Articles

This article delves into the underlying memory system of ChatGPT, pointing out that it does not employ RAG (Retrieval-Augmented Generation) technology, but rather uses clever system prompts to manage user data. The article divides ChatGPT's memory system into four major components: interaction metadata (including device information and usage habits), recent conversation content (storing only user messages), model setting context (information explicitly told by the user, with the highest priority), and user knowledge memory (AI-generated condensed summaries, capturing user patterns). The author compares these four components to the training process of LLMs, emphasizing that OpenAI relies on more powerful models and larger context windows to handle memory, rather than complex engineering techniques. The article also discusses the challenges of real-time refreshing of user knowledge memory in the future, and mentions the memory strategies of other AI vendors.

The article delves into key AI industry issues via an interview with OpenAI Chairman Bret Taylor. Taylor notes the 'performative' nature of many AI applications and a significant AI bubble, while affirming AI's long-term economic potential. He advocates for solution-centric AI companies, rather than AGI pursuit or self-developed models. Furthermore, he anticipates reduced fine-tuning importance due to increasing context windows and improved rule adherence. Taylor is particularly optimistic about AI Agents disrupting customer service, revolutionizing digital interaction with voice. He also introduces Sierra's pay-for-results model, shares insights on GPT-5's advancements, the evolving AGI definition, and perspectives on 'super intelligence' and safety.

The article delves into the contrasting philosophies behind the memory system design of AI assistant Claude and ChatGPT. ChatGPT tends towards automation, comprehensively recording user interactions to build user profiles, catering to individual users; while Claude adheres to the principles of statelessness and direct data access, granting users high control and focusing more on developer tools and professional workflows. The article details Claude's two memory tools: conversation retrieval and short-term dialogue retrieval. In addition, the article also reports that shortly after its reverse engineering analysis was published, Claude launched a new memory function for teams and enterprise users, including project-based memory segmentation, memory summaries, incognito chat, and enterprise management permissions, further strengthening its positioning in enterprise collaboration solutions. This reflects the vast diversity of AI memory design space and the profound impact of product strategy on technological implementation paths.

This article provides an in-depth review of OpenAI's latest 'semi-released' model, GPT-5-Codex, a version of GPT-5 fine-tuned for AI-assisted programming. It introduces the model's core innovation: dynamically allocated 'thinking time,' which significantly enhances performance in agentic programming and code refactoring. The review details the tools integrated into GPT-5-Codex, including VS Code plugins, Codex CLI, Codex Cloud, and GitHub integration. It highlights six key improvements: training for code review, dynamic adjustment of thinking time, enhanced code refactoring, mobile website optimization, reduced irrelevant comments, and shorter system prompts. The article also mentions how to use Codex CLI. A real-world test, referencing developer Theo Browne's video review, confirms that Codex addresses the GPT-5 issue of 'fast Token consumption' by intelligently allocating Tokens based on task complexity and performing well in code review. However, the article also identifies shortcomings, including a poor search function, an immature UI, disappointing cloud testing, and a tendency to generate incorrect or nonsensical outputs in certain situations. Finally, the article discusses OpenAI's challenges in programming ecosystem integration, noting that its toolchain has 'all the parts, but easily falls apart when put together,' emphasizing ecosystem coherence as a major issue.

This article covers Rust core contributor Nicholas Nethercote's layoff from Futurewei due to budget cuts (partially due to AI's resource drain) and his subsequent job search. This event sparked debate about Rust's job prospects and AI's impact on traditional tech. Nicholas clarified Rust's continued relevance, citing its use in operating systems, compilers, and GPU programming. He joined VectorWare, a startup improving GPU programming with Rust, showcasing the need for technical talent to remain adaptable and learn new technologies, heralding a 'GPU-native software era'.

Academician Wang Jian's speech at the Bund Conference in Shanghai explored the future of Artificial Intelligence. He stated that open source has evolved from code openness to the availability of data and computing resources. He also posited that OpenAI's closed source approach is a historical choice rather than a strategic error, emphasizing open resources as key to AI industry advancement. Furthermore, he revealed for the first time Zhejiang Lab's successful deployment of 8B AI models on 12 satellites, creating a 'Three-Body Computing Constellation' that enables AI model operation and satellite interconnection in space. Academician Wang Jian envisions sending computing satellites to the L5 point in solar orbit in the coming years, highlighting AI and computing power's crucial role in enabling humanity to venture beyond Earth, thus providing computing support for deep space exploration.