Articles

This article uncovers the true reasons behind Sequoia Capital CEO Roelof Botha's departure. Botha's exit was not a voluntary handover but rather a forced removal, orchestrated by partners leveraging Sequoia's unique internal governance mechanism, which allows partners to initiate a removal vote with greater weight given to more senior partners. Key reasons included Sequoia repeatedly missing early-stage investment opportunities in hot AI unicorns like OpenAI and Cursor, leading to widespread partner dissatisfaction. Additionally, Botha's mishandling of the Klarna board controversy and the COO's departure, coupled with the ill-timed launch of the Evergreen Fund program, fueled internal conflict and external criticism. Despite Botha generating over $50 billion in returns for LPs during his tenure and investing in successful projects like YouTube and Instagram, internal strife and external pressures ultimately led Sequoia to transition leadership. Alfred Lin and Pat Grady, two prominent partners, subsequently took over, and Sequoia immediately announced the establishment of a new early-stage fund to heavily invest in AI, signaling a major strategic and leadership shift.

The article explores how product managers can improve their abilities and professional acumen. It begins with a case study of a product manager facing challenges and frustration, leading to the core question: how to gain respect and effectively manage oneself in daily work. The author proposes five concrete suggestions: acknowledging the 'outmoded practices' of current business systems and gradually optimizing them; embracing novelties and not being constrained by past experience; exercising appropriate emotional intelligence and avoiding arbitrary criticism of others; learning to leverage existing resources to avoid reinventing the wheel; and engaging in deep, systematic thinking rather than simply measuring work by time. Each suggestion is accompanied by specific case analyses, aiming to correct some misconceptions in the product manager industry, enhance professional influence, and underscore the learning ability, adaptability, communication and collaboration skills, and deep thinking ability that product managers should possess.

Based on a survey and code analysis of various RL training frameworks, the author delves into the pain points and innovations of existing frameworks to address their Agentic RL training needs for VL Multimodal Models. The article first highlights the limitations of current frameworks in adapting to Agentic RL training environments and introduces the advantages of the AReaL framework in custom workflows. Subsequently, it discusses in detail core technical issues and solutions in RL training, including GPU orchestration (synchronous vs. asynchronous), data flow and data structures, decoupled training and inference, model weight update mechanisms, and the gaps between the training engine and inference engine (on-policy degenerating to off-policy, inconsistent token distribution). Finally, the article also touches upon perennial RL training issues such as RL algorithms (sequence-level vs. token-level rewards), reward and environment design, and evaluates the usability and engineering quality of various frameworks, providing readers with a panoramic view for choosing and understanding RL training frameworks.

In intricate micro-frontend and Monorepo architectures, frontend developers frequently encounter the challenge of quickly pinpointing component source code. This article introduces and implements a "Source Code Traceability" solution. First, it thoroughly explains Vite plugin implementation methods and lifecycle hooks (particularly the `transform` stage), laying the groundwork for the core functionality. Subsequently, it elaborates on how to inject encoded absolute file paths into Vue component root elements via custom Vite plugins and propagate this path information to child components through Vue compiler extensions. Finally, by integrating Tampermonkey scripts, browser-side interaction logic is realized, allowing users to click any component on the page to instantly view its hierarchy and corresponding source code path. This solution significantly boosts development efficiency and offers clear guidance on the fundamental use of Vite plugins and Tampermonkey scripts.

This podcast features an interview with Richard Bown, who shares his experiences working as a software engineer for 30 years, including his career path from an individual contributor (IC) to Head of Engineering, and ultimately back to an IC role as a DevOps engineer. He highlights that true leadership is not confined to management positions but can be exercised effectively from any role, emphasizing the importance of self-awareness in career choices. Bown discusses the aspirational nature of "humane engineering organizations," noting that despite advancements in sociotechnical understanding, the ideal often remains elusive due to commercial pressures and the inherent human cost of continuous engineering operations. He also touches upon his novel, "Human Software: A Life in I.T.," explaining how it delves into the emotional toll and work-life balance challenges faced by engineers, while also exploring broader themes like climate change, the impact of AI, and organizational vanity projects. The interview concludes with advice for emerging leaders to seek knowledge, read influential books, and connect with respected mentors.

Against the backdrop of rapidly escalating competition in artificial intelligence, this article focuses on the fierce rivalry between Google and OpenAI. Recently, Google released the reimagined Gemini Deep Research Agent and, for the first time, unveiled its Embedded Research Agent API, emphasizing breakthroughs in long-chain reasoning stability, ultra-large context processing capability, and traceable citations. Almost simultaneously, OpenAI released the highly anticipated GPT-5.2 (codenamed Garlic), which enhanced logical consistency, tool-calling stability, and agent autonomy. The article provides a deep comparison of the 'close combat' between the two companies in foundation model capabilities and agent platform capabilities, pointing out that Google's Deep Research Agent is completely rewritten based on Gemini 3 Pro, incorporating multi-step reinforcement learning into its reasoning chain and providing full-chain citations. This aims to upgrade it from a 'report generation tool' to a 'professional agent capable of executing complete research tasks'. Concurrently, Google introduced the Interactions API, allowing developers granular control over agent behavior. The article also cites discussions from the tech community about this release, particularly the attention to Google's new Agent slightly outperforming GPT-5.2 in benchmarks like the 'Human's Last Exam' (HLE).

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.

This article details the journey of Jenny Shao, a Chinese-American doctor, who resolutely resigned from her surgical residency at Harvard to establish Robyn, an emotional AI application. Witnessing patients die alone and recognizing the healthcare system's prioritization of insurance company interests, she was driven to leverage AI to combat loneliness and foster emotional connection. Robyn successfully secured $5.5 million in seed funding within 18 months of its inception. The article highlights Robyn's distinct product positioning: it functions neither as a friend nor a therapist, but rather as a tool for emotional reflection and self-awareness, developed with strong medical ethics and memory research principles. Robyn emphasizes robust safety mechanisms and boundary design to prevent overstepping, with user feedback validating its value in helping individuals understand themselves and improve interpersonal relationships. The piece also explores the challenges Jenny faces as an atypical entrepreneur and her unique leadership philosophy, concluding with an optimistic outlook on the future of emotional AI, stressing the critical balance between technological advancement and ethical security.

The article uses "10 snippets" to outline the significant achievements of DeepMind, an Artificial Intelligence company. It covers key milestones from its founding vision to AlphaGo's victory over the Go World Champion and AlphaFold's solution to the protein folding problem. The article emphasizes DeepMind's pursuit of Artificial General Intelligence (AGI) and how its groundbreaking research has transformed science and human society. Through these snippets, readers can understand DeepMind's reflections and practices in AI ethics, scientific exploration, and technological application, inspiring imagination about the future development of AI and fostering appreciation for their rigorous approach.

This article details Nanbeige4-3B, a small language model from Boss Zhipin's Nanbeige Large Model Lab. With just 3 billion parameters, it achieves breakthrough performance through extreme data and algorithm optimization. The article explains its innovative Hybrid Data Filtering System and fine-grained Warmup-Stable-Decay (WSD) Scheduler used during pre-training, along with a four-stage progressive post-training process: Cold-Start SFT, Overall SFT, Dual-Level Preference Distillation (DPD), and Phased Reinforcement Learning (RL). Evaluation results demonstrate that Nanbeige4-3B not only significantly outperforms Qwen3-4B and Qwen3-8B (of comparable scale) in tasks like mathematics, scientific reasoning, tool calling, and human preference alignment, but also rivals larger models such as Qwen3-30B-A3B in several high-difficulty benchmarks. The article highlights that Nanbeige4-3B establishes a new 'lightweight flagship' paradigm for small language models. Its Base and Thinking models, along with a technical report, have been open-sourced to promote the democratizing AI.