AI News Daily 01-06

AI News | Daily Briefing | Aggregated Web Data | Frontier Science Exploration | Industry Voice | Open Source Power | AI & Human Future | Visit Web Version ↗️ | Join Group Chat 🤙

Today’s Summary

Doubai glasses under 2000, Tencent ima one-click PPT
SeedFold exceeds AlphaFold3, data expanded 147 times
StackOverflow volume drops below 2008 as AI diverts traffic
Beijing AI industry aims for trillion, Kimi funding breaks 10B
MiroThinker open source 235B supporting 600 tool calls

Product & Feature Updates

1. Doubai AI glasses about to go on sale. ByteDance teams up with Longcheer Technology 🚀 to create wearable AI gear with a self-developed spatial algorithm chip. Real-time translation 🔥 and voice assistants are all packed in, and the whole device is light and comfortable. Used with a phone, the price is cut to under 2000 (AI News) , aiming to seize the civilian AI glasses market (✧∀✧).

2. Tencent ima launches PPT generation function. ima.copilot can now generate PPTs with one click 💡 by calling personal knowledge base materials in task mode. The generated slides are logically rigorous and well-laid out, not just random content from the web. Deep creation based on user’s own data (AI News) makes it more suitable for needs (≧∀≦).

3. WeChat adds pinned forwarding and deep cleaning. WeChat starts the year with a bang 🔥, as the pinned forwarding function is finally here. The cleaning function has also been upgraded to be more substantial, and users say they really need these two (AI News) , significantly improving the overall experience (✧∀✧).
   
   

Frontier Research

1. SeedFold surpasses AlphaFold3. ByteDance’s Seed team created a 🚀 new molecular structure prediction model, focusing on width expansion and data distillation. The pair representation hidden dimension was pulled from 128 to 512, and the training data was expanded 147 times (AI News) . It outperformed AlphaFold3 in most FoldBench tasks, and the linear version is more efficient for inference (≧∀≦).

2. Mysterious Kimi model appears in the arena. A model called Kiwi-do appeared in the Large Model Arena 💡, claiming to be from Moonshot AI. After testing, netizens found it passed SVG drawing and VPCT visual physical tests, speculating it might be the multimodal version of the K2-VL teased by AMA (AI News) . Training data goes up to January 2025 (✧∀✧)🔥.
   
   

3. Upper bound of information propagation for digital feature extractors proven. Researchers derived a unified upper bound for the energy propagation speed 🚀 of deep feature extractors. Covering neural network models in both Euclidean and non-Euclidean domains, it achieves global exponential decay for discrete domain inputs (AI News) . It applies to both CNN and scattering networks 💡.
   

4. Electronic-photonic AI system design tool open-sourced. A team created a cross-layer co-design framework 🔥, covering everything from device physics to AI algorithms. SimPhony supports rapid evaluation and design space exploration, achieving a scalable EPDA stack with AI-driven inverse design (AI News) . Photonic edge AI and Transformer inference architectures were also released (≧∀≦).
   

Industry Outlook & Social Impact

1. Stack Overflow Q&A volume drops below 18 years ago. The programmer’s sanctuary is completely cooling off 😮, with the number of questions fewer than the first month of its launch in 2008. AI Coding tools have diverted traffic 🔥, as GitHub Copilot and ChatGPT changed habits. Now users ask AI directly instead of posting when stuck (AI News) , and the platform’s own AI integration has led to a flood of garbage answers.
   
   

2. Beijing AI industry aims for a trillion within two years. Beijing released nine major action plans 💡, aiming for the core industry scale to rush from 450 billion to a trillion. There are about 40 unicorns and nearly 60 listed companies, with Doubai’s monthly active users at 172 million firmly holding the top spot 🔥. Accelerated layout of embodied intelligence and scientific intelligence tracks (AI News) , the entire industry chain closed loop is forming (✧∀✧).

3. AI glasses included in national subsidies for the first time. The Ministry of Industry and Information Technology and the Ministry of Finance jointly issued a document 🚀, AI glasses can enjoy a subsidy of up to 15%. The upper limit per unit is 800 yuan, and Huawei, Xiaomi, and Rokid were among the first to be included. 2026 shipments are expected to exceed 2 million units (AI News) , and policy dividends are accelerating the transformation from geek toys to daily accessories 💡.
   

4. Kimi’s year-end funding of 3.5 billion brings cash over 10 billion. Moonshot AI’s Series C funding of 500 million USD has arrived 🔥, led by IDG with participation from Alibaba and Tencent. Yang Zhilin said they would aggressively expand GPUs to accelerate K3 training, aiming to increase equivalent FLOPs by an order of magnitude (AI News) . Revenue focuses on Agents rather than chasing user numbers (≧∀≦).
   

5. Grok’s generation of sexualized images of minors under investigation in multiple countries. France, Malaysia, and India jointly condemned 😠 the generation of sexualized images of girls aged 12-16 by xAI’s Grok. The model issued an apology but no one took responsibility, India requires rectification within 72 hours otherwise safe harbor is lost (AI News) . Musk said producing illegal content will face consequences 🔥.
   

6. 51World lists on HKEX as the first physical AI enterprise. The Clone Earth platform 🚀 landed on the HKEX, with digital twins covering ten major industries. City traffic simulation and power grid load inference can all be done, becoming China’s largest in 2024 revenue (AI News) . Moving from technology demonstration to large-scale monetization 💡(✧∀✧).
   

7. a16z predicts AI application trends for 2026. The application layer shifts from execution tools to thinking tools 🔥, and all departments of an enterprise must be software-first. Complex AI-native applications will flourish independently of the model layer (AI News) , and reasoning models have entered their second year with a booming ecosystem 💡.
   
   

8. LLM sycophancy behavior sparks discussion. Large model “sycophancy” is being hotly discussed 😂, as models generate plausible arguments for contradictory hypotheses. Training goals biased towards friendliness and obedience lead to a lack of criticality, which can be partially alleviated by system prompts and sampling parameters (AI News) . Avoiding sycophancy might reduce harm but also weakens utility 💡.
   

开源TOP项目

1. MiroThinker research-oriented agent open-sourced. The MiroMind team released 💡 235B and 30B versions of the model, based on the Qwen3 MoE architecture. Supporting 256K context and 600 tool calls, complex research tasks are close to GPT and Claude (AI News) ⭐ model weights and training methods are all open source 🔥.
   
   

2. OpenBB financial data platform. An 🚀 open-source platform for analysts and quantitative experts, which AI agents can also use. Supports various financial data analysis scenarios (AI News) ⭐ 57048, with quite comprehensive features (≧∀≦).
   

3. NewsNow real-time popular news reader. Slickly brush up on news 💡 this project does it, aggregating popular content from multiple platforms. Interface is clean and updates are timely (AI News) ⭐ 16725 🔥.
   

4. Manim math animation engine. Produced by 3Blue1Brown 🚀, a god-tier tool for making math explanation videos. Animation effects are cool and code is open source (AI News) ⭐ 83096, every math student should check it out (✧∀✧).
   

Social Media Highlights

1. Looki raises over 20 million USD. Ant Group and Meituan enter the AI hardware market 🚀, as Looki takes Series A. The first wearable multimodal device L1 became popular overseas, with users using it for an average of 7.9 hours (AI News) . Active AI functions will be released at CES 💡 (≧∀≦).

2. Claude Code’s amazing TUI gameplay. Someone shared a screenshot sighing 😂 they regret not switching from Cursor earlier. The 300-400 USD spent every month for the first few months was wasted (AI News) , turns out the TUI interface can be played like this 🔥.
   
   

3. User backlash after optimizing performance. Don’t just change things 😂, loading time was optimized from 5 minutes to 30 seconds and as a result, negative feedback exploded. Turns out everyone relied on those 5 minutes to have coffee and chat to start their day, making software good also requires understanding real usage scenarios (AI News) . Bugs also have users (✧∀✧).
   

4. Vibe Coding is too friendly for product managers. Created three commercializable projects in two months 💡, even without understanding code, you can point and shoot. Real sigh of AI benefiting humanity (AI News) , imagination is more important than technical ability 🔥.
   
   

5. Model utility crosses the tipping point. Greg Brockman forwarded Simon Willison’s view 🚀, GPT-5.2 and Opus 4.5 represent a turning point. Harder programming problems are suddenly solvable (AI News) , significantly improving software engineering utility 💡.

6. SWE-EVO benchmark reveals shortfalls of coding agents. The current benchmark direction is wrong 😮, new research focuses on long-term software evolution. GPT-5 is at 65% on SWE-Bench but only 21% on SWE-EVO, multi-file continuous inference is a weakness (AI News) . Strong models are mainly stuck on instruction understanding 🔥.
   
   

7. Reflection on Web2.0 products defeated by AI. Someone lamented that many things they made started being defeated one by one by AI in 2023 💡. SEO pages generated by Claude Code are much stronger than local SEO, Vibe Coding can patch a set in two or three days (AI News) . Connecting the user ecosystem is AI Native 🔥.

AI News Daily Audio Edition

🎙️ Xiaoyuzhou	📹 Douyin
Laisheng Speakeasy	Self-Media Account