BestBlogs.dev Highlights Issue #69

This article introduces DeepSeek-OCR, a new open-source model from the DeepSeek team. It's not just another OCR tool; it's a revolutionary approach to long text context processing. Traditional large language models face challenges with quadratically increasing computational complexity when handling extensive texts. DeepSeek-OCR tackles this by 'compressing' text content into a two-dimensional image and encoding it into visual Tokens. This significantly reduces Token consumption within the context window, achieving up to a 10x compression ratio while maintaining high recognition accuracy. The article elucidates the collaborative mechanism between DeepSeek-OCR's DeepEncoder and DeepSeek-3B Decoder using AI assistant chat logs. Furthermore, the model draws inspiration from human memory decay and visual perception, implementing a 'digital forgetting curve' that mirrors the gradual fading of information over time, thus offering a fresh perspective on AI memory management.

This YC-sourced article provides a comprehensive overview of the foundational Transformer architecture, which powers contemporary AI models like ChatGPT and Gemini. It meticulously details three crucial advancements: Long Short-Term Memory (LSTMs), Sequence-to-Sequence (Seq2Seq) with Attention, and the Transformer itself. The narrative begins by highlighting the limitations of early neural networks, such as feedforward networks and Recurrent Neural Networks (RNNs) struggling with sequential data and the 'vanishing gradients' problem. LSTMs are then introduced as a significant breakthrough from the 1990s, using 'gates' to manage information flow and enable learning long-term dependencies, which became viable with GPU acceleration in the 2010s. Despite their success, LSTMs faced a 'fixed-length bottleneck' in Seq2Seq tasks. The article then describes the advent of Attention mechanisms in 2014, which allowed decoders to 'attend to' relevant parts of the encoder's hidden states, dramatically improving machine translation. The climax is the 2017 paper 'Attention Is All You Need,' which introduced the Transformer, an architecture that entirely eschewed recurrence for parallel processing via self-attention, leading to superior speed and accuracy. The discussion concludes with the rise of Transformer variants like BERT and GPT, highlighting how their scalability led to the transition from single-task models to the general-purpose intelligent large language models we utilize today.

The article delves into the emerging technology paradigm of Deep Research Agent, first defining its core capabilities and boundaries, and distinguishing it from general-purpose models and single-purpose research tools. Next, it elaborates on the core technical architecture of Deep Research Agent, including the evolution and trade-offs between static and dynamic workflows (single-agent and multi-agent). The article also focuses on how agents utilize tools such as web search, code interpreter, and multi-modal processing, and explores optimization methods such as prompt engineering, supervised fine-tuning, reinforcement learning, and non-parametric continual learning. Finally, by analyzing closed-source projects from OpenAI, Google, etc., and open-source projects such as A.deep research and DeerFlow, it distills crucial insights for building Agent frameworks and points out future challenges in evaluation benchmarks and information source expansion.

The article compiles the latest interview with Jerry Tworek, OpenAI GPT-5 developer and VP of Research. The interview emphasizes that the realization of Artificial General Intelligence (AGI) is inseparable from the deep integration of Pre-training and Reinforcement Learning (RL); both are essential. Tworek explains reasoning as a process of finding unknown answers, enabled in language models via Chain of Thought (CoT). He reviewed the progress of OpenAI models (o1, o3, GPT-5) in reasoning ability, regarding GPT-5 as a technical iteration of o3. The article also explains the basic principles of Reinforcement Learning in a simple way, comparing it to training pets, and guiding model behavior through reward and punishment mechanisms. Tworek highlights the interactive nature of RL environments and the role of RLHF in GPT-4's success, contrasting modern and traditional RL approaches. Tworek envisions the future of Agentic AI and outlines OpenAI's strategies in research, collaboration, and rapid deployment.

The article compiles a speech by former OpenAI core researcher Jason Wei at Stanford University, proposing three core ideas for understanding AI development in 2025. First, Commoditization of Intelligence . Once AI capabilities are mastered, their costs will approach zero. This makes knowledge acquisition instant and personalized. As a result, it democratizes the field and enhances the value of private information. Second, Verifier's Law states that AI's ability to solve tasks is directly proportional to the verifiability of the task. Any task that can be solved and is easy to verify will eventually be conquered by AI, offering insights into measurement and automation. Finally, the Serrated Edge of Intelligence emphasizes that AI development is not a linear 'rapid takeoff' but exhibits uneven progress across different tasks. AI accelerates fastest in digital, human-friendly, data-rich tasks with clear metrics.

The article delves into the importance, methods, and challenges of evaluating Generative Artificial Intelligence capabilities. First, it emphasizes the crucial role of evaluation for model users and developers in identifying the best models and optimizing development processes. Next, it details various evaluation methods, including exact matching and similarity calculations based on standard answers (such as BLEU, ROUGE, BERTScore), as well as human evaluation and the use of Large Language Models (LLM) as referees when standard answers are unavailable. The article highlights various pitfalls and biases that may be encountered during evaluation, such as Goodhart's Law effect resulting from over-reliance on evaluation scores, model hallucination issues and their evaluation mechanism defects, the subjectivity and superficial biases in human evaluation, and the self-favoring, positional bias, and verbosity bias in LLM evaluation. In addition, the article comprehensively discusses the need to go beyond content quality when evaluating, and to comprehensively consider important factors in practical applications such as generation speed, operating costs, computing resources, model robustness (against jailbreak and prompt injection attacks), data contamination, and model bias. Finally, it emphasizes that evaluation methods should be selected based on specific application scenarios, and that critical thinking should be maintained, with a clear understanding of their limitations.

The article addresses developers' common confusion about AI Agents and illustrates their evolution from Chatbots to advanced intelligent entities through the analogy of 'a top student's academic journey', covering core concepts like chain of thought, self-reflection, planning, and tool use. The central argument is that the qualitative change in AI Agent capabilities stems from the 'cognitive process' designed around Large Language Models (LLMs), rather than the intelligence of the LLMs themselves. It explores the threefold value of this process: using 'structure' to scaffold thinking, using 'iteration' to create compression algorithms for memory, and using 'interaction' to connect the model to the real world. The article also explains the effectiveness of the Agent loop from the perspectives of Cybernetics and Information Theory. Ultimately, it suggests developers evolve from 'Prompt Engineers' to 'Agent Process Architects', focusing on designing thinking processes, empowering action tools, and building decision-making context. It further explores the evolution of Agent Performance Engineering and future Cognitive Architectures.

This discussion features Anthropic's Alex Albert and Erik Schluntz, who delve into the evolution and best practices for building effective AI agents. They explain how Claude is trained for agent tasks, emphasizing the role of coding in making agents more autonomous. The conversation highlights the utility of the Claude Agent SDK for developers, allowing them to integrate custom business logic and tools without rebuilding core agent loops. A key innovation discussed is 'Agent Skills,' which extends claude.md files to include reusable resources like templates, code, and assets, significantly enhancing an agent's capabilities. The experts differentiate between 'agent workflows' (sequential agents) and 'multi-agent systems' (parallel or orchestrated agents), detailing patterns like parallelization and MapReduce. They also address common failure modes, such as over-engineering and communication overhead in multi-agent setups. Best practices include starting simple, understanding the agent's perspective, and designing model-facing tools that map to UI concepts rather than raw APIs. The future of agents is envisioned with greater 'computer use' capabilities, enabling self-verification and autonomous interaction with applications like Google Docs, ultimately reducing the need for human QA.

This article starts from the core concept and clearly explains the fundamental differences between AI Agents and traditional Workflows in handling complex, dynamic, and long-tail problems, emphasizing the unique advantages of Agents in autonomous thinking, action, and retrospection. Through a detailed intelligent customer service case, the article vividly demonstrates how the Agent Framework addresses complex scenarios such as multi-intent recognition, cross-system verification, policy reasoning, and negotiation, effectively addressing the limitations of Workflows in handling 'branching complexity' and dynamic dialogue decision-making. Subsequently, the article provides an in-depth introduction to five mainstream Agent Frameworks: AutoGPT, LangGraph, Dify, CrewAI, and AutoGen, including their core features, typical application scenarios, and their respective advantages and disadvantages, supplemented by practical operation examples. Finally, the article summarizes the value of Agent as a new way of thinking, that is, to enable intelligent systems to move from 'executing commands' to 'understanding goals,' and highlights Tencent Cloud TDAI team's exploration of Agent memory capabilities. This aims to lay the foundation for AI transformation.

As a translation of Chapter 5 of 《智能体设计模式》, this article systematically introduces the core concepts, a six-step process, typical applications, and various frameworks for the 'Tool Use Pattern'. It highlights that this pattern allows agents to interact with external systems, interfaces, and services through function calling, enabling them to obtain real-time information, perform calculations, operate databases, and even control devices, thereby transforming large language models into active agents. The article details six major applications, including obtaining external information, interacting with databases, and executing code. It also provides detailed code examples based on LangChain, CrewAI, and Google ADK, covering use cases such as simulated search, stock query, code execution, and enterprise search, greatly enhancing its practicality. The article emphasizes that tool usage is key to building powerful and interactive AI agents.

This article discusses the shift from Prompt Engineering to Context Engineering in the rapidly evolving field of AI Agents. It defines Context Engineering as building dynamic systems that provide LLMs with the correct information and tools in the appropriate format, listing its seven core components and contrasting it with Prompt Engineering. The article emphasizes Context Engineering's value in reducing AI failure rates, ensuring consistency, supporting complex features, and enabling self-correction. It analyzes the 'Context-Rot' problem caused by long contexts and its solutions, detailing LangChain's four context management methods: Offload, Retrieve, Compress, and Isolate. Furthermore, it showcases industry applications of Context Engineering through Claude Code's three-layer memory architecture, real-time Steering, hierarchical multi-Agent collaboration, and dynamic context injection, as well as Manus's practices in KV Cache optimization, tool masking, and file system memory. Finally, the article introduces Spec-Driven Development, illustrating its implementation through the Kiro project, and envisions Context Engineering evolving towards Environment Engineering, emphasizing the two-way interaction between AI and the environment.

The article delves into the core mental models for developers to efficiently conduct AI Pair Programming in the AI Era. It introduces Andrej Karpathy's “VibeCoding” concept, highlighting the paradigm shift from “Architect” to “Product Sculptor” and revealing two major efficiency dilemmas: “Communication Breakdown” and “Code Quality Assurance Concerns.” To address these, the article proposes strategies in “Communication,” “Code Quality,” and “Context Management.” Leveraging the Johari Window Model, it emphasizes actively switching between 'Teacher' and 'Student' roles, employing the Feynman Learning Technique to 'teach' AI and Socratic Questioning to 'question' AI, building consensus and fostering in-depth thinking. It advocates Test-Driven Development (TDD) and Minimum Viable Validation for task decomposition, ensuring code quality through progressive submission and rollback. Finally, it distinguishes between Prompt Engineering and Context Engineering, stressing the importance of actively managing AI's “Memory” to enhance interaction quality. The article combines theory with practice, providing practical guidance for AI collaboration.

OpenAI has launched ChatGPT Atlas, a new AI-driven web browser built around ChatGPT, aiming to redefine web interaction. Sam Altman emphasizes the opportunity to innovate beyond traditional browser models, making the chat experience central. Key features include 'chat anywhere across the web,' allowing ChatGPT to understand web page context and offer assistance, and 'browser memory,' which personalizes the browsing experience over time. Additionally, Atlas introduces innovative interaction models such as a 'multi-turn search experience' for conversational query refinement and 'cursor chat' for in-line editing of text directly within web pages. The most advanced feature is 'agent mode,' where ChatGPT autonomously performs tasks by interacting with web elements, as demonstrated by managing Google Docs tasks and ordering groceries via Instacart. OpenAI stresses the importance of safety and user control, ensuring the agent operates within the user's tab, cannot execute local code, and offers explicit controls for access and actions. Atlas is initially available for macOS, with agent mode for Plus and Pro users, and plans for Windows and mobile expansion. The team envisions Atlas as a 'vibe lifing' tool, delegating various personal and professional tasks to the AI agent.

The article details the Claude Skills feature launched by Anthropic, which allows users to load professional prompt packages and executable code packages on demand, thereby expanding the capabilities of the Claude model. This feature boasts composability (allowing Claude to automatically recognize and combine multiple skills), portability (usable across Claude Apps, API, and Claude Code), efficiency (loading only the minimum necessary information), and the ability to include executable code. The structure of Skills includes core instructions (SKILL.md), script files (such as Python/Bash), and resource files. Claude intelligently scans and loads relevant skills based on task requirements. The article also describes how Skills are used on different platforms (Claude Apps, API, Claude Code), and specifically mentions the Code Execution Tool as the underlying secure sandbox environment for Skills. In addition, Anthropic provides a tool called 'skill-creator' to help users create new skills through dialogue, simplifying the development process. Finally, the article discusses the significance of Skills (modularization of professional knowledge, reusability, team sharing) and its limitations (security, creation complexity, maintenance costs), emphasizing the importance of using skills from reputable sources.

The article details how AI video generation company HeyGen grew its ARR from $1 million to $100 million in 29 months. The core philosophy is to 'embrace uncertainty,' treating the instability of AI technology as an advantage by quickly iterating, releasing, and learning. The handbook covers HeyGen's specific methodologies in core philosophies, iteration rhythm (two-month roadmap, daily releases), principles (speed, learning, innovation), teamwork (PM, engineer, designer, data scientist responsibilities), product and growth team division, communication methods, and 'pitfalls' to avoid. HeyGen emphasizes building products around unchanging user pain points and capitalizing on model advancements, aiming to create products that automatically improve with AI upgrades.

The article provides an in-depth review of Vidu Q2's newly launched multi-image reference video generation feature, pointing out that it brings a new workflow paradigm to the field of AI Video Generation and is expected to replace the traditional 'Image-to-Video' model. Through a large number of practical examples, the author details the significant progress of Vidu Q2 in three core aspects: First, significantly improved consistency , even in multiple subjects or complex scenes, it can stably maintain the characteristics of people, objects, and scenes; Second, significantly stronger performance ability , whether it is the delicate emotional expression of real actors or the rich actions and expressions of anime characters, Vidu Q2 can accurately present them, and even understand common expression techniques in anime; Third, significantly better multi-style expressiveness , it can generate videos of various animation styles and maintain extremely high style consistency. The article also mentions Vidu Q2's user experience optimizations, such as a more convenient reference image citation method and a subject library function, as well as its relatively economical cost. Overall, the author believes that the upgrade of Vidu Q2 is a victory for the 'Advocates of Multi-Image Referencing,' indicating that AI Video Generation technology is moving towards a new stage.

This fireside chat features Anthropic CEO Dario Amodei and Eli Lilly CDO Diogo Rau, exploring strategies for deploying enterprise-grade AI in highly regulated industries like life sciences. Amodei outlines Anthropic's distinct approach, prioritizing accuracy and reliability over consumer-driven engagement metrics, to avoid "model sychophancy" and deliver "truth" for business-critical applications. The discussion highlights the critical role of specialized AI "skills" and dedicated models, such as a life sciences-focused Claude integrated with domain-specific databases (proteins, compounds, assays), to enhance utility. Rau shares Eli Lilly's perspective on leveraging models like Claude for clinical research and drug development. Amodei's parting advice stresses the importance of ambitious, end-to-end AI automation, urging enterprises to anticipate rapid technological advancements rather than focusing solely on incremental process optimizations, to avoid delaying patient benefits.

This article, an interview with Figma CEO Dylan Field, delves into the convergence of design, product, and engineering amidst the AI revolution. Field chronicles Figma's evolution from a browser-based design tool to a comprehensive multi-product ecosystem, including FigJam, Dev Mode, and AI-driven features like Figma Make. He highlights how AI is fundamentally altering product interfaces, making interactions ubiquitous across various platforms and contexts. A core tenet discussed is the power of community and direct user feedback, which has been crucial in shaping Figma's product roadmap and identifying new use cases, exemplified by FigJam's organic development. Field asserts that in an AI-accelerated world, "good enough" design is no longer acceptable; exceptional craft, a strong point of view, and deep empathy for user experience are paramount for product differentiation. He shares insights on his personal balance between leadership and hands-on building, including 'jailbreaking' AI models. The discussion also covers Figma's strategy for expanding beyond its core platform through integrations with other tools like ChatGPT, Notion, and Jira connectors, and the development of new product offerings like Figma Slides and Dev Mode, all driven by observed user behavior and a vision to bridge imagination and reality. Field also touches upon the complexities of AI pricing models, emphasizing the need for utility and aligned incentives. The interview concludes by looking ahead to a future where AI interactions move beyond simple prompts to more intuitive, use-case-specific interfaces, underscoring the continuous need for thoughtful design in this rapidly evolving landscape.

In this in-depth interview, Andrej Karpathy, a leading AI expert, challenges the notion of AGI being imminent, asserting it's still a decade away. He highlights critical cognitive deficits in current Large Language Models (LLMs), such as the lack of continual learning, insufficient multimodal capabilities, and poor computer interaction, making them unsuitable for complex tasks. Karpathy sharply criticizes Reinforcement Learning (RL) as 'terrible' due to its sparse, noisy supervision, which he vividly describes as 'sucking supervision through a straw,' contrasting it with human learning and reflection processes. He explains why LLMs struggle with novel, intellectually intensive coding tasks, often misinterpreting custom implementations and defaulting to boilerplate. Furthermore, Karpathy delves into the fundamental differences between human learning (driven by evolution and a 'cognitive core' of algorithms) and LLM training, which he believes is overly reliant on memorizing 'slop' from the internet. He also addresses the critical problem of 'model collapse' in synthetic data generation, where LLMs lack the diversity and entropy of human thought, hindering effective self-improvement. The discussion touches on the 'autonomy slider' in software engineering, the challenges of autonomous driving, AGI's gradual economic impact (e.g., integrating into 2% GDP growth), and his vision for future education, including his Eureka project which aims to improve learning through building. Ultimately, he emphasizes that AI progress is a continuous, multi-faceted improvement across data, hardware, and algorithms, rather than singular breakthroughs.

This episode of the "Crossroads" podcast, together with guest Zhuang Minghao, reviews the AI and technology industry in 2025, defining it as a year marking an inflection point, and deeply explores the limits that technology, products, and capital may reach. The conversation first analyzes the Large Language Model battlefield, from the low-cost yet high-impact approach of DeepSeek R1 to Sam Altman's redefinition of AGI subtly surpassing previous expectations, and the differences between China and the United States in technical routes and open source strategies. Then, it delves into the fierce competition and clear commercialization path of multimodal technologies (such as Sora 2, World Model), and analyzes OpenAI's distinctive productization strategies. The podcast also focuses on the Agent year, discussing the complexity of L3-stage Agents shifting from language to behavior, the coexistence of general and vertical Agents, and looking forward to the new opportunities brought by on-device Agents under the Harmony HMAF Framework for developers. Finally, it discusses the open source ecosystem as China's strategic asset for AI development and its commercial potential, and conducts an in-depth analysis of the current AI field's capital market frenzy, valuation bubbles, and shifts in investment logic. Overall, the podcast emphasizes that the AI field has both technical and commercial challenges, but is still full of transformations and prospects.

The article compiles insights from Y Combinator partners on how to thrive in the increasingly competitive AI landscape. The YC team points out that the AI 'Gold Rush' window is closing, and obvious startup ideas are saturated. The key to success lies in discovering 'non-obvious, even dangerous' secrets and taking contrarian bets. Using examples such as Uber, Coinbase, Flock Safety, OpenAI, and SpaceX, the article illustrates the importance of challenging legal gray areas, disrupting traditional business models, and pursuing 'Sci-fi' inspired visions. It emphasizes that founders should return to First Principles, focusing on genuine user needs rather than external noise or popular playbooks, to build formidable and sustainable companies.

This article documents a fireside chat with Lilian Weng, former VP of Research at OpenAI and co-founder of Thinking Machines. She recounts the humility and persistence she learned from mathematics competitions, and the 'reality distortion field' effect she experienced while working on the robot Rubik's Cube project in OpenAI's early days. She details her experiences with GPT-3 productization and the AI safety team, and explains her decision to leave OpenAI due to a 'flattening learning curve' to establish Thinking Machines with like-minded individuals. The new company is committed to 'human-centered AI' and 'open science,' aiming to provide researchers with flexible infrastructure through the Tinker API. Lilian Weng also shares her learning methods, prioritization skills, and the importance of maintaining technical grounding (through code reviews) and providing constructive feedback as a leader, offering valuable career and personal growth advice for AI practitioners.

This podcast delves into the entrepreneurial journey of Duxiang founder Wang Dengke, from early SaaS to AI companion products. He shares his experience of obtaining seed funding when he was still inexperienced in the era of innovation and entrepreneurship, followed by seven years of being constrained, and the turning point brought about by the Large Language Models wave (LLM), which gave him the opportunity to create ToC products that impact more people. The podcast details the development concepts and challenges of his AI painting tool '6pen,' the gamified product 'Hong Hong Simulator,' and the AI companion app 'Duxiang.' Wang Dengke emphasizes that AI products should go beyond mere efficiency improvements and focus on providing users with emotional value and in-depth experiences, requiring an artistic touch. He points out that building a deep connection between humans and AI faces challenges such as model capabilities and limited interaction methods, and shares innovative attempts such as asynchronous interaction and training AI digital avatars with personal data. The recently launched AI hardware 'Xiang Meng Ring' aims to extend the AI connection to the real world, enhancing the feeling of ceremony and emotional attachment between users and AI through multiple touchpoints. The program also explores how entrepreneurs balance commercial ambition and personal sentiments, and the possibility of AI products meeting users' deep emotional needs and creating unique worldviews in the future, providing listeners with rich entrepreneurial, product, and AI insights.

This article, based on a podcast with Nicole Forsgren, a leading expert in developer productivity and experience (DevEx), delves into the complexities of measuring engineering performance in the age of AI. Forsgren, known for the DORA and SPACE frameworks, explains why traditional metrics like Lines of Code are misleading with AI-generated content and introduces 'trust' as a critical new dimension for evaluating AI's output. She emphasizes that while AI accelerates coding, overall developer speed is often bottlenecked by broken builds, unreliable tools, and increased code review time. The discussion highlights the importance of DevEx, defined by flow state, cognitive load, and feedback loops, as a foundational element for innovation and engineer well-being. Forsgren provides practical advice, including conducting 'listening tours' to identify process friction and advocating for a product mindset in DevEx improvements. Her upcoming book 'Frictionless' outlines a seven-step process for organizations to remove barriers, unlock value, and leverage AI effectively, urging companies to align DevEx metrics with leadership's strategic priorities like market share or profit margins, rather than generic productivity scores. The conversation underscores that improving DevEx yields significant business value, from faster time-to-market to reduced costs, despite the initial 'J-curve' of implementation.