Articles

This article evaluates Zhipu AutoGLM 2.0, which aims to create a standardized cloud environment to avoid real-world complexity. Through analyzing cloud phone and computer configurations and testing tasks like shopping, flight booking, and PPT production, the article reveals the Agent's limitations when facing complex login verifications, frequent ad pop-ups, mechanical understanding of multi-conditional instructions, and file storage and uploading issues in domestic applications. While Zhipu's 'standardization' approach solves some compatibility issues, the dynamic complexity of the real world remains a major obstacle for Agents to become general-purpose assistants.

This article provides an in-depth analysis of the diverging strategies adopted by OpenAI and DeepSeek following the release of GPT-5. It observes that while GPT-5 demonstrated top-tier performance in benchmarks, it failed to deliver a paradigm shift or fundamental technological advancements. Persistent issues like hallucinations and multimodal understanding suggest that large language models may be approaching the inherent constraints of the Transformer architecture. Consequently, OpenAI has opted to prioritize the refinement and productization of existing capabilities, focusing on building a comprehensive 'super app' platform to enhance user experience and commercial value. Conversely, DeepSeek, under Liang Wenfeng's leadership, is leveraging the current plateau in large model technology to pursue an 'independent' strategy: training state-of-the-art models on domestic chips to reduce reliance on NVIDIA GPUs, despite the considerable engineering hurdles involved. Simultaneously, DeepSeek is committed to driving fundamental technological innovation and open-source accessibility through research contributions like 'Native Sparse Attention,' underscoring its dedication to pushing technological boundaries and fostering an open ecosystem. The article posits that these two companies exemplify distinct trajectories for the future of the AI industry.

The article provides a detailed review of the Looki L1 wearable AI life-logging camera, positioning it as a 'Personal Life Recorder'. Through personal experience, the author introduces the core functions of Looki L1, including Story Mode for all-day automatic recording, For You for automatic Vlog generation, Chat AI for AI-powered memory retrieval, and Lifelog video schedule. The article emphasizes the difference between Looki L1 and traditional action cameras, namely its AI-powered footage processing and organization, which greatly simplifies user operation. At the same time, the article points out shortcomings in hardware design (such as charging method, button logic) and user experience (such as storage reminders). After discussing the privacy challenges faced by such devices, the article proposes potential solutions such as AI-assisted blurring of pedestrians' faces. Finally, the article regards Looki L1 as a real-world data acquisition carrier needed for future 'World Models' and Multimodal AI models, and envisions its important role in AI development.

The article uses the release of DeepSeek V3.1 as a starting point, deeply analyzing the "UE8M0 FP8" parameter precision mentioned in the comments. It first explains the importance of floating-point numbers (FP32, FP16, FP8) in large model training and the advantages of FP8 in terms of memory usage and training speed. It then points out that NVIDIA's FP8 standard is not fully open, and its optimization is deeply tied to hardware. The article further explains that DeepSeek's adoption of the 'UE8M0 FP8' range-priority variant format is to adapt to the incompatibility of domestic chips with NVIDIA in the underlying design, thereby ensuring that large models can run stably on domestic hardware. Finally, the article lists domestic chip manufacturers such as 沐曦 (MuXi) and 燧原科技 (Enflame) and their FP8-supporting chips, emphasizing that this marks China's AI industry entering a critical stage of software and hardware collaborative development to reduce dependence on foreign computing power.