Articles

This article deeply analyzes the multifaceted challenges Meta faces internally, following Zuckerberg's strategic pivot from the Metaverse entirely to Artificial Intelligence. The core conflict lies between the newly formed elite AI team, TBD Lab, and the original core business units: TBD Lab is dedicated to developing "god-like artificial superintelligence," prefers a closed-source approach, and refuses to use social data to train models. In contrast, Meta's long-term executives hope AI will serve the core social media and advertising businesses, favoring an open-source strategy. This ideological clash embodies Meta's internal "civil war." To support its expensive AI ambitions, Meta has significantly cut the budget for its perennially loss-making Metaverse division (Reality Labs) and may proceed with layoffs. By reviewing historical cases like Microsoft missing the mobile era and Yahoo's "Peanut Butter Dilemma," the article warns of the strategic risks Meta currently faces and ponders whether Meta can avoid repeating the fate of Google+ or replicate the miracle of Google Gemini. Although Meta officially denies internal conflicts, the shifting resources and ideological differences continue to fuel tensions. This situation will determine whether Meta can control its own destiny in the AI era.

The article explores the pervasive issue of ableist design, arguing that it stems from societal systems that inherently exclude disabled people. It defines ableism and internalised ableism, illustrating how systemic norms, often driven by capitalism and perfectionism, shape designers' approaches and lead to a lack of accessibility. The author debunks common misconceptions about disability prevalence and the perceived conflict between aesthetics and accessibility, citing real-world examples like Figma Sites and Awwwards to highlight how design tools and evaluation criteria often deprioritise accessibility. The article then shifts to practical solutions, encouraging designers to ask more questions, accept temporary answers, challenge assumptions with design fiction, and conduct inclusive user research. It stresses normalising the needs of disabled people and adopting an anti-ableist stance beyond mere compliance, advocating for truly welcoming and respectful digital spaces.

This article announces the release of `huggingface_hub` library v1.0, marking five years of its maturity and stability. This version introduces several major updates, including the adoption of `httpx` as the new backend request library, a redesigned `hf` command-line tool (replacing `huggingface-cli`), and the complete migration of file transfers to `hf_xet`, entirely phasing out the old `hf_transfer` tool. The article reviews `huggingface_hub`'s evolution from initially providing model sharing capabilities for the `transformers` library to its current status as a core component supporting over 2 million models, 500,000 datasets, and 1 million Hugging Face Spaces. It also highlights the integration of Model Context Protocol (MCP) and Tiny-Agents to simplify AI Agent development. The new version focuses on modernizing the HTTP architecture, enhancing CLI tools, and clearing technical debt, providing detailed migration guides to address the challenges posed by rapid AI development.

The article deeply analyzes ChatGPT's memory system through reverse engineering. It highlights that instead of employing complex traditional Retrieval Augmented Generation (RAG) methods for its massive user base, ChatGPT manages user memory via an ingenious four-layer context stack, achieving high efficiency and resource savings. This system incorporates: temporary session metadata (environmental information), a permanent personal profile (key points the user asked to remember), a summary of recent interests (based on chat titles rather than full text), and a sliding window of the current conversation. This layered design enables ChatGPT to provide a personalized and coherent conversational experience for up to 800 million users efficiently. It achieves this by avoiding the high latency and token costs typically associated with RAG, thus delivering an excellent user experience. The article details how these four layers work in conjunction, optimizing for personalization needs, system performance, and computational cost.

The article addresses the unique challenges of debugging "deep agents," which are more complex and long-running than simple LLM applications. It highlights how deep agents involve longer prompts, multi-step traces, and multiple turns of human interaction, making traditional debugging difficult. LangSmith provides tracing tools to capture execution data (runs, traces, and threads) for visibility into agent behavior. Two new solutions are introduced: Polly, an AI assistant integrated into LangSmith's UI, which analyzes trace and thread data to help identify inefficiencies, mistakes, and suggest prompt improvements; and LangSmith Fetch, a CLI tool that allows developers to access LangSmith traces and threads in their local IDEs or with other coding agents for bulk analysis and dataset creation. These tools aim to simplify the process of debugging, understanding, and refining deep agents, thereby enhancing their performance and reliability.

This article details Rokid's outstanding performance and key achievements in the smart glasses industry in 2025. Facing explosive growth in the smart glasses market, Rokid has meticulously refined its core product capabilities, distilling features like photography, translation, teleprompting, navigation, and payment into compact, lightweight glasses. It has also innovated with magnetic charging, optimized a dual teleprompter mode, addressed light leakage issues, upgraded AI quick commands, and enhanced the payment experience. Furthermore, Rokid has paid close attention to the special needs of users with myopia and the visually impaired community. Its products have achieved skyrocketing sales in both domestic and international markets, setting multiple records. Rokid has also built a robust ecosystem through collaborations with platforms like Alipay and Amap. Beyond its success in the consumer market, Rokid's products have been adopted in professional, high-intensity scenarios by China Central Television (CCTV), the Haier Group CEO, and others, transitioning smart glasses from 'experimental devices' to 'trusted devices' and cementing Rokid's role as an industry leader.

This article provides an in-depth exploration of the crucial role of Heterogeneous Treatment Effect Analysis (HTE) in A/B experiment data analysis. First, it clarifies the concept of heterogeneity with intuitive examples and defines Average Treatment Effect (ATE), Conditional Average Treatment Effect (CATE), and Individual Treatment Effect (ITE). Subsequently, it emphasizes the profound business implications of heterogeneity analysis, including uncovering business logic, identifying optimal user sub-populations, and providing dynamic support for optimal user segments. The article also outlines dimension selection principles (T⊥X) and offers an overview of common HTEA methods. Finally, using JD Retail's practices as a case study, it demonstrates how their team-developed Python toolkit facilitates CATE drill-down exploration, presenting a real experimental scenario. This case illustrates that even when overall metrics are not statistically significant, HTEA can uncover sub-populations with positive or negative significant effects, thereby guiding product managers toward targeted optimization.

The article deeply explores the disruptive impact of AI-assisted programming (Vibe Coding) on the software development paradigm. As a programming teacher with 25 years of teaching experience, the author expresses concerns about Vibe Coding diminishing the joy of programming, but also acknowledges the powerful capabilities and feasibility of AI-generated code. His core argument is: traditional programming languages are designed for "human" readability and convenience, but when AI becomes the primary code generator, are these features still necessary? The author questions why, if AI can generate complex, functional systems, AI shouldn't directly output C language or even x86 assembly, which are more machine-friendly. The article further proposes that the future may require a "Vibe-Oriented Programming Language" (VOPL), which could be a "pseudocode" language, implement "literate programming," or be close to a natural language programming paradigm. Finally, the author likens the current debate over AI programming to historical technological revolutions, pointing out that a new computational paradigm is emerging.

This article details the INTELLECT-3 Mixture-of-Experts model released by Prime Intellect. This 106B parameter model was fine-tuned and trained using reinforcement learning based on GLM 4.5 Air. INTELLECT-3 achieved the best performance for models of its scale in various benchmarks, including mathematics, code, science, and reasoning. The article highlights its core technologies, including the self-developed distributed RL framework PRIME-RL, the unified RL environment and evaluation ecosystem Environments Hub, the high-performance code execution system Prime Sandboxes, and the computing power scheduling solution implemented on 512 H200 GPUs. Prime Intellect has open-sourced the model weights, training framework, datasets, and evaluation system, aiming to lower the threshold for post-training advanced models and promote community research and application of large-scale reinforcement learning. The article concludes by outlining future efforts in expanding agentic RL, enriching RL environments, and developing long-term agents.

The author showcases a peculiar scenario where AI (Claude Code) controls another AI (AutoGLM), which in turn operates a mobile phone. Initially, Claude Code was employed to deploy AutoGLM, and remarkably, Claude Code independently resolved all deployment challenges. Subsequently, the author instructed Claude Code to summarize these deployment issues, utilizing the deployed AutoGLM to dispatch these summaries to the developers. This synergy leverages Claude Code's 'thinking' and 'scheduling' prowess alongside AutoGLM's 'visual perception' and 'action' capabilities, forging a formidable mobile Agent. The article specifically highlights that this solution differs from ByteDance's 'Doubao Phone,' which is a system-level integration. In contrast, the AutoGLM-Claude Code combination represents a maverick solution, operating via debugging interfaces.