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 technical trends and learn about innovative AI product applications.
New AI products Click to learn more:https://app.aibase.com/zh
1. Luma AI releases Uni-1 image model, which uses a autoregressive architecture to generate text and pixels simultaneously
Luma AI has released the Uni-1 image generation model based on an autoregressive architecture, supporting the simultaneous generation of text and images, and offering various visual styles. The model performs well in benchmark tests, and its API pricing is reasonable, making it suitable for enterprise use.
【AiBase Summary:】
🧠 Autoregressive architecture enables simultaneous generation of text and images
🎨 Supports 76 visual styles, including realistic photography and comics
💰 API charges by token, with lower costs than competitors
2. Meituan Longcat releases open-source mathematical theorem proving model LongCat-Flash-Prover
The Meituan Longcat team has released a deep learning model called LongCat-Flash-Prover, specifically designed for mathematical formalization and theorem proving. Through three capabilities—automatic formalization, draft generation, and proof generation—the model achieves a paradigm shift from probabilistic prediction of answers to rigorous logical proofs, and has achieved excellent results in multiple benchmark tests.
【AiBase Summary:】
🧠 The model achieves a paradigm shift in logical reasoning through automatic formalization, draft generation, and proof generation.
📊 In the MiniF2F-Test benchmark, it achieves a 97.1% pass rate with only 72 inference budgets, setting a new SOTA record for open-source Prover models.
🔒 It uses a "hybrid expert iteration" framework based on TIR, effectively solving logic vulnerabilities and code deception issues.
Details: https://github.com/meituan-longcat/LongCat-Flash-Prover
3. Alibaba DAMO Academy releases Xuantie C950: Single-core performance exceeds 70 points, breaking the global RISC-V record
The Xuantie C950 processor released by Alibaba DAMO Academy has made significant breakthroughs in performance, security, and AI capabilities, marking the strength of the RISC-V architecture in high-performance computing fields.
【AiBase Summary:】
🚀 Single-core performance exceeds 70 points, breaking the global RISC-V record
🧬 Native support for confidential computing security isolation and data protection
🧠 Integrated self-developed AI acceleration engine, supporting billion-parameter large models
4. Successfully "chewed" 40 billion parameters! iPhone 17 Pro challenges running ultra-large models locally, but the speed is only 0.6 tokens per second
The article discusses the technical challenges and significance of the iPhone 17 Pro successfully running a 40-billion parameter large language model, demonstrating the possibility of running ultra-large models locally, although the speed is relatively slow.
【AiBase Summary:】
🧠 Data is streamed via SSD to break memory limits
⚡ Generates one word every two seconds, with slow speed
🔒 Local execution improves privacy protection and offline feasibility
5. Claude Assistant upgrade! Anthropic introduces new computer control features
The upgrade of Claude Assistant marks the transformation of AI from "speaking" to "acting," and its ability to directly control user computers has sparked discussions on security and efficiency.
【AiBase Summary:】
🧠 Claude Assistant upgrade, with the capability to directly control user computers.
🔒 Anthropic emphasizes permission confirmation and active defense mechanisms for security.
🚀 AI agents will fully emerge in 2026, with major players vying for control over the user desktop.
6. West Lake University releases "Titan o1": The world's first robot "general cerebellum" is born
The Titan o1 humanoid robot released by West Lake University is equipped with the world's first action generalization large model GAE system, achieving millisecond-level action replication, action generalization, and remote synchronization control, marking the transition of humanoid robots from toys to productivity tools.
【AiBase Summary:】
🤖 Millisecond replication: Titan o1 can precisely synchronize human actions, enabling one person to control hundreds of robots.
🧠 Action generalization: The GAE system allows robots to have human-like action adaptability without rigid programming.
🚀 Spatiotemporal breakthroughs: Supports remote synchronization control, allowing the coordination of hundreds or even thousands of robots to complete complex collaborative tasks.
7. Search rules changing? OpenAI applies to join Android "selection screen", ChatGPT may become the new default in Chrome
OpenAI submitted an application to the UK Competition and Markets Authority, hoping to include ChatGPT in the list of optional default search engines for Google's Android system and Chrome browser. This move marks the entry of AI-native applications into traditional search entrances and may reshape the mobile search ecosystem.
【AiBase Summary:】
✅ OpenAI applied to make ChatGPT a default search engine for Android and Chrome, challenging traditional search entrances.
💡 OpenAI believes that generative AI platforms should be included in search qualification standards, not limited to traditional engines.
🔄 If approved, it will break Google's long-term dominance over core traffic entrances, promoting a trend toward decentralization in the search market.
8. National Data Administration confirms the official Chinese name of 'Token' as "word unit"
The National Data Administration has officially named the basic unit of large models 'Token' as 'word unit'. This naming ends the debate in academia and industry about the Chinese translation of 'Token', marking further standardization of the national level of AI technology terminology. As the smallest unit for AI to understand human language, the 'word unit' has a granularity between characters and words, allowing more precise coverage of semantic slices when large models process text. With the volume of calls reaching 140 trillion, unified terminology standards will help reduce industry communication costs and lay the foundation for subsequent AI infrastructure statistics and policy formulation.
【AiBase Summary:】
🧠 The 'word unit' is the smallest unit for AI to understand human language, with a granularity between characters and words.
📊 China's AI industry has exceeded 140 trillion word unit calls per day.
🌐 Unified terminology standards help reduce industry communication costs, laying the foundation for subsequent AI infrastructure statistics and policy formulation.
