AI Daily: DouBao Tests Social Features; AutoHome Tests Daima Enters AI Programming; Sina VibeThinker-3B Open Source

Welcome to the "AI Daily" section! This is your guide to exploring the world of artificial intelligence every day. Every day, we present you with the latest content in the AI field, focusing on developers to help you understand technology trends and learn about innovative AI product applications.

Fresh AI products Click to learn more:https://app.aibase.com/zh

1. Doubao Tests Social Function: Integrates Feishu Account, Is AI Assistant Also Doing Familiar Social Networking?

Doubao tests social functions, integrates Feishu account, is AI assistant also doing familiar social networking?

AiBase Summary:

📱 Doubao adds a new "Conversation" page, supporting adding "Doubao friends" or "Feishu friends".

👥 Doubao account can log in to Feishu web version, information synchronizes to Feishu, but core office functions are not yet available.

🚀 Industry analysis believes that Doubao has the potential to become a "Super App", connecting Feishu aims to enter the workplace social scene.

2. Smaller Models, Equal Capabilities: Sina VibeThinker-3B Open Source, Bringing New Ideas to AI Inference "Lightweight"

The article discusses whether the number of parameters in an AI model must be larger to be better, using Sina's open-source VibeThinker-3B model as an example, demonstrating its outstanding performance in complex tasks. The model achieves a balance between performance and scale through unique training strategies, offering new optimization ideas for the AI industry.

AiBase Summary:

🧠 VibeThinker-3B model achieves high performance through multi-stage post-training.

💡 Research proposes the "Parameter Compression-Coverage Hypothesis," revealing the compressibility of AI capabilities.

🚀 The open-sourcing of this model provides developers with a lightweight and efficient tool, promoting the optimization paradigm of the AI industry.

3. Gaode Tests "Daima" to Enter AI Programming, Generating Usable Applications in Minutes with a Single Sentence

Gaode Map quietly tests the "Daima" product, aiming to let users quickly generate usable applications without programming through natural language-driven methods. The product mainly serves small and medium-sized enterprise owners and independent creators, supporting dialogue-style modifications and an inspiration square function, significantly reducing the cost of turning ideas into reality. The product is currently in the testing phase and has not yet announced specific launch time or commercial details. With the popularity of AI programming tools, the development threshold has significantly decreased, and the era of mass application development is accelerating.

AiBase Summary:

🤖 Gaode Map tests the "Daima" product, generating applications through natural language-driven methods.

💡 Provides convenient development tools for small and medium-sized enterprise owners and independent creators.

🚀 Competition in the AI application generation market is fierce, and Gaode focuses on small-scale needs.

4. Baidu Opensources 3B Model Unlimited OCR: Star Count Exceeds 10,000 in 5 Days, Setting a New Record for Long Document Parsing

Baidu's 3B parameter end-to-end OCR model, Unlimited OCR, has achieved breakthroughs in long document parsing. Its innovative Reference Sliding Window Attention mechanism effectively improves processing efficiency and accuracy, and it shows significant speed advantages in practical applications, providing new insights for document digitization and large model long-term memory management.

AiBase Summary:

📚 Introduces Reference Sliding Window Attention mechanism, enabling one-time parsing of long documents.

📈 Achieves a high score of 93.92% in the OmniDocBench v1.6 benchmark test.

⚡ The inference speed is 12.7% faster than DeepSeek OCR, showing even greater advantages in long outputs.

5. Ant Afu Announces Progress on Losing 100 Million Catties: Already Lost 100,000 Catties, Netizens: I Am Still Holding Back Again

The article introduces the progress of the "Scientific Weight Loss of 100 Million Catties" project initiated by Ant Afu, analyzes the weight loss situation among different groups and regions, and emphasizes that scientific weight loss requires long-term persistence.

AiBase Summary:

🍎 Ant Afu launched the "Scientific Weight Loss of 100 Million Catties" project, with 184,700 netizens participating, and a total weight loss of 102,600 catties so far.

📈 Middle-aged men are the main force in weight loss, but the older the age, the slower the weight loss, with people over 50 years old losing only 0.38 catties on average.

🌏 Southern people show more significant weight loss effects, with the highest weight loss quantity in the East China region, while the Northwest and Northeast regions have less weight loss.

6. Report: Baidu Kunlun Chip Will Go Public in Hong Kong, Targeting a Valuation of $50 Billion

Baidu's AI chip business entity, Kunlun Chip, is advancing the process of going public in Hong Kong, targeting a valuation of up to $50 billion. As a core component of Baidu's AI hardware ecosystem, this IPO is an important milestone for capitalization and injects new catalysts into the global AI computing power market. Kunlun Chip has successfully launched multiple generations of self-developed AI chips and has achieved large-scale application in cloud-based inference scenarios.

AiBase Summary:

🚀 Baidu Kunlun Chip plans to go public in Hong Kong, targeting a valuation of $50 billion.

🧠 Kunlun Chip is a core part of Baidu's AI hardware ecosystem, driving the development of the global AI computing power market.

💡 Kunlun Chip has launched multiple generations of self-developed AI chips and has achieved widespread application in various computing power scenarios.

7. Key Breakthrough in Computing Power Acceleration: Peking University and DeepSeek Jointly Open Source Large Model Inference Framework DSpark

The article introduces the DSpark framework jointly developed by Peking University and DeepSeek, which solves the issues of response delay and computing waste in large model inference through dual optimization mechanisms, significantly improving generation efficiency and throughput.

AiBase Summary:

🧠 Uses a semi-autoregressive architecture to improve the balance between generation efficiency and quality.

📊 Introduces confidence scheduling verification mechanism to reduce unnecessary calculations.

🚀 Engineering optimization achieves a leap in throughput.

8. Musk Says Monthly Release of New Large Model Grok4.5, Internal Test Performance Comparable to Claude Opus

Musk revealed that his latest large language model, Grok4.5, has started Beta testing within SpaceX and Tesla, and plans to release a new base model every month, further intensifying competition with leading companies such as OpenAI and Google.

AiBase Summary:

🧠 Grok4.5 is built on a 1.5 trillion parameter V9 base model and incorporates Cursor data for additional training.

🚀 Musk states that Grok4.5's performance is already comparable or even exceeds Anthropic's flagship model Claude Opus, and reinforcement learning is still being optimized.

🚀 SpaceX plans to release a completely newly trained base model every month, intensifying competition with companies like OpenAI and Google.