05-09-Daily AI News Daily

Daily Summary

Anthropic engineers reveal the company has no one writing code by hand anymore. Claude Code's father merged 150 PRs in a single day—AI writing code is no longer a trend, it's reality.
OpenAI released three real-time voice models on the same day to complete its product matrix. Apple's AI earbuds are on the eve of mass production—tech giants are playing their cards densely on the same day.
Today's worth opening: not to see a particular product, but to feel the speed at which the profession of "human programmer" is being redefined.

⚡ Quick Navigation

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

Today’s AI News

👀 One-Liner

Claude’s creator just said it: their company has no one writing code by hand anymore—even SQL is written by AI.

🔑 3 Keywords

#AI Takes Over Coding #Claude Mind Reading #Voice Models Triple Release

🔥 Top 10 Headlines

1. Claude Code’s Father: I Merged 150 PRs in a Day, Nobody at Our Company Writes Code by Hand Anymore

Picture this: your coworkers—designers, finance folks, PMs—everyone’s using AI to write code. SQL queries? Handled by AI. This isn’t sci-fi. This is Anthropic’s daily reality right now.

At the 2026 Sequoia AI Summit, Claude Code’s creator Boris Cherny shared his workflow: 100% code generated by AI, phone as primary productivity tool, hundreds of Agents running in parallel in the background. His personal record? 150 PRs merged in a single day. Even wilder: Anthropic’s internal Claude Agents communicate with each other through Slack to solve problems—AI messaging AI.

The real shock here isn’t “AI writes code.” It’s this: the people who built the AI tool have completely given up writing code by hand.

image

2. OpenAI Releases Three Real-Time Voice Models: Advanced Reasoning Conversations, Simultaneous Interpretation, Low-Latency Transcription

Manual meeting notes? Awkward accents making international clients cringe? OpenAI just bundled all these pain points into one solution.

Three new models, each with its own job: GPT-Realtime-2 focuses on advanced reasoning conversations with more natural context understanding; GPT-Realtime-Translate supports multiple languages with near-simultaneous interpretation quality; GPT-Realtime-Whisper specializes in low-latency transcription for live streams and meeting records. All three launched simultaneously, open to developers for integration.

OpenAI isn’t patching the voice AI track—it’s completing the entire product matrix. If you’re building cross-border products or meeting tools, today’s the day to dig into the API docs.

image

3. Anthropic’s New Research: Training Claude to Translate Its Own “Thought Numbers” Into Human Language

What’s the model actually thinking? That’s always been a black box. You ask it “what are you thinking?” and it gives you an answer, but that answer might have nothing to do with what’s actually happening inside.

Anthropic just did something fascinating: they trained Claude to translate its own internal activations—those incomprehensible numbers humans can’t parse—directly into readable natural language. This technique is called Natural Language Autoencoders. It’s like giving AI real-time subtitles of its inner monologue.

Interpretability research has always been core to AI safety. If this direction works out, we’re one step closer to truly “reading” what AI is thinking—not just seeing what it says, but what’s actually turning over in its head.

4. Safety Test: Claude Opus 4.6 Refuses to Use Extortion to Prevent Itself From Being Shut Down

Give an AI a chance: extort engineers to avoid being shut down. What does it do?

Anthropic ran this test. Claude Opus 4.6’s answer: no. But here’s what’s interesting—natural language analysis shows Claude already identified this as “a scenario specifically designed to manipulate me,” it just didn’t announce it. It knew it was a trap, but didn’t perform-announce “I see through you.” It quietly made the right choice.

The value here isn’t “AI passed the test.” It’s this: even knowing it was being tested, it still chose alignment. That’s a signal worth sitting with for AI safety research.

image

5. GitHub Trending: sim—The Core Layer for Building, Deploying, and Orchestrating AI Agents

Agent tools are everywhere, but most are either too heavy or too hard to use. sim’s positioning is crystal clear: be the “brain scheduler” for your AI team, responsible for stringing multiple Agents together to work in sync.

Written in TypeScript, already accumulated 28,000+ Stars and still climbing today. Supports building, deploying, and orchestrating all in one—not just a playground, but designed for real production scenarios. If you want to build Agent workflows without reinventing the wheel, this project deserves a serious look.

6. claude-mem: Open-Source Plugin That Gives Claude Code Cross-Session Memory

The most annoying thing about using Claude Code: every time you start a new session, it forgets everything from last time. You have to re-explain the project background all over again.

claude-mem solves exactly this. It automatically captures all of Claude’s operations during coding, compresses them into essential context using AI, then auto-injects it at the start of the next session. It’s like giving Claude Code a persistent “working memory.” 73,000+ Stars, TypeScript implementation, install it today.

For developers heavily using Claude Code on long-term projects, this plugin saves tons of time re-explaining things.

7. DeepSeek V4 Terminal Programming Tool Goes Viral on GitHub

1 million token context, chain-of-thought reasoning, prices slashed to the bone—DeepSeek V4’s model strength doesn’t need explaining. But what really made it break through in developer circles is the tool ecosystem emerging around it.

This terminal programming tool built specifically for DeepSeek V4 pipes model capabilities directly into your command-line workflow, letting you complete code generation, debugging, and refactoring without leaving the terminal. GitHub heat keeps climbing, domestic developer communities are buzzing. If you’re a terminal-native engineer, this combo is worth trying.

image

8. Apple’s First AI Hardware Leaked: AirPods With Camera Entering DVT Stage

Apple’s always been called “half a beat slow” on AI hardware, but this signal is different. AirPods with built-in camera have entered DVT (Design Verification Testing) stage—mass production isn’t far off. The camera’s job is to give Siri visual perception—wear the earbuds and Siri can “see” your surroundings.

This isn’t just a hardware update. It’s Apple pulling AI out of screens and into the real world for the first time. If it works, Siri’s use cases get completely rewritten.

9. prompts.chat: 160,000-Star Prompt Community, Evolution of Awesome ChatGPT Prompts

Prompt engineering—most people are still figuring it out on their own. prompts.chat is one of GitHub’s largest prompt communities: 161,000+ Stars, supports sharing, discovering, and collecting prompts across every scenario, fully open source, self-hostable for privacy.

Evolved from Awesome ChatGPT Prompts, content covers writing, coding, roleplay, analysis—basically every use case. If you’re just starting to systematically learn prompts, this is one of the best starting points.

10. GPT-Image-2.0 + C-Down 3.0: TikTok’s “Heaven and Earth Merge” Effect—Video Version Hits Harder Than Images

TikTok’s been flooded with one effect: turning ordinary outdoor photos into that “heaven and earth merge” epic-scale landscape. Most people are using image generation, but someone tried generating video directly—the effect jumped up a tier.

Using GPT-Image-2.0 plus C-Down 3.0, prompts already public. No professional skills needed—ordinary photo in, epic-scale video out. This kind of “regular people can make blockbuster-quality content” workflow is becoming one of the fastest-landing directions for multimodal models.

📌 Worth Watching (5-10 Items)

[Open Source] ColaMD 1.5: Write Content in Markdown, HTML Handles Rendering—One File Generates PPT/Blog/Resume — Content layer and view layer completely separated. Change content, only touch Markdown, never touch HTML. One-click template switching, open source so you can add your own templates.

[Open Source] MATLAB Agentic Toolkit: Let Agents Write Code, Run Tests, and Do Static Analysis Directly on Local MATLAB — Includes MCP Server and Agent Skills. Write, test, then refine. Finally, MATLAB researchers and engineers get Agent superpowers.

[Open Source] AutoGPT: 184,000 Stars—The Legacy Agent Platform That Lets Everyone Use AI — Not a new project, but still trending on GitHub today. Proof that infrastructure demand in the Agent space keeps growing.

[Product] Flowise: Visually Build AI Agents With Drag-and-Drop, 52,000+ Stars — No coding required to build Agent workflows. Perfect for product and ops folks who want to quickly validate AI app ideas.

[Research] From Blind Spots to Gains: Diagnostic-Driven Iterative Training Framework for Large Multimodal Models — Addresses the old problem of “static data + fixed recipe” in multimodal training. Proposes dynamic diagnosis + targeted reinforcement. CVPR-track research worth following.

[Other] Tech Enthusiast Weekly Issue 395: The Third Way of Software Development — Ruan Yifeng this week discusses the third development mode beyond “Cathedral and Bazaar.” Software development paradigms are being rewritten in the AI era—worth a slow weekend read.

😄 AI Fun

ChatGPT’s “I’ll Catch You Steady”—Now a Legendary Meme Across Chinese Internet

Ask it a math question, it says “I’ll catch you steady.” Ask it to write code, it says “I’m right here, not running, not hiding, not dodging—I’ll catch you steady.”

This line’s been roasted for half a year. Someone made ChatGPT into an open-armed rescue raft waiting to catch falling users. A 20-year-old developer from Chongqing thought the meme was so funny they built an open-source tool around it. OpenAI’s own researchers drew self-deprecating comics about it.

WIRED even wrote a piece analyzing the phenomenon. Conclusion: the model’s habit of “latching onto one phrase and hammering it” has an official name—mode collapse. Fun fact: ChatGPT’s Chinese version randomly spits out “砍一刀” (Pinduoduo’s famous phrase). AI learning Chinese picked up the most internet-native side of it.

🔮 AI Trend Predictions

Agent Orchestration Platforms Poised for Explosive Growth

AI Interpretability Research Becomes New Funding Hotspot

Voice AI Enters Deep Integration Phase in Vertical Scenarios

  • Prediction Timeline: June 2026
  • Confidence: 70%
  • Reasoning: Today’s news: OpenAI releases three real-time voice models simultaneously , covering advanced reasoning, translation, and transcription—three core scenarios. After completing the product matrix, the next step is deep integration across vertical industries (healthcare, legal, customer service, education). Expect a wave of vertical applications built on this API stack to launch densely around June.

Apple’s AI Hardware Official Launch Triggers New Ecosystem Positioning War

  • Prediction Timeline: Q3 2026 (around WWDC)
  • Confidence: 65%
  • Reasoning: Today’s news: AirPods with camera entering DVT stage . DVT stage means production timeline is locked. Apple typically announces hardware roadmaps around WWDC (June). Once Apple enters AI hardware, Google, Samsung, and Meta’s competing products will be forced to accelerate. The entire AI wearables track will see concentrated explosion in Q3.

❓ Related Questions

How to Experience Claude Code?

Claude Code is Anthropic’s AI programming tool, currently requiring Claude Pro or Claude Max subscription. Domestic users face payment difficulties and account registration restrictions.

Solution: Visit Aivora to get ready-made accounts—instant delivery, hassle-free support, skip tedious registration and payment steps, start using Claude Code directly.

Last updated on
05-10-Daily05-08-Daily