BestBlogs.dev Highlights Issue #42

This article introduces the latest updates to the Google Gemini API, including the release of Gemini 2.5 Flash and Pro models, which feature enhanced reasoning capabilities and low latency. Notably, Gemini 2.5 employs a "thinking models" approach. Veo 2 is now officially released and supports the generation of high-quality videos from text and images. The Live API has also been updated, adding support for more languages and configurable Voice Activity Detection (VAD), making it suitable for building real-time interactive applications. For example, Wolf Games reduced the number of iterations needed by 60% by using Veo 2. These updates are designed to help developers build more powerful and efficient AI applications.

This article provides an in-depth look at the AI updates from Google Cloud Next 25, including 5 AI models, 1 AI protocol, and 6 other updates. It highlights the Gemini 2.5 Flash inference model, emphasizing its cost-effectiveness and inference capabilities; the Veo 2 video generation model, highlighting its applications in video editing and creation; and the Chirp 3 audio understanding and generation model, focusing on its applications in speech synthesis and transcription. In addition, it introduces the improved image generation and editing capabilities of Imagen 3, as well as the A2A protocol designed to enable seamless collaboration between Agents. Firebase Studio cloud-based AI programming tools, ADK Agent development framework, Google Workspace integrated AI services, and Ironwood TPU AI chips are also mentioned. Google's 601 customer AI case studies offer guidance for AI application developers. This update highlights Google's ongoing AI investment and innovation, potentially shaping the future of AI technology.

Meta has unveiled its latest Llama 4 series AI models, including Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. These models leverage a Mixture of Experts (MoE) architecture, significantly boosting training and inference efficiency. They also boast multimodal capabilities, support multiple languages, and demonstrate excellent performance across various benchmarks. The Llama 4 Scout features 17 billion active parameters and a 10M token ultra-long context window, positioning it as an industry frontrunner and unlocking novel applications in memory, personalization, and multimodality. Llama 4 Maverick, with 17 billion active parameters and 128 experts, surpasses GPT-4o and Gemini 2.0 in several benchmarks. Llama 4 Behemoth, with 288 billion active parameters, stands as one of Meta's most powerful models to date, serving as a teacher model for knowledge distillation in smaller models. Llama 4 Scout and Llama 4 Maverick are available for download on llama.com and Hugging Face, furthering the advancement of open-source AI.

The article introduces Jina AI's new generation multimodal multilingual re-ranker, jina-reranker-m0. This model is based on the Qwen2-VL-2B architecture, with a total of 2.4 billion parameters. It adopts a pairwise comparison approach to simultaneously evaluate the relevance of visual and textual elements in the input document to the query, enabling efficient document ranking. Compared to its predecessor, jina-reranker-m0 not only adds the ability to process visual information, but also further improves performance in pure text re-ranking scenarios, targeting multilingual content, long documents, and code search tasks. The article also highlights the model's leading performance in multimodal benchmarks such as ViDoRe. It further provides API call examples (user-friendly) and Hugging Face usage tutorials (more flexible, supports images as queries).

The article introduces the Kimi team's open-sourced visual language model Kimi-VL and its inference version Kimi-VL-Thinking. This model is based on the MoE architecture, with a total of 16B parameters but activates only 2.8B during inference. It has powerful multimodal reasoning and Agent capability, supporting a 128K context window. The model training used three major categories of datasets, including pre-training data (caption data, image-text interleaved data, etc.), instruction data, and reasoning data. The article showcases Kimi-VL's outstanding performance in visual understanding, reasoning, OCR Character Recognition, and multi-turn Agent interaction tasks through multiple examples, and it surpasses GPT-4o in specific benchmark tests. The article also introduces the technical details of the model, including the model architecture and training process. Finally, the article also mentions that the Kimi team may soon launch the K1.6 model.

This article provides a systematic review of Test-Time Scaling (TTS), which dynamically allocates computing power during the inference phase to maximize the potential of large language models (LLMs). The article proposes a four-dimensional orthogonal analysis framework, including What to scale (e.g., CoT length, number of samples, etc.), How to scale (e.g., Prompt, Search, etc.), Where to scale (e.g., mathematics, code, etc.), and How well to scale (e.g., accuracy, efficiency, etc.). Based on this framework, the article reviews existing literature and summarizes the three major development directions of Inference-Time Scaling technologies. It also provides specific technical selection guidelines and future development directions for typical scenarios. This review offers valuable references for research and application in the field of TTS.

This article is a transcript of Professor Lee Hung-yi's course on AI Agent at NTU, which introduces the concept, operation mode, and how to use Large Language Models (LLM) to build AI Agents with clear explanations and in-depth analysis. The article first clarifies the definition of AI Agent: when humans only give the goal, the AI can autonomously plan and execute multiple steps to achieve it. Then, the article elaborates on the operation process of AI Agent, including observing the environment, analyzing the situation, and taking actions, repeating until the goal is achieved. The article emphasizes that LLM is an important driving force for the current development of AI Agent and analyzes its advantages, such as the vast range of possible actions and no need to manually set rewards. The article also discusses the key role of RAG technology in AI Agent's memory and learning and lists several application examples of AI Agent, including AI villagers, AI controlling computers, and AI training AI models. In addition, the article also emphasizes the importance of AI Agent adjusting its behavior based on experience and introduces the method of using RAG technology to achieve long-term memory. This article provides a solid foundation for understanding AI Agents and their potential impact on various fields.

Cloudflare has released the open beta of AutoRAG, a fully managed Retrieval-Augmented Generation (RAG) pipeline designed to help developers more easily integrate context-aware AI into their applications. AutoRAG eliminates the complexity of building and maintaining RAG pipelines by automatically handling steps such as data ingestion, chunking, embedding, vector storage, semantic retrieval, and response generation. It also offers continuous monitoring of data sources and automatic indexing to ensure the AI knowledge base is up-to-date. AutoRAG is built on Cloudflare's Vectorize database and Workers AI, leveraging Cloudflare's serverless platform to provide developers with high performance and scalability. AutoRAG is currently in the Open Beta phase and is free to use.

This article delves into the application of RAG (Retrieval-Augmented Generation) technology in intelligent question answering systems, especially in the field of cloud services. The article first reviews the development history of large language models (LLMs), and then points out the key role of RAG technology in solving LLM hallucinations and verticalization of domain data. Then, starting from four core propositions: addressing data value challenges, the leap of heterogeneous retrieval, the optimization of generation control, and the reconstruction of the evaluation system, the article elaborates on the technical challenges and solutions faced by RAG technology in detail. The article also introduces the specific methods and experimental results adopted by the author's team in actual business, such as building a hierarchical knowledge graph, optimizing retrieval strategies, and introducing Retrieval-Augmented Relevance (RAR) technology, including the use of agent/user feedback and high-parameter LLM analysis to refine evaluation criteria, design chain-of-thought prompting, dynamic few-sample learning and other steps. Finally, the article looks forward to the future development trends of RAG technology in multi-modal retrieval, deep search, and evaluation optimization. Overall, this article provides a comprehensive and in-depth analysis of RAG technology, providing valuable reference for technical personnel in related fields.

This is a summary of the Prompt Engineering whitepaper officially released by Google. The whitepaper details the concept of Prompt Engineering, LLM output configurations (including output length, sampling controls: temperature, Top-K, and Top-P), and various prompting techniques (such as zero-shot, one-shot, few-shot, system prompt, context prompt, and role prompt). In addition, it explores code prompting and multimodal prompting, and provides best practices: providing examples, concise design, and specifying output formats. The whitepaper has a clear structure, starting with basic concepts (introduction and fundamentals), delving into specific technologies, covering key applications, mentioning future directions, and ending with practical suggestions. The whitepaper aims to help users better understand and apply Prompt Engineering.

Cloudflare aims to lead in the AI Agent ecosystem by strategically enhancing its Agents SDK. The updates include simplified integration with external services through remote MCP clients with built-in authentication, integration with Stytch, Auth0, and WorkOS for authentication and authorization, and McpAgent hibernation for optimized resource utilization. A Durable Objects free tier lowers the barrier to entry, while Workflows GA and AutoRAG facilitate production-ready, context-aware AI applications. These enhancements enable AI Agents to securely connect to external services and perform actions on behalf of users efficiently, benefiting both developers and the broader AI Agent ecosystem.

The article introduces the Agent Development Kit (ADK), an open-source framework released by Google, designed to simplify the full-stack end-to-end development of agents and multi-agent systems. ADK enables developers to build production-ready agentic applications with greater flexibility and precise control by providing capabilities across the agent development lifecycle, such as building, interacting, evaluating, and deploying. Key features of ADK include multi-agent design, a rich ecosystem of models and tools, built-in streaming, flexible orchestration, an integrated developer experience, and easy deployment. Additionally, the article compares ADK with Genkit and highlights ADK's optimized integration with Google Cloud, particularly with Gemini models and Vertex AI. Google encourages developers to use ADK to build the next generation of AI applications.

The article delves into the potential for Agent technology to grow in 2025 and the challenges it faces. First, it explains how advancements in model reasoning capabilities, multimodal model capabilities, and code capabilities are driving Agent technology development, enabling Agents to better understand user needs, process image information, and efficiently generate code. Second, it analyzes the challenges faced by General Agents at the engineering and model levels, including the architecture and design of Agents, evaluating Agent performance, solving memory management in long-duration tasks, and improving the model's instruction following ability, long context ability, reasoning planning, and reflection ability. The article highlights the representative role of Devin and Cursor in promoting Agent development. Finally, the article puts forward the view that General Agents will not be replaced by models and emphasizes the important role of Agents in human-computer collaboration. The article also introduces the author's practical experience in the internal R&D of intelligentization at Alibaba, as well as the speech sharing at the QCon Global Software Development Conference.

This article analyzes OpenAI's Operator and Deep Research Agent products, compares the differences between Agent and Workflow, and combines Anthropic's research to emphasize that the core competitiveness of Agent lies in End-to-End Optimization and versatility, with simplicity over complexity. It also provides practical suggestions for Algorithm Engineers in the Agent development wave, including accumulating test sets and fine-tuning techniques.

The article explores the needs of AI Agents regarding browsers, pointing out the shortcomings of traditional browsers in automated scraping, interaction, and real-time data processing. It analyzes the drawbacks of existing headless browsers, such as performance issues, deployment complexity, and script fragility. It proposes a blueprint for building a new generation of headless browsers using LLM and VLM technologies. This enhances AI's understanding and adaptability to web pages through the conversion of natural language instructions into Playwright code. Additionally, the article discusses market opportunities, emphasizing the growing demand for browser automation tools due to the proliferation of AI Agents. It addresses GTM strategies, potential risks, and the competitive landscape, highlighting the importance of the open-source community and developer experience. Browserbase's StageHand framework, which allows developers to interact with web pages using natural language, exemplifies this approach.

Alibaba Cloud Bailian officially launches the full-lifecycle MCP service, covering service registration, cloud hosting, Agent orchestration, and process orchestration, aiming to solve the pain points of tool integration in AI application development. This service transforms tool orchestration from a model provider's proprietary feature into a universal capability and has a complete product form, which is an important step in Alibaba Cloud's AI commercialization. Developers can use MCP through either the official hosting service or self-managed service. The official hosting service has zero threshold, and you can directly call it in Agent or process by opening and filling in the API key, while the self-managed service is suitable for enterprise developers to integrate internal services with MCP. The difference between MCP and Plugin lies in protocol openness, service deployment method, and invocation paradigm. MCP aims to allow all models to understand the same "service language", the service is hosted by the platform, and supports multi-step scheduling and multi-tool combination. The emergence of MCP transforms AI orchestration of external tools from a cumbersome engineering task into a standardized platform capability, shifting the focus from human developers to AI itself, and the service is designed to be easily understood and used by AI.

The article opens with an internal memo from Shopify's CEO, stating that proficient use of AI is a basic requirement for employees, emphasizing AI's central role in corporate survival and competition, and a resource allocation philosophy that prioritizes AI. Specific measures include incorporating AI usage into performance evaluations. Subsequently, the article elaborates on Vibe coding, an emerging programming method that involves completing coding tasks in collaboration with AI through natural language and Prompt Engineering. It summarizes 12 practical tips for Vibe coding, including communicating requirements to AI, selecting a simple technology stack, providing ample context, and using image examples. Finally, the article shares a case study using Prompts to build an MCP Server using TypeScript, showcasing the application of Vibe coding in real-world development.

Captions is an AI-powered audio and video creation application that rapidly gained users through innovative features like AI-generated 3D avatars, AI editing, AI lip-syncing, automatic subtitles, and eye contact correction, utilizing a subscription-based business model. The article analyzes Captions' strategies, including addressing user pain points, continuous model optimization via a Data Flywheel (a system where data fuels continuous improvement), and a unique 'secret roadmap.' Captions emphasizes solving practical problems and explores AI applications in video creation to reduce production costs and barriers. Their future focus is on character generation models. The article also discusses the challenges and opportunities for AI startups regarding business models and Technical Debt.

This article provides an in-depth review of Midjourney V7 Alpha and compares it with GPT-4o. V7 Alpha has improved in image quality, with new personalization features and a draft mode. The personalization feature generates images based on user aesthetic preferences, while the draft mode improves rendering speed and reduces costs. V7 Alpha excels in style diversity, including realistic, illustration, 3D, and surrealistic styles. However, V7 Alpha struggles to accurately generate images based on prompts. In text rendering, GPT-4o performs the best, with a significant gap in the text generated by V7 Alpha. The draft mode is currently unstable during testing, resulting in a suboptimal user experience. Overall, V7 Alpha has made progress in image generation but still needs improvement in adherence to prompts and text rendering.

The author of this article shares various creative methods and prompts for image generation using GPT-4o, providing a large number of directly usable prompts and prompt ideas. The content covers multiple aspects such as the microscopic world, 3D icon design, journaling style, photo graffiti, game character integration, cartoonized classic movies, and gradient color extraction. Each method provides detailed prompts and effect display images. The author emphasizes that creativity does not arise out of thin air, but through consciously connecting elements from different fields, breaking conventional thinking, and re-observing daily things. AI tools only expand the boundaries of creative realization capabilities, and true creativity stems from human emotions, experiences, and thinking.

The article is an in-depth interview with Rabbit founder Lyu Cheng about RabbitOS Intern. Lyu Cheng emphasizes that Rabbit is not a hardware company, but is committed to building a new operating system based on AI Agents, directly competing with challengers such as Manus, aiming to overcome the limitations of the traditional APP model. RabbitOS Intern is a key step in realizing this vision, using cross-platform general Agents to control the underlying operation of machines with natural language, subverting the existing GUI interface. Lyu Cheng believes that the core of AI Agents lies in reshaping human-computer interaction, giving machines control over planning, reasoning, and execution, achieving more efficient and intelligent task processing. He also firmly believes that there should only be one operating system in the cloud in the future, which can flow into any device. He shared his thoughts on industry trends, competitive landscape, and product pricing, as well as the impact of Raven's experience on this venture, and believes that the real barrier is not technology, but the ability to execute and solve detailed problems, and firmly believes that Rabbit will take a leading position in the AI field.

This article introduces Day.ai, an AI-Native CRM company founded by Christopher O'Donnell, former HubSpot Chief Product Officer, aimed at resolving issues like incomplete data and complex workflows in traditional CRM systems. Day.ai utilizes AI to automatically extract data from emails, meeting records, and other sources, building a comprehensive understanding of customer relationships. Analyzing email content and meeting records, it automatically generates to-do items, prompting timely follow-ups. A conversation between Sequoia Capital partner Pat Grady and Christopher O'Donnell reveals that Day.ai's AI-Native design inherently bypasses the data compression issues of traditional CRMs, providing a CRM experience closer to genuine customer relationships. Day.ai aims to revolutionize sales CRM, similar to how Spotify transformed music consumption, enabling CRM to truly serve sales personnel, enhancing their efficiency and job satisfaction. Day.ai is poised to lead the AI-Native CRM space through ongoing technological innovation and user feedback.

In response to GSR Ventures' Zhu Xiaohu's doubts about the commercialization prospects of humanoid robots, the article surveys 10 leading domestic and foreign companies such as Unitree Robotics, EngineAI Robotics, and UBTECH Robotics, analyzing the common problems faced by the humanoid robot industry in its rapid development, such as difficulties in commercialization, high costs, and limited application scenarios. The article discusses the rationale behind the 'humanoid' form factor and the differences and challenges in industrial and home environment applications, pointing out that humanoid robots need to solve actual needs and technical bottlenecks. The article concludes that although the humanoid robot industry has broad prospects, it still requires long-term exploration and tackling to achieve true maturity and popularization, crossing the divide from factories to homes.

The 2025 AI Index Report from Stanford University offers a comprehensive analysis of AI's current state, covering technological advancements, applications in daily life, business investments, and the evolving global landscape. It highlights progress in complex tasks and video generation, widespread adoption in healthcare and transportation, and the narrowing performance gap between Chinese and American AI models. The report addresses ethical concerns like data bias and misinformation, emphasizing responsible AI governance, and analyzes AI's impact on the economy, education, and public awareness, serving as a valuable reference for decision-makers and the public.

To stay ahead in the AI era, Shopify CEO Tobi Lütke released an internal email emphasizing the integration of AI into the company culture, requiring all employees to learn and apply AI in their daily work, even incorporating it into performance evaluations. The email points out that AI should be a core tool in the prototyping phase of GSD (Get Stuff Done) projects. Additionally, teams need to demonstrate that their goals cannot be achieved through AI before requesting more hiring slots and resources. The email also shares Shopify's practices in the AI field and encourages employees to share their AI usage experiences. This move aims to empower merchants and take a leading position in the AI-driven e-commerce future. Tobi believes AI will revolutionize Shopify's operations. He encourages employees to explore entrepreneurial opportunities in an AI-driven world.

OpenAI CEO Altman responded to questions about the popularity of GPT-4o's Ghibli style and the 'wrapper' business model of AI startups in a recent interview. He believes that technology-driven changes have lowered the barriers to entrepreneurship, and the emergence of AI will solve the huge global gap in software demand. AI Agents will transform the development workflow, enabling developers to generate complete functional code by simply describing requirements in natural language. Altman emphasized that AI is more of an empowering tool than a complete replacement for humans, and predicts that AI will bring disruptive breakthroughs in programming and AI agent fields. He also suggested that practitioners actively embrace AI technology, adapt to the new ways of working in the AI era, prioritize environments that can access cutting-edge technologies, and make active adoption of AI the primary criterion for evaluating employers. Altman believes that AI is empowering humans in a more mature way, enhancing creativity, addressing specific societal challenges, and reshaping our lives.

This podcast interview delves into Duozhuayu founder Maozhu's eight-year entrepreneurial journey, covering AI applications and insights in the C2B2C secondhand book trading platform, such as Vector Models in the search system and Alibaba models in clothing classification. Maozhu shares her experiences facing market changes, financing challenges, and personal life changes, along with deep insights into business, team leadership, user value, and social impact. She emphasizes the importance of personality, interests, and social insights in entrepreneurship, and reflects on balancing corporate development and personal value realization, and the shift from perfectionism to embracing imperfection. The discussion also explores the digitalization of Chinese content, the value of out-of-print books, and the importance of curation in the era of information overload, reflecting Duozhuayu's value in knowledge dissemination and social culture. Finally, Maozhu shares her experiences in entrepreneurship, financing, corporate governance, and personal growth, providing valuable experiences and insights for entrepreneurs.

This report provides an in-depth analysis of the US AI startup ecosystem. By mining data from 2443 AI startups and 802 investors, it reveals the financing characteristics, industry distribution, geographical pattern, and investor preferences of early-stage AI startups in the US. It provides Chinese AI entrepreneurs with a clear navigation map of the AI capital market, helping them understand the development trends and investment opportunities in the US AI field, to better formulate their startup and financing strategies. The report points out that US AI startups generally adopt a “small and fast” financing strategy, To B Enterprise Applications and AI Infrastructure Layer technology stacks are the mainstream directions, and Silicon Valley dominates geographically. In addition, the report also analyzes the characteristics of investors behind star projects, including emerging funds, angel investors, and CVCs.

This edition of Last Week in AI covers the latest developments in the AI field. Meta launched the Llama 4 series, which uses a Mixture of Experts (MoE) architecture, but it has faced criticism for underperforming. Amazon introduced Nova Act, an AI Agent capable of controlling a web browser. Adobe Premiere Pro also released AI-powered video extension features. Additionally, OpenAI's ChatGPT saw significant growth in paid users and revenue. Other news includes the release of Runway's Gen-4 video generation model and updates to Microsoft Copilot and Google AI products. In business, Nvidia's H20 chips are in demand by Chinese tech giants, and there is substantial investment in AI drug discovery and video generation. Research highlights DeepMind's AGI safety approaches, and studies suggest LLMs can pass the Turing test. On the policy front, the UN warns of AI potentially widening the digital divide, and publishers are urging governments to stop AI copyright theft by requiring AI companies to pay for content used in training.

This issue of the deeplearning.ai Batch discusses the impact of U.S. tariff policies on global trade and the development of artificial intelligence. The author argues that tariffs are generally detrimental to AI development but may provide some impetus for the domestic robotics and automation industries. Research from Anthropic reveals that large language models (LLMs) can perform reasoning even without explicit training, and it presents a method for examining their internal reasoning processes. Meta's release of the Llama 4 series models features ultra-long context windows, and Alibaba's Qwen2.5-Omni 7B further demonstrates the potential of open-source models in multimodal tasks.