Formal Skill paper + get-shit-done hitting 63k stars appearing simultaneously signals developer demand for “how do Agents execute reliably” has hit critical mass. Next 3 months will see dense waves of frameworks, specs, and tooling around Agent workflows, with de facto standards starting to compete.
AI Tools’ “Backlash Sentiment” Will Force Product Repositioning
- Timeline: Q2-Q3 2026
- Confidence: 65%
- Reasoning: Today’s graduation boo incident + Gallup showing public AI excitement dropping. This signal cascades to product layer: “AI replaces you” narratives get harder to sell, while “AI does the boring stuff you hate” tools (auto-fill forms, auto-organize docs) gain traction instead.
3D Generation Tools Enter Game/Film Production Mainstream
- Timeline: Q3 2026
- Confidence: 70%
- Reasoning: Today’s Rodin Gen-2.5 hitting tens of millions of polygons , precision now at “direct import to pro 3D software” level. Historically, every time production tool precision crosses the “good enough” threshold, commercial adoption speed jumps suddenly. Game and film studio procurement cycles run 3-6 months; expect visible real-world cases in H2.
AI Coding Tools Shift from “Write Code” to “Full Workflow Automation”
- Timeline: Q2 2026
- Confidence: 72%
- Reasoning: Today’s Codex auto-completing Chrome extension store submission , including writing its own script to bypass obstacles. Shows AI coding tool capability boundary already expanded from “write code” to “complete an entire dev task.” More tools will productize this, extending from code generation through testing, deployment, and store submission.
❓ Related Questions
How do I access Claude Code?
Claude Code is Anthropic’s command-line AI programming tool, currently requiring a Claude Pro or Claude API account. For domestic users, registering an Anthropic account needs an overseas phone number, and payment requires a foreign credit card—high barrier.
Solution: Visit Aivora to get a ready-made account—instant delivery, worry-free support.