BestBlogs.dev Highlights Issue #66

Anthropic has launched Claude Sonnet 4.5, positioned as the world's leading model for coding, complex agent development, and computer interaction, demonstrating substantial improvements in reasoning and math. This release is coupled with significant product enhancements, including checkpoints and a VS Code extension for Claude Code, and advanced context management for the Claude API. A key offering is the Claude Agent SDK, providing developers access to Anthropic's foundational agent-building infrastructure. The article highlights Sonnet 4.5's state-of-the-art performance on the SWE-bench Verified for coding and a leading 61.4% on OSWorld for real-world computer tasks, alongside dramatically better domain-specific knowledge across finance, law, medicine, and STEM. Anthropic also emphasizes Sonnet 4.5 as its most aligned frontier model, with reduced misaligned behaviors and enhanced defenses against prompt injection, operating under AI Safety Level 3 protections. The model is immediately available via API and apps at existing Sonnet 4 pricing, with a temporary research preview called "Imagine with Claude" demonstrating on-the-fly software generation.

OpenAI has launched Sora 2, its next-generation video and audio generation system, alongside the new Sora app. Sora 2 demonstrates significant advancements in handling physical interactions, generating longer and more coherent narratives, and integrating audio directly with video. A highlight is the 'Cameo' feature, enabling users to place themselves or others into AI-generated scenes. The Sora app offers a social media-like interface designed to foster a new form of video-based communication, emphasizing user-generated AI content. The platform incorporates robust safety measures, including identity verification for Cameo, content moderation, and visible watermarking using C2PA standards to clearly label AI-generated content. OpenAI plans to roll out APIs and creator tools, expanding the model's utility beyond the app, initially launching the iOS app in the US and Canada via an invitation system.

DeepSeek officially released the experimental model DeepSeek-V3.2-Exp, whose core innovation is the introduction of the DeepSeek Sparse Attention (DSA) mechanism. This mechanism significantly improves the training and inference speed in long text scenarios without substantially affecting the model's output. The article points out that V3.2-Exp performs comparably to the previous generation V3.1-Terminus on public evaluation datasets. To promote the democratization of technology and community development, DeepSeek simultaneously reduced API prices by more than 50%, and open-sourced the V3.2-Exp model, related research papers, and innovative TileLang and CUDA GPU kernels. We encourage users to compare the new model and provide feedback to further verify its performance in real-world application scenarios.

This article announces the launch of Zhipu's latest flagship code model, GLM-4.6, emphasizing its comprehensive advancement in coding proficiency, a 27% improvement over GLM-4.5. GLM-4.6 excels in advanced coding, long context processing (200K), reasoning, search, and writing capabilities, performing strongly in public benchmarks and real-world programming tasks. Its performance is comparable to Claude Sonnet 4 on certain leaderboards, establishing it as a leading domestic code model. Furthermore, GLM-4.6 achieves over 30% reduction in average token consumption compared to its predecessor and pioneers FP8+Int4 mixed-precision quantization deployment on domestic chips like Cambricon and Moore Threads, paving the way for on-device execution of large models on domestic hardware. Zhipu has also upgraded the GLM Coding Plan, offering personal and enterprise versions, and has released select test data for industry validation.

This article announces the general availability of Google's Gemini 2.5 Flash Image model for production environments, highlighting its state-of-the-art capabilities in image generation and editing. Key new features include support for 10 diverse aspect ratios, facilitating content creation across various formats from cinematic to social media, and the ability to specify image-only output. The model empowers users to seamlessly blend multiple images, maintain consistent characters for richer storytelling, perform targeted edits using natural language, and leverage Gemini's extensive world knowledge. It is accessible through the Gemini API, Google AI Studio, and Vertex AI for enterprise use. The article showcases real-world applications from companies like Cartwheel and Volley, demonstrating the model's effectiveness in providing unparalleled character control, maintaining poses, and delivering low-latency, aesthetically guided image generation for live applications. It also provides developers with resources such as documentation, a cookbook, examples of AI-powered apps built with the model, pricing details, and a Python code sample to encourage immediate adoption.

This video course is the third lecture of 'Generative AI and Machine Learning Introduction 2025,' focusing on deconstructing the internal workings of LLMs. The article details the entire process from how an input sentence undergoes tokenization (词元化), embedding table lookups, to processing through multi-layer Transformers (self-attention, feedforward networks), and finally generating the next token's probability through the LM Head (Language Model Head) and Softmax. It particularly emphasizes the concept of 'unembedding,' where the LM Head (Language Model Head) reuses the embedding table, enabling the model to find the representation closest to the target token's embedding during prediction. The article also explores the semantic meaning of the model's intermediate layer representations, how 'representation engineering' can modify model behavior, and introduces advanced tools like Logic Lens and Patch Scope for visualizing and analyzing changes in the model's understanding of context, handling of polysemous words, and internal 'thinking.' Finally, the theoretical explanation is verified through a practical dissection of the parameter structures of Llama and Gemma models, demonstrating the visualization of token similarity, representation changes, and attention weights.

Anthropic has rolled out significant enhancements to Claude Code, aiming to boost its autonomy and efficiency for complex development tasks. Key updates include a new native VS Code extension, offering a richer, graphical interface with real-time diffs directly within the IDE, complementing the refreshed terminal interface with improved status visibility and searchable prompt history. The article also introduces the Claude Agent SDK (formerly Claude Code SDK), providing developers with core tools to build custom AI agents, now supporting subagents and hooks for greater customization. A crucial addition is the checkpointing feature, which automatically saves code states before each change, enabling users to confidently delegate long-running tasks and revert to previous versions if needed. This, combined with subagents for parallel development, hooks for automated actions, and background tasks for non-blocking processes, collectively empowers Claude Code to handle sophisticated refactoring and feature exploration more reliably. These advancements are powered by Claude Sonnet 4.5, making Claude Code a more capable and integrated AI development assistant.

This video features Anthropic's Alex Albert, Brad Abrams, and Katelyn Lesse, who delve into the evolution of building AI agents using Claude. They introduce the rebranded Claude Developer Platform, which now comprehensively integrates APIs, SDKs, documentation, and a suite of advanced tools like Prompt Caching, Batch API, Web Search, Web Fetch, and code execution. A central theme is the concept of "unhobbling the model," advocating for empowering models with necessary tools to autonomously decide actions, execute tasks, process results, and iterate, rather than constraining them with excessive scaffolding. This approach ensures agents can fully leverage increasing model intelligence. The Claude Code SDK, originally developed for coding, is highlighted as a robust, general-purpose agentic framework, ideal for rapid prototyping due to its built-in automation for tool calling and loops. The discussion also covers best practices for identifying valuable agentic use cases, effective context management (e.g., automatic removal of stale tool call records with "tombstones"), and the implementation of agentic memory. Looking ahead, Anthropic plans to introduce higher-level abstractions, enhanced observability for long-running tasks, and a visionary concept of providing Claude with a "persistent computer" to unlock even greater autonomous problem-solving capabilities.

This presentation explores the practical integration of AI functionalities into Java applications, moving from experimental stages to production. It covers essential AI patterns such as providing short-term and long-term memory for Large Language Models (LLMs), integrating LLMs with APIs via Tool Calling to build advanced agents, and implementing Guardrails to protect AI workflows from prompt injection and sensitive data leakage. The talk also delves into Multimodality, demonstrated through image interaction, and leverages advanced retrieval techniques (RAG) to augment LLM context with custom data. Through live demonstrations using Spring AI and Vaadin, the speakers offer immediate practical insights for developers, applicable whether workloads run locally or in the cloud. Key topics include configuring system prompts, streaming responses, sentiment analysis with structured output, and the Model Concept Protocol (MCP) for tool publication.

This article provides an in-depth analysis of the application and practice of AI pair programming in the entire software development process. Based on the author's experience transitioning from a traditional IDE to Cursor, the author introduces in detail the PDCA (Plan-Do-Check-Act) cycle methodology and uses it as the core strategy for AI collaboration. The article distinguishes between human-computer collaboration modes based on development scenarios (production delivery, rapid verification, experimental exploration), and provides specific practical solutions such as prompt engineering and context management (through Cursor Rules and Memories), aiming to improve the certainty and efficiency of AI-assisted development. Finally, the article looks forward to the evolution of AI-driven business development from experience-driven to knowledge engineering, and to a new paradigm of human and AI Agent matrix collaborative production, calling on developers to strengthen system thinking, architectural capabilities and innovation, and continue to invest in knowledge engineering construction.

This article analyzes in detail Alibaba Tongyi Lab's open-source Tongyi DeepResearch Web Agent project. It positions DeepResearch as an open-source, high-performance Web Agent and analyzes its core components, including the model, inference code, and evaluation code. Next, it elaborates on DeepResearch's three-stage training process: Continual Pre-Training (CPT), Supervised Fine-Tuning (SFT), and Reinforcement Learning (RL), while also delving into data synthesis strategies like WebFrontier and the crucial role of the IterResearch pattern in long-cycle Agent tasks. Finally, the article explores the reference value of DeepResearch's design for AI Agent field research and offers more instructive technical insights and practical directions for technical practitioners from different backgrounds, based on established and emerging concepts.

This interview features OpenAI's Chief Scientist Jakub Pachocki and Chief Research Officer Mark Chen, offering deep insights into GPT-5's strategic goals and its emphasis on integrating advanced reasoning for broader user accessibility. They elaborate on how OpenAI measures research progress beyond conventional, often saturated, benchmarks, instead focusing on models' abilities to make novel discoveries in challenging domains like mathematics and programming competitions. The discussion projects a future where "automated researchers" are capable of generating economically impactful new ideas, highlighting the importance of extending reasoning capabilities over longer timeframes and ensuring consistency in agent systems. The conversation also touches upon the surprising and sustained success of reinforcement learning when combined with sophisticated language models, the transformative impact of AI coding tools like CodeX (leading to concepts like "Vibe Coding"), and the essential qualities of top-tier researchers. Furthermore, they outline strategies for fostering a mission-driven, resilient research culture that effectively balances foundational breakthroughs with product development, underscoring the critical role of compute resources in advancing frontier AI.

This article details the new "OK Computer" Agent mode launched by Kimi LLM, now in limited beta testing. It aims to significantly improve AI intelligence through enhanced reasoning, the utilization of over 20 tools, and native Agent capabilities enabling autonomous task execution. Kimi Agent can independently plan and execute the entire process from requirement analysis to product solutions, interaction design, and front-end development and deployment, delivering multi-page website prototypes, mobile Web applications, and high-quality presentations. Through three specific examples—pet website development, financial data visualization analysis, and movie-themed PPT production—the article demonstrates the powerful capabilities of "OK Computer" in handling complex tasks. It transforms into roles such as project manager, product manager, designer, and front-end engineer, automatically sourcing materials, writing code, and deploying websites. This mode trains the K2 model using end-to-end reinforcement learning for LLM training, enabling it to master tools like file systems, browsers, terminals, code, and image/audio generation to handle various task scenarios and flexibly address emergencies. The article notes that "OK Computer" will help users become full-stack knowledge workers, but due to intensive computing power demands, the beta test range will be gradually expanded.

This article presents a comprehensive evaluation and comparison of ByteDance's newly launched Doubao Vision Large Model Seed-1.6-vision. The article first introduces the four image processing tools built into Seed-1.6-vision: POINT, GROUNDING, ZOOM, and ROTATE. Subsequently, through multiple sets of real-world cases, such as blurry license plate recognition, ship location positioning, rotated text recognition, and finding differences in images, Seed-1.6-vision is compared in detail with GPT-5. The test results show that the Doubao model performs excellently in image detail processing, multi-step reasoning, and accuracy, especially with a clear advantage in using image tools to assist analysis. The article highlights how these integrated tools lower the adoption threshold for users and increase the model's practical utility.

This article is an insightful interview with Dylan Field, CEO of Figma, exploring the profound impact of AI on the design landscape. Field articulates Figma's strategic evolution, particularly through initiatives like Figma Make, which aims to streamline the entire design-to-app creation process. He introduces a three-tiered model of design composability: high-level design abstraction (Figma Make), granular component editing (Figma Design), and foundational design systems. The discussion highlights Figma's transition into a 'Figma to Code' company, leveraging AI to automate design implementation and handoff, exemplified by their MCP server. Field shares his journey to becoming 'AI-pilled' and offers a forward-looking perspective on natural language interfaces, predicting their eventual evolution into more intuitive, multi-modal interaction methods for exploring 'Latent Space.' A core theme is the paradox that as AI accelerates software creation, human elements like taste, craft, and distinct design vision become paramount differentiators. The interview also addresses the 'blank canvas' challenge, positioning Figma as a tool to lower the barrier to entry for creative endeavors, making design more accessible and efficient for a broad audience.

This article provides an in-depth discussion of OpenAI's latest ChatGPT Pulse feature and its impact on the AI industry. Pulse is considered OpenAI's second data flywheel after Memory, aiming to shift Large Language Model interaction from passive Q&A to proactive push, thereby significantly lowering the barrier to user adoption and making it a potentially widely adopted AI application. The article gathers opinions from multiple industry experts, analyzing how Pulse enhances user context understanding and increases Daily Active Users through extreme personalization, as well as its profound impact on computing power demand, recommendation system evolution, and the industry's competitive landscape. The discussion also touches on key challenges and opportunities such as data ownership, user privacy protection, and the development of on-device AI.

Liam Ottley, an experienced AI entrepreneur, debunks common misconceptions about AI side hustles by ranking 18 different options. He evaluates each based on five key criteria: skill acquisition, income potential, client acquisition difficulty, long-term viability, and beginner-friendliness. The analysis reveals that most online AI side hustles are time-wasters, often requiring significant upfront investment or offering low returns. Ottley strongly recommends two 'S-tier' options: building an AI Automation Agency (AAA) and 'vibe coding' websites for local businesses using AI tools. These are highlighted for their significant skill development, high income potential, manageable client acquisition, and strong long-term growth prospects, which can lead to genuine financial freedom. The video aims to guide aspiring entrepreneurs towards genuinely valuable AI-driven business opportunities and away from ineffective ventures.

This article, based on an interview with Nesrine Changuel (ex-Spotify, Google), delves into the strategic importance of 'product delight' beyond mere feature additions. Changuel defines delight as creating products that satisfy both functional and emotional user needs, specifically combining 'joy and surprise.' She outlines three core pillars for achieving this: removing friction (e.g., Uber's easy refund), anticipating needs (e.g., Revolut's eSIM), and exceeding expectations (e.g., Edge's automatic coupons). The core of her approach is a practical four-step 'Delight Model': 1) identifying user functional and emotional motives, 2) transforming these motives into product opportunities, 3) using a 'Delight Grid' to categorize solutions (surface, low, deep delight), and 4) validating ideas with a 'Delight Checklist' (including inclusivity and familiarity). The discussion also covers applying delight to B2B products ('Business to Human'), a '50-40-10' rule for prioritizing delight features, fostering a delight culture, and guarding against the 'habituation effect' and instances where delight backfires (e.g., Apple Reactions). Real-world examples from Spotify, Chrome, and Google Meet illustrate the framework's application, highlighting delight's role in user retention and team motivation.

The article provides an in-depth report on NVIDIA CEO Jensen Huang's interview on the BG2 podcast, which took place after a series of high-profile activities, including Huang's announcement of an investment partnership with Intel and plans to invest in OpenAI's 'Stargate.' Huang denied the AI bubble theory, pointing out that the market severely underestimates the exponential growth potential of AI's cognitive inference, and believes that NVIDIA has transformed from a simple GPU company into an AI infrastructure company. He emphasized that by 'Extreme Co-design' — simultaneously optimizing the entire system, including CPU, GPU, network chips, etc. — combined with extreme supply chain scale, performance improvements far exceeding Moore's Law can be achieved, thereby reducing Token costs and maintaining a competitive advantage. Huang also responded to the doubts about investing in OpenAI's 'Recurring Revenue,' saying it is a strategic investment in a future trillion-dollar hyperscale company. In addition, he discussed ASIC competition, the importance of Sovereign AI, the necessity of the United States attracting top talent, and looked forward to the future of AI merging with Mechanical and Electrical Engineering and biology.

The interview features John Collison of Stripe and Des Traynor, co-founder of Intercom, delving into Intercom's significant transformation into an AI-first customer service company. Traynor explains the rapid development and success of Fin, Intercom's AI customer service agent, which now resolves over a million queries weekly with a 65% resolution rate, generating substantial revenue. He highlights the challenges of selling AI products in a crowded market, emphasizing the need for concrete results and guarantees over superficial marketing. The discussion also covers the importance of deep, differentiated AI technology, focusing on solving real customer problems, and the pitfalls of over-reliance on data without direct user feedback. Traynor shares insights into Intercom's shift to usage-based AI pricing, which, despite initial risks, proved successful by aligning cost with value delivered. He also offers advice to startups on aligning funding with market potential, focusing on core competencies, and avoiding premature expansion. The conversation underscores the accelerated pace of growth for successful AI companies and the unique dynamics of co-founder relationships in navigating such transformations.

This article presents an exclusive interview with Tencent Senior Executive Vice President Dowson Tong, reviewing Tencent's strategic adjustments in AI over the past six months. It highlights Yuanbao's transfer from TEG to CSIG, marking a shift from technology-driven to product-driven, under Tong's leadership. Yuanbao has moved to a multi-model strategy, integrating DeepSeek. He discusses Tencent's view of AI Chatbots as a new consumer entry point, the DeepSeek integration, the relationship between Large Language Models and search, the rise of Agents, and WeChat's support for Yuanbao. He also shares Tencent's insights on AI talent, To B Agent implementation, organizational management, and emphasizes the company's proactive embrace of the AI era.

The article features an in-depth interview with Wu Minghui, founder of Mininglamp Technology. Drawing on his mathematics and AI background from Peking University, Wu Minghui presents a unique technological idealism, discussing the essence and future trends of AI entrepreneurship, and Mininglamp's strategic transformation. Wu Minghui emphasizes that AI founders need to be personally involved in model development and uses "publishing papers" as a simple criterion to judge whether they remain relevant in the field. He believes that open-source large models such as DeepSeek will drive companies to fully transition to the "Agent" model, greatly reducing the barrier to Agent development, enabling companies to build professional skills services based on this. Mininglamp Technology exemplifies this shift, transitioning from data analysis to proprietary models in vertical fields, centered on 'data-driven trusted productivity.' Wu Minghui advocates for an idealistic innovation path of "finding nails with a hammer," believing that great success often stems from a belief in the potential of technology rather than pre-set scenarios. The article also discusses the importance of data as a moat in the AI era, and the impact of AI on organizational forms and production relationships. Currently, Mininglamp is pursuing an IPO in Hong Kong to further invest in AI R&D and build leading proprietary model capabilities in specialized areas.

The article features an interview with Elias Torres, co-founder of Drift and Agency, who highlights a core paradox: businesses want AI but reject it due to an expectation of 'perfect' accuracy, a legacy from pre-AI technologies. Torres shares his entrepreneurial journey from humble beginnings in Nicaragua, through building successful companies like HubSpot and Drift, which provided insights into the limitations of human-centric scaling for customer experience. This led him to found Agency, aiming to create truly AI-led customer experiences by automating the backend of business, enabling unprecedented scale and efficiency. He criticizes current AI tools for merely assisting humans rather than fully automating tasks and envisions a future where AI empowers humans to focus on higher-level decision-making, leading to a lean, product-focused company structure. Torres also touches upon the high costs of LLMs and his goal to build a billion-dollar company with fewer than 100 employees by leveraging AI for core operations.

This video features an in-depth discussion between Maria Sha, Wes Roth, and Dylan, exploring the evolving landscape of programming in the AI era. Key themes include the enduring importance of programming skills despite the rise of Large Language Models (LLMs), Python's suitability for beginners, and the critical role of emotional intelligence in tech. They delve into AI's profound influence on future job markets, the limitations of traditional academic AI education, and the necessity of lifelong learning. The conversation also covers philosophical aspects like AI consciousness, privacy concerns with data ownership, the potential for personalized AI, and the benefits of open-source AI development. The participants emphasize AI's potential to enhance productivity across various industries and advocate for greater transparency in AI development and a re-evaluation of copyright laws in the context of AI training data, concluding with an optimistic outlook on AI's future for human potential.