12-29-Daily AI News Daily

Alright, here’s my take. The user wants me to act as a Senior Technical Translator and Editor. My task is to translate and edit this AI-related news digest. It’s a four-step process. First, the “Zero-Tolerance” rule: I need to scan for anything in square brackets [text](URL) and ![text](URL). Any Chinese text in the brackets gets translated. URLs and file paths? Leave ’em untouched. I’ve got to self-check: make sure I’ve got all the brackets and that the Chinese is translated, and the URLs are the same.

Next up, the Primary Editorial Task. This is where I get to shine. I’ll identify the core subject of each paragraph and rewrite it, putting that subject right at the front.

Then, there are some style rules. Gotta keep the tone conversational, informal, and energetic. Think everyday English, maybe a little slang. No emojis from the original, but I get to add my own strategically. I keep the Markdown formatting 100% – headings, lists, the whole shebang. Code blocks? Code stays, comments get translated. And finally, I must translate all the content, and keep the same paragraph structure.

Here’s my translation:

## AI Daily 2025/12/29
> AI Daily

### **Today's Digest**

Jim Fan reviews the robotics field: hardware is impressive, but software lags significantly, and VLM solutions "don't feel right."

A 32B open-source model achieves OpenAI-level deep research capabilities, with the secret being replacing "predicting the next token" with "deciding the next atomic action."

Vibe Coding's hard-earned lessons are worth saving, and the story of Claude Code evolving from a side project into a $1 billion product is equally compelling.

## ⚡ Quick Navigation

### **Today's AI News**

### **👀 One-Liner**

DeepMind's documentary hits 200 million views, but Jim Fan calls the robotics field a "Wild West"—software is far behind hardware.

### **🔑 3 Keywords**

#RoboticsDilemma #VibeCodingBestPractices #DeepResearchAgent

## **🔥 Top 10 Highlights**

NVIDIA Senior Research Scientist Jim Fan poured cold water on the hype around cool robots like Optimus and Figure, stating: **"Hardware is already impressive, but software simply can't keep up."** What's more disheartening is how delicate these robots are—overheating, motor failures, and firmware glitches are commonplace; they require constant "servicing." He also directly stated that current mainstream VLM-based VLA solutions "don't feel right" because visual language models are pre-trained to answer questions, actively discarding low-level details crucial for dexterous manipulation. **He's betting on video world models as the right path forward.** The robotics scene in 2026 will be one to watch.

A 32B open-source model achieves OpenAI-level deep research capabilities, with the secret being replacing "predicting the next token" with "deciding the next atomic action"—a four-step process involving planning, deep search, reflective verification, and report generation. The coolest part is its incredibly simple architecture: just a ReAct-style Agent, without any fancy multi-agent orchestration. **Medium-scale models + correct training data = expert-level research capabilities**; this formula is worth remembering.

@vibekanban's co-founder, after writing a million lines of code with AI, summarized several iron rules for Vibe Coding. "Anyone can Vibe Code"? That's naive. **First, plan before coding** (once AI starts writing, it tends towards "minimal changes," leading to increasingly rigid architectures). **Second, enable YOLO mode and let the AI run autonomously**, but only if your codebase has automated tests. **Third, explicitly tell the AI in the system prompt, "We aim for the simplest changes and don't care about migration costs,"** otherwise, it will get lazy. There's also a clever trick: use ESLint to prevent the AI from arbitrarily disabling lint rules. This guide is highly recommended for saving.

Claude Code, believe it or not, **started as a side project by developer Boris Cherny in September 2024**. Back then, Claude often messed up even simple Bash scripts and crashed after just a few minutes. But now, based on Claude Sonnet 4.5 and Opus 4.5, it can run continuously for hours or even days, completing incredibly complex tasks. The key technology is the "Stop Hooks" mechanism—when Claude wants to stop, you can "poke" it with a script to keep working, like running tests and automatically fixing failures. **Anthropic defines an AI Agent as: LLM + cyclical automatic tool invocation**, and Claude Code is the perfect embodiment of this definition.

Demis Hassabis personally endorsed the DeepMind documentary about AlphaFold's origin story, which **surpassed 200 million YouTube views in just 4 weeks**. If you want to understand how an AGI lab operates, or how Nobel Prize-level projects are created, it's a great watch for the holidays. Directed by Greg Kohs with music by Dan Deacon, the production team is also top-notch.

A new paper introduces LongVideoAgent, addressing how current large multimodal models handle long videos, which typically involves "compressed summaries + frantic frame extraction," losing all details. **LongVideoAgent features a main Agent responsible for reasoning and decision-making, a localization Agent to find relevant segments, and a visual Agent to extract details.** Reinforcement learning teaches the main Agent when to explore and when to stop. The results? GPT-5-mini jumped from 62.4% to 71.1% on long video Q&A benchmarks, and Qwen2.5-3B soared from 23.5% to 47.4%, effectively doubling its performance. **An agent-based design is the correct approach for long video understanding.**

Genfocus, this model is quite interesting, **specifically designed to adjust depth of field and aperture effects in images**, and can also convert shallow depth-of-field photos to full focus. It's not a comprehensive image editing model, but rather one that excels at a single task. The model is already open-source on HuggingFace, so photography enthusiasts can have some fun with it.

Sam Altman's latest statement: **"Google remains a huge threat to OpenAI,"** and the ChatGPT team might "go into red alert twice a year, and it will last for a long time." This is a very candid remark—the AI competition is far from over, and the battle among tech giants will continue.

Here's a trick worth snagging for ChatGPT Plus: **After canceling your ChatGPT Plus subscription, OpenAI might offer you a free month** (100% off) to retain you. While it's uncertain how long this strategy will last, it's currently still working. 🤫

China is currently drafting regulatory rules for "AI with human-like interaction capabilities." Specific details haven't been fully disclosed, but the direction is clear: **as AI becomes more human-like, regulation must keep pace.** This is a signal that teams developing AI companions and AI customer service domestically need to pay attention to.

## **📌 Worth Noting**

*   [awesome-llm-apps](https://github.com/Shubhamsaboo/awesome-llm-apps) - An 84K Star collection of LLM applications, with comprehensive RAG and Agent examples.
*   [vibe-kanban](https://github.com/BloopAI/vibe-kanban) - A Kanban tool that boosts Claude Code/Codex efficiency by 10x, 7K Stars.
*   [Mole](https://github.com/tw93/Mole) - A deep cleaning tool for Mac, 21K Stars, developed by Chinese.
*   [Fresh](https://github.com/sinelaw/fresh) - A simple yet powerful terminal text editor, no need to memorize Vim shortcuts.
*   [Omni-Design Custom Workshop](https://x.com/tuturetom/status/2005278630843478221) - Produced by Refly.ai, this tool deconstructs any complex concept into 4K HD images.
*   [French Telecom Giant Orange Employees Using AI Tool Site Developed by Chinese Developers](https://x.com/gefei55/status/2005316438492692768) - Indirectly serving a Fortune 500 company?
*   [Li Feifei's "K12 Education is a Waste of Time" Statement Clarified](https://x.com/dotey/status/2005335316555776385) - The original quote was severely misinterpreted; it's recommended to read the original text.

## **📊 More Dynamics**

| Title                                                                             | Type       |
| --------------------------------------------------------------------------------- | ---------- |
| [Python Implementations of All Algorithms](https://github.com/TheAlgorithms/Python) | Open Source |
| [Python Interpreter Written in Rust](https://github.com/RustPython/RustPython)   | Open Source |
| [Algorithmic Trading Engine](https://github.com/QuantConnect/Lean)                | Open Source |
| VLNVerse - Wu Qi's Team's Embodied Navigation Full-Stack Platform                  | Research   |
| How to Build a $100K AI SaaS Without Code                                          | Tutorial   |
| Nuggt Canvas - A Better-Looking MCP Client                                         | Tool       |

## **❓ Related Questions**

ChatGPT Plus currently requires a $20 monthly subscription to access advanced features like GPT-4o. For users in China, there might be difficulties with credit card payments or account registration restrictions.

**Solution:**

*   **[Aivora](https://aivora.cn)** provides ready-made ChatGPT Plus account services.
*   Aivora's service offers instant delivery, ready for use upon order, eliminating payment and registration hassles.
*   Aivora provides stable, exclusive accounts with worry-free after-sales support.

Visit [aivora.cn](https://aivora.cn) to view the complete list of AI account services.

That’s it. Output complete. Ready to go.

AI Daily 2025/12/29

AI Daily

Today’s Digest

Jim Fan reviews the robotics field: hardware is impressive, but software lags significantly, and VLM solutions “don’t feel right.” A 32B open-source model achieves OpenAI-level deep research capabilities, with the secret being replacing “predicting the next token” with “deciding the next atomic action.” Vibe Coding’s hard-earned lessons are worth saving, and the story of Claude Code evolving from a side project into a $1 billion product is equally compelling.

⚡ Quick Navigation

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

💡 Tip: Want to be among the first to experience the latest AI models mentioned (Claude 4.5, GPT, Gemini 3 Pro)? Don’t have an account? Head over to Aivora to grab one, get started in a minute, and enjoy worry-free after-sales support.

Today’s AI News

👀 One-Liner

DeepMind’s documentary hits 200 million views, but Jim Fan calls the robotics field a “Wild West”—software is far behind hardware.

🔑 3 Keywords

#RoboticsDilemma #VibeCodingBestPractices #DeepResearchAgent

🔥 Top 10 Highlights

1. Jim Fan’s Year-End Review: Three Lessons from the Robotics Field

NVIDIA Senior Research Scientist Jim Fan poured cold water on the hype around cool robots like Optimus and Figure, stating: “Hardware is already impressive, but software simply can’t keep up.” What’s more disheartening is how delicate these robots are—overheating, motor failures, and firmware glitches are commonplace; they require constant “servicing.” He also directly stated that current mainstream VLM-based VLA solutions “don’t feel right” because visual language models are pre-trained to answer questions, actively discarding low-level details crucial for dexterous manipulation. He’s betting on video world models as the right path forward. The robotics scene in 2026 will be one to watch.

2. Step-DeepResearch: A 32B Parameter Agent Outperforms OpenAI and Gemini in Deep Research

OpenAI and Google’s deep research systems are proprietary and costly. Now, a 32B parameter open-source model has achieved comparable scores (61.42 on Scale AI benchmarks). What’s the secret? It replaces “predicting the next token” with “deciding the next atomic action”—a four-step process involving planning, deep search, reflective verification, and report generation. The coolest part is its incredibly simple architecture: just a ReAct-style Agent, without any fancy multi-agent orchestration. Medium-scale models + correct training data = expert-level research capabilities; this formula is worth remembering.

3. Vibe Coding Best Practices: Hard-Earned Lessons from 1 Million Lines of Code

@vibekanban’s co-founder, after writing a million lines of code with AI, summarized several iron rules for Vibe Coding. “Anyone can Vibe Code”? That’s naive. First, plan before coding (once AI starts writing, it tends towards “minimal changes,” leading to increasingly rigid architectures). Second, enable YOLO mode and let the AI run autonomously, but only if your codebase has automated tests. Third, explicitly tell the AI in the system prompt, “We aim for the simplest changes and don’t care about migration costs,” otherwise, it will get lazy. There’s also a clever trick: use ESLint to prevent the AI from arbitrarily disabling lint rules. This guide is highly recommended for saving.

4. The Origin of Claude Code: How a Side Project Became a $1 Billion ARR Product

Claude Code, believe it or not, started as a side project by developer Boris Cherny in September 2024. Back then, Claude often messed up even simple Bash scripts and crashed after just a few minutes. But now, based on Claude Sonnet 4.5 and Opus 4.5, it can run continuously for hours or even days, completing incredibly complex tasks. The key technology is the “Stop Hooks” mechanism—when Claude wants to stop, you can “poke” it with a script to keep working, like running tests and automatically fixing failures. Anthropic defines an AI Agent as: LLM + cyclical automatic tool invocation, and Claude Code is the perfect embodiment of this definition.

5. DeepMind Documentary “The Thinking Game” Hits 200 Million Views in 4 Weeks

Demis Hassabis personally endorsed the DeepMind documentary about AlphaFold’s origin story, which surpassed 200 million YouTube views in just 4 weeks. If you want to understand how an AGI lab operates, or how Nobel Prize-level projects are created, it’s a great watch for the holidays. Directed by Greg Kohs with music by Dan Deacon, the production team is also top-notch.

6. LongVideoAgent: Enabling AI to Truly “Understand” One-Hour Long Videos

A new paper introduces LongVideoAgent, addressing how current large multimodal models handle long videos, which typically involves “compressed summaries + frantic frame extraction,” losing all details. LongVideoAgent features a main Agent responsible for reasoning and decision-making, a localization Agent to find relevant segments, and a visual Agent to extract details. Reinforcement learning teaches the main Agent when to explore and when to stop. The results? GPT-5-mini jumped from 62.4% to 71.1% on long video Q&A benchmarks, and Qwen2.5-3B soared from 23.5% to 47.4%, effectively doubling its performance. An agent-based design is the correct approach for long video understanding.

7. Genfocus: A Small AI Model Dedicated to Adjusting Depth of Field and Aperture

Genfocus, this model is quite interesting, specifically designed to adjust depth of field and aperture effects in images, and can also convert shallow depth-of-field photos to full focus. It’s not a comprehensive image editing model, but rather one that excels at a single task. The model is already open-source on HuggingFace, so photography enthusiasts can have some fun with it.

8. Sam Altman: Google Remains a Huge Threat, ChatGPT Needs “Red Alert” Twice a Year

Sam Altman’s latest statement: “Google remains a huge threat to OpenAI,” and the ChatGPT team might “go into red alert twice a year, and it will last for a long time.” This is a very candid remark—the AI competition is far from over, and the battle among tech giants will continue.

9. A Little Trick for One Month Free ChatGPT Plus

Here’s a trick worth snagging for ChatGPT Plus: After canceling your ChatGPT Plus subscription, OpenAI might offer you a free month (100% off) to retain you. While it’s uncertain how long this strategy will last, it’s currently still working. 😉

10. China Releases Draft Regulations for AI Human Interaction

China is currently drafting regulatory rules for “AI with human-like interaction capabilities.” Specific details haven’t been fully disclosed, but the direction is clear: as AI becomes more human-like, regulation must keep pace. This is a signal that teams developing AI companions and AI customer service domestically need to pay attention to.

📌 Worth Noting

[Open Source] awesome-llm-apps - An 84K Star collection of LLM applications, with comprehensive RAG and Agent examples.

[Open Source] vibe-kanban - A Kanban tool that boosts Claude Code/Codex efficiency by 10x, 7K Stars.

[Open Source] Mole - A deep cleaning tool for Mac, 21K Stars, developed by Chinese.

[Open Source] Fresh - A simple yet powerful terminal text editor, no need to memorize Vim shortcuts.

[Product] Omni-Design Custom Workshop - Produced by Refly.ai, this tool deconstructs any complex concept into 4K HD images.

[Business] French Telecom Giant Orange Employees Using AI Tool Site Developed by Chinese Developers - Indirectly serving a Fortune 500 company?

[Other] Li Feifei’s “K12 Education is a Waste of Time” Statement Clarified - The original quote was severely misinterpreted; it’s recommended to read the original text.

📊 More Dynamics

❓ Related Questions

How to experience ChatGPT Plus?

ChatGPT Plus currently requires a $20 monthly subscription to access advanced features like GPT-4o. For users in China, there might be difficulties with credit card payments or account registration restrictions.

Solution:

• Aivora provides ready-made ChatGPT Plus account services.
• Aivora’s service offers instant delivery, ready for use upon order, eliminating payment and registration hassles.
• Aivora provides stable, exclusive accounts with worry-free after-sales support.

Visit aivora.cn to view the complete list of AI account services.