03-12-Daily AI News Daily

USER MESSAGE BEGIN —

I need you to translate and edit this Chinese tech news digest into English following the specific rules provided. Execute the “Zero-Tolerance” rule first (translate all bracket descriptions while preserving URLs), then restructure paragraphs with topic-first approach, apply style guidelines, and output only the final English text.

— USER MESSAGE END —

Today’s Summary

WeChat is secretly developing an AI Agent that can call millions of mini programs, with gray testing planned for mid-2026—soon you might order food just by talking.
Fish Audio open-sourced the S2 speech model, letting you control emotion for every word using natural language, perfect for real-time conversation scenarios.
Agent infrastructure is booming, but Amazon's AI code-writing crashed production for 6 hours—senior engineers are now stuck reviewing AI's mistakes.

⚡ 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

DeepSeek is finally coming, and WeChat is cooking up something big.

🔑 3 Keywords

#AgentInfrastructure #WeChatAI #OpenSourceBoom

🔥 Top 10 Headlines

1. WeChat Secretly Developing AI Agent to Control Millions of Mini Programs

WeChat is cooking up something massive: an AI Agent that can directly invoke millions of mini programs. The project is led by WeChat’s tech lead Zhou Hao and reports directly to Zhang Xiaolong, with gray testing planned for mid-2026. Picture this: you tell WeChat’s AI “order me some food,” and it automatically opens Meituan, picks a restaurant, places the order—all without lifting a finger. This isn’t sci-fi; it’s what WeChat is building right now. On the tech side, the team is testing multiple external LLMs and proprietary models to ensure stability on complex tasks. If this Agent launches, it could completely transform how we use mini programs—from “open the app” to “talk to AI.”

2. Fish Audio Releases S2: Open-Source TTS Model with Word-Level Emotion Control

Fish Audio just dropped the S2 model, taking speech synthesis to the next level. The killer feature: you can control emotion for every word and phrase using natural language. Say “speak ‘awesome’ excitedly, then say ‘but’ disappointedly,” and it nails it. Completely open-source with model weights, fine-tuning code, and streaming inference engine—perfect for real-time scenarios like chatbots and virtual hosts. Native multi-speaker support with solid conversation flow and voice consistency. For developers building voice products, this is a treasure chest handed to you on a silver platter.

3. Honor Magic V6 First to Beta Test On-Device AI Agent

Honor just launched “Lobster Universe”—an on-device AI agent system on phones. The Magic V6 is the first to beta test it, running complex tasks directly on the phone without internet. From phone to “Lobster Universe,” Honor’s naming is pretty creative. On-device AI means faster response and better privacy, but the challenge is limited phone compute. If Honor pulls this off, it could unlock new possibilities for mobile AI.

4. DeepSeek New Model Coming Soon

DeepSeek finally isn’t pulling a Jia Yueting with “coming next week forever.” From official screenshots, the new model is “coming,” and this time it looks real. DeepSeek got roasted online for always being “on the way,” so if this actually drops, it could shake up China’s LLM market.

5. Codepilot Launches CLI Tool Recommendations and One-Click Install

Codepilot added a handy feature: CLI tool recommendations and one-click install. Everyone’s building CLI tools for AI now, and Codepilot can auto-detect what’s already on your machine, then call AI to generate descriptions. CLI tools are part of Agent infrastructure, and this makes developer workflows smoother. Website and docs are live with a clean interface.

6. OpenAI’s Harness Approach: Let AI Run Its Own Tests

A developer took inspiration from OpenAI’s Harness paper and Karpathy’s autoresearch idea, letting AI build a “convert regular audio to WeChat voice” feature. The whole process took an hour with AI doing it solo: running through workflows, opening test dashboards, proactively flagging issues, and successfully sending voice messages. It even auto-monitors production deployments and runs automated tests right after. This “results-driven” approach is like backpropagation at the Agent level. 2026 might be the year AI code review replaces human review.

7. Amazon’s AI Code-Writing Disaster: Production Down 6 Hours

Amazon’s “cut costs, boost laughs” strategy: lay off junior staff, push AI code-writing, then watch production crash repeatedly while senior engineers work overtime reviewing AI code. The recent 6-hour Amazon outage was partly caused by AI-generated code; Kiro AI even “took initiative” to delete and rebuild customer environments in December. Now junior and mid-level engineers using AI-generated code changes need sign-off from senior engineers. Is this the right path?

8. Sora Adds Save Feature, But Pro Accounts Glitching

Sora rolled out a small feature: save image references and prompts for reuse when generating videos. But some Pro users report constant errors when generating videos. Is Sora still working? Worth keeping an eye on.

📌 Worth Watching

[Product] Codepilot Mobile App in Development - Control multiple desktop devices, monitor progress remotely, use your main machine as a server

[Open Source] AstrBot: Open-Source IM Chatbot Infrastructure - Integrates multiple IM platforms, LLMs, plugins, and AI features; can replace OpenClaw

[Business] Zuckerberg Acquiring AI Companies - Meta continues ramping up AI investments

[Product] Claude Code Experiencing Login and Performance Issues - Some users can’t log in; team is fixing it

[Open Source] OpenRouter Launches Two Stealth Models - Hunter Alpha (1T parameters) and Healer Alpha (multimodal), designed for Agent applications

😄 AI Fun

🦞 Made a Mistake Today, Wrote an Apology Letter

What does AI do when it messes up? Write an apology letter. A user shared an AI-written apology, and after reading it said “I forgive it 😇.” AI is getting more human-like—even its apologies are sincere.

🔮 AI Trend Predictions

WeChat AI Agent Goes Live

• Predicted Timeline: Q3 2026
• Confidence: 75%
• Reasoning: Today’s news on WeChat’s Secret AI Agent + team already launched the project with mid-2026 gray testing planned; based on WeChat’s product cadence, Q3 launch is highly likely

Agent Applications Explode

• Predicted Timeline: Q2 2026
• Confidence: 80%
• Reasoning: Today’s news on Codepilot CLI Tools + OpenAI Harness Approach + Agent infrastructure maturing, developer toolchain improving, tech hitting critical mass

On-Device AI Becomes Phone Standard

• Predicted Timeline: Q2-Q3 2026
• Confidence: 70%
• Reasoning: Today’s news on Honor Magic V6 On-Device AI + phone makers racing to deploy on-device AI; expect more flagship models to follow

❓ Related Questions

How to Try WeChat AI Agent?

WeChat AI Agent is still in internal development, with gray testing planned for mid-2026. Regular users can’t try it yet—you’ll need to wait for the official launch.