BestBlogs.dev Highlights Issue #43

OpenAI has released the GPT-4.1 series of models, including GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano. These models are available to all developers via API calls, with performance comprehensively outperforming GPT-4o, especially with significant improvements in programming and instruction following. GPT-4.1 supports a context window of up to 1 million token, and improves long context understanding, while innovating in long context reasoning, such as the OpenAI-MRCR and Graphwalks datasets. In various benchmark tests, GPT-4.1 demonstrates superior performance in programming, instruction following, and long context understanding. GPT-4.1 mini achieves a significant leap in small model performance, while GPT-4.1 nano is OpenAI's fastest and lowest-cost model currently. OpenAI has also reduced the price of the GPT-4.1 series and increased instant cache discounts. The GPT-4.1 series of models are also very powerful in image understanding, with GPT-4.1 mini frequently outperforming GPT-4o in image benchmark tests.

Google has released an early preview of Gemini 2.5 Flash, accessible through Google AI Studio and Vertex AI. Building upon 2.0 Flash, this version significantly upgrades reasoning capabilities while maintaining speed and cost-efficiency. Gemini 2.5 Flash is the first hybrid reasoning model, allowing developers to enable or disable 'thinking' and set a thinking budget to balance quality, cost, and latency. It demonstrates strong performance on complex tasks and offers fine-grained control over reasoning. The article showcases the model's reasoning performance across tasks of varying complexity and provides API examples and documentation links for experimentation.

The article introduces Tongyi Wanxiang's latest open-source Start-and-End Frame Video Generation Model, based on Wan2.1 Text-to-Video 14B. It generates 5-second 720p High-Definition videos from start and end frames, showcasing smooth transitions and excellent detail, realistic action, and prompt adherence across various scenarios. The article also covers the model's architecture with advanced AI and Semantic Feature Technology for temporal and spatial consistency, optimization strategies like Data Parallelism and Model Partitioning for High-Definition video generation, and the use of DiffSynth-Studio for Model Inference with reduced GPU Memory requirements.

ByteDance's Doubao team has released the new Seed-Thinking-v1.5 Inference Model, featuring 200B total parameters and a MoE architecture that activates 20B parameters per iteration. Seed-Thinking-v1.5 demonstrates superior performance in benchmarks such as AIME 2024, Codeforces, and GPQA, even surpassing the 671B parameter DeepSeek-R1. To achieve this, the model incorporates optimizations in data construction, reinforcement learning frameworks, and infrastructure. These include the creation of the BeyondAIME mathematics benchmark, the introduction of VAPO and DAPO reinforcement learning frameworks, and the development of a streaming inference architecture. For efficient large-scale training, the model utilizes various parallel strategies, dynamic workload balancing, and memory optimization techniques.

Zhipu has open-sourced the GLM series 32B and 9B base, inference, and rumination models under the MIT License Agreement, allowing free commercial use. Among them, the inference model GLM-Z1-32B-0414 has performance comparable to DeepSeek-R1, with an inference speed of up to 200 tokens/second, and a price of only 1/30 of DeepSeek-R1. At the same time, Zhipu has launched a new domain name Z.ai, integrating three types of GLM models as the interactive portal for the latest models. The base model GLM-4-32B-0414 has 32 billion parameters and excels at code generation and Artifacts generation. Representing Zhipu's exploration of AGI, the rumination model GLM-Z1-Rumination-32B-0414 solves complex problems through in-depth thinking and integration of search tools. The base and inference models have also been launched on the Zhipu MaaS open platform, providing API services.

This article delves into the role of Long Chain of Thought (Long CoT) in reasoning Large Language Models (LLMs). First, the article compares the essential differences between Long Chain of Thought and Short Chain of Thought, proposes a new classification framework for reasoning paradigms, and emphasizes the advantages of Long Chain of Thought in terms of depth, breadth, and refinement. Second, the article analyzes in detail the six core reasoning phenomena of Long Chain of Thought, such as reasoning boundary, overthinking, and aha moment, and discusses their impact on model reasoning efficiency and answer quality. Next, the article comprehensively organizes the current mainstream optimization strategies for Long Chain of Thought, including key technologies such as reinforcement learning and retrieval-augmented generation (RAG). Finally, the article highlights future development directions of Long Chain of Thought, including multi-modal reasoning, cross-lingual reasoning, agent interaction, efficiency optimization, knowledge enhancement, and security assurance. This review aims to provide a unified perspective for the research of Long Chain of Thought, promote its further development in theory and practice, and play an important role in promoting the development of artificial intelligence.

This article details how to implement various Prompt Engineering techniques using the Spring AI framework for Java developers. It begins by explaining LLM configuration, including selecting providers like OpenAI and Anthropic, and adjusting generation parameters such as temperature and maxTokens. The article then demonstrates Zero-Shot, Few-Shot, System, Role, and Contextual Prompting with Java code examples. Spring AI's advantages include its ease of configuration and its ability to map LLM responses directly to Java objects using the entity() method, facilitating structured data processing. Aimed at Java developers, this guide showcases how to leverage Spring AI for efficient Prompt Engineering.

The article delves into the prevalent length bias issue in text vector models, where longer text vectors tend to receive higher similarity scores, even if the content is not truly relevant. Through experiments using Jina AI's jina-embeddings-v3 Model and the CISI Dataset, the author demonstrates the impact of length bias on cosine similarity threshold settings and explains the reasons for the bias: longer texts usually contain more information points, causing their vectors to be more spread out in the semantic space. The article also discusses mitigation methods such as asymmetric encoding and proposes hybrid solutions combining re-rankers and large language models to more accurately assess relevance. Finally, the author emphasizes the importance of understanding model limitations, focusing on real-world applications, and leveraging strengths and mitigating weaknesses.

The article systematically explains the concepts and evolution of Tool, MCP (Model Context Protocol), and Agent in the AI field clearly and concisely. It uses the 'brain in a vat' analogy to highlight the initial limitations of LLMs in text processing. It then introduces how 'function calling' or 'tool use' empowers LLMs to interact with external systems. Subsequently, it highlights Anthropic's MCP protocol, which standardizes how models interact with tools, addressing issues of redundancy and reusability. Building upon this, the article discusses the rise of Agents, which leverage LLMs and Tools to enable more intelligent and efficient AI tool utilization. Finally, the article forecasts the Agent ecosystem's future, suggesting that Vertical Agents offer near-term implementability and practical benefits. It predicts that 2025 will be a pivotal year for Agents, marked by significant technological advancements and commercial prospects. The advantages and potential of Vertical Agents are particularly emphasized.

Elasticsearch 9.0 and 8.18 are officially released, with key highlights including: BBQ (Better Binary Quantization) vector quantization technology is now GA, offering significant improvements in query speed and throughput compared to traditional methods and OpenSearch (up to 5x faster); support for multi-stage interaction models such as ColPali and ColBERT; integration of ELSER and e5 multilingual dense vector models, and support for JinaAI's embeddings and reranking capabilities, making semantic search easier for users. In addition, the new version enhances hybrid search capabilities and introduces the ES|QL Join command, improving the flexibility of cross-data querying.

This article announces significant updates to Cloudflare's Workers AI platform aimed at improving inference accessibility and efficiency. Key announcements include speeding up inference by 2-4x using techniques like speculative decoding and prefix caching, introducing an asynchronous batch API for handling large workloads more efficiently, and expanding LoRA support for greater model customization. The article also covers a new dashboard, updated pricing, and the addition of several new AI models to the platform.

This article details how Kuaishou E-commerce leverages LLMs to empower business-side merchants, enhancing their operational efficiency and service quality. Addressing the diversity and complexity of e-commerce scenarios for business, Kuaishou E-commerce tackles factual issues in product understanding and content creation by constructing an e-commerce LLM foundation, comprising the application, capability, solution, and architecture layers. Meanwhile, Zhilin Engine and Qianji Platform lower the barrier to LLM application development, enabling no-code, configuration-based delivery. Additionally, Retrieval-Augmented Generation (RAG) technology optimizes intelligent assistants, improving accuracy by approximately 17% in intelligent customer service scenarios. Multi-agent collaboration addresses complex business scenarios such as pre-sales, mid-sales, after-sales, and policy consultation. Finally, the Hongru Platform ensures the reliability and compliance of LLM applications through evaluation and monitoring. The overall goal is to make AI an assistant for merchants, influencers, and operations personnel, promoting innovation and development in the e-commerce industry, especially with in-depth practices in engineering and evaluation systems.

Keling AI (Keling Artificial Intelligence) has released version 2.0 of its video generation model and image generation model, marking a new era in AI video creation. The Keling 2.0 video generation model has significantly improved in semantic understanding, dynamic quality, and aesthetic appeal, enabling it to better understand complex prompts and generate more fluid and natural video content. The Ketu 2.0 (Ketu Image 2.0) image generation model has been upgraded in instruction-following ability, cinematic aesthetic expression, and style diversity, supporting nearly a hundred styles and providing features such as local re-painting, image expansion, and style transfer. A key feature of Keling AI is 'instant access upon release,' allowing global members to experience it immediately. The underlying technology adopts a new DiT Architecture. Through technological innovation and upgraded training strategies, Keling AI has surpassed competitors such as Google Veo2 and Sora in multiple evaluations, establishing its leading position in the global AI video generation field. Keling AI also released a new interaction concept, Multi-modal Visual Language (MVL), aimed at improving communication efficiency between humans and AI, enabling more precise creative expression. Simultaneously, it launched the 'Keling AI NextGen New Image Venture Program,' investing millions to increase support for AIGC creators.

The article introduces significant upgrades made by Anthropic to its AI Assistant, Claude. This update mainly includes three aspects: First, the launch of the Research Feature, which is currently in early Beta Testing and features an Agentic Search Framework, Cross-Source Information Integration, Systematic Problem Exploration, and Verifiable Comprehensive Answers, aiming to enhance information processing capabilities. Second, deep integration with Google Workspace, connecting core applications such as Gmail, Google Calendar, and Google Docs to achieve Automated Context Acquisition and Context-Aware Driven Assistance, simplifying user interaction steps with AI. Third, the upcoming Voice Mode, Anthropic is catching up, competing with competitors such as OpenAI in the field of Multimodal Interaction. These updates aim to enhance Claude's practicality and intelligence, allowing for greater possibilities in practical applications like market research and academic research. This update represents a significant advancement for Claude in becoming a more intelligent and user-friendly AI Assistant.

This article introduces the upgrade of Google Veo 2 and its powerful functions in video creation. Veo 2 can generate high-quality, cinematic video clips through simple text prompts, democratizing video creation. The article showcases Veo 2 in various applications, highlighting its advantages in lighting, camera movement, and detail processing. In addition, the article also introduces Freepik AI Suite, a creative toolkit used in conjunction with Veo 2, which can further improve the efficiency and quality of video creation. Overall, the article aims to demonstrate the great potential of AI technology in the field of video creation, as well as its benefits for video creators and AI enthusiasts.

The article reviews Dia, the AI browser launched by the Arc team, arguing that it changes the traditional browser interaction model through deep integration of AI technology, appealing to users interested in AI and novel information acquisition. Dia allows users to have conversations with webpage content and process contextual information from multiple webpages simultaneously, thereby quickly obtaining high-quality answers. The article also introduces innovative designs such as Dia's Smart Cursor, which are designed to make AI an extension of user thinking rather than a standalone tool. The author believes that Dia represents a new form of future browsers, transforming the browser from a 'document center' to a 'dialog center', empowering users to acquire information through intent expression, not just operations. Although Dia currently only supports Mac M1+ chips and is still in its early stages, it has demonstrated the enormous potential of AI in the browser field.

This article shares a set of prompt templates for Ji Meng AI, aimed at helping users quickly generate various styles of text design by inputting text content. The template is easy to use and supports various design aesthetics. Users can quickly get started by following the operation steps provided in the article. The article details the construction ideas of the prompts, that is, by analyzing high-frequency prompts in high-quality images and combining font effect descriptions to form a system that AI can understand and generate drawing prompts. The article provides various styles of text design cases, including abstract, e-sports, Chinese style, sweet, etc. Ji Meng AI shows potential for further development in incorporating advanced professional font design elements.

This article interviews Zhou Changyin, the founder of Vozo AI, sharing his experience from being a researcher at Google X to a successful AI video tool entrepreneur. It details Vozo's feature iteration and market strategy, including using AI technology to re-dub videos, translate (including technologies such as voice cloning, speech synthesis, and AI lip sync), and edit videos. It also covers achieving a cold start through Product Hunt, Vozo's technology selection considerations—such as avoiding general models and focusing on specific professional needs—and balancing innovation with user needs to achieve product-market fit. Additionally, Zhou Changyin shares his Google X experience and lessons from his first venture, emphasizing the need to focus on clear user needs and choose the right business model. The article concludes with Vozo's future development strategies, including product consolidation and unified branding.

This article is OpenAI scientist Yao Shunyu's interpretation of the second half of AI. The core view is that AI is shifting from problem-solving to problem definition, and model evaluation will be more important than model training. The article reviews the characteristics of the first half of AI, which focused on algorithm and model innovation, such as Transformer, AlexNet, GPT-3, etc., points out the key role of Reinforcement Learning (RL) in realizing Artificial General Intelligence (AGI), and underscores the significance of prior knowledge. The author believes that the second half of AI needs to rethink evaluation methods, break assumptions such as automated evaluation processes and independent and identically distributed (i.i.d.), and focus on real-world utility to realize the true value of AI. The article also mentions that in fields such as computer use and network navigation, RL Agents still need to improve their zero-shot ability. Finally, the article encourages AI researchers and practitioners to focus on practical applications, break inherent thinking patterns, transform intelligence into useful products, and create companies with huge commercial value.

The article delves into the development trends of AI Avatars, providing a comprehensive analysis from technological evolution and application scenarios to future prospects. The article reviews the evolution of AI Avatars, from early CNN and GAN models to current Transformer and Diffusion models, highlighting significant improvements in generation quality and capabilities. It then explores the applications of AI Avatars across consumer, SMB, and enterprise sectors, including character creation, advertising, learning and development, and content localization. In addition, the article analyzes the key elements of AI Avatars, including face, voice, lip-sync, body, and background, and proposes possible future development directions, such as the stability and deformability of characters, more natural facial movements and expressions, body movements, and interactions with the real world. Finally, based on the author's personal testing of more than 20 AI Avatar products, the article provides an in-depth analysis of industry development trends.

This article summarizes Google Chief Scientist Jeff Dean's speech at ETH Zurich, focusing on Google's foundational research contributions to AI over the past fifteen years. The speech covered the development of key technologies including Neural Networks, Backpropagation, DistBelief, Word2Vec, Sequence-to-Sequence Learning Model, and TPUs. Google has significantly contributed to AI hardware, notably the development of TPUs. These technologies form the cornerstone of modern AI and have driven the development of advanced models like Gemini. Jeff Dean also emphasized the positive impact of AI on society, stressing the importance of continuous research and innovation for a future augmented by AI.

This article interprets Richard Sutton and David Silver's latest paper, 'Welcome to the Era of Experience,' highlighting the shift from the 'Human Data Era' to the 'Era of Experience' in AI development. It highlights the limitations of current AI's reliance on human data. To achieve superhuman intelligence, AI needs to interact with the environment and learn from its own experiences, forming 'Experience Streams.' Experience is infinite, can break through the boundaries of human knowledge, and is the native language of intelligent agents. The future direction of AI development is a cycle of 'action + feedback' rather than 'prompt + knowledge base.' A key feature of the 'Era of Experience' is the close integration of agent actions and observations with the environment, with reward mechanisms derived from environmental experiences. The article also discusses the importance of Experience Streams for the long-term development of AI, as well as the potential risks and challenges of experience learning.

This article presents Qu Kai's interview with Wu Yi, Assistant Professor at Tsinghua University's Institute for Interdisciplinary Information Sciences, delving into Reinforcement Learning (RL) and its latest advancements. Wu Yi differentiates RL from traditional machine learning, emphasizing its advantages in multi-step decision problems, particularly in solving LLM instruction-following with RLHF. Wu Yi also shares OpenAI's exploration in RL and its application in the Agent paradigm. Furthermore, the importance of infrastructure for developing RL talent is discussed, along with insights into life decisions, emphasizing hands-on skills, open-mindedness, and proactive exploration, and noting that startups should avoid a fixed, end-game strategy.

This issue of deeplearning.ai The Batch covers three main topics. First, it emphasizes the importance of iteratively building evaluation systems for GenAI applications, starting with small, imperfect evaluations and gradually improving them. Second, it introduces Google's new Gemini 2.5 Pro Experimental model, which outperforms competitors in several benchmarks and incorporates chain-of-thought training in all new models, demonstrating that AI progress has not slowed. Third, it discusses OpenAI's support for the Model Context Protocol (MCP), an open standard that facilitates connecting LLMs to various tools and data sources, promoting the development of agentic applications. Finally, the newsletter reviews the behind-the-scenes events of Sam Altman's brief firing and reinstatement as OpenAI CEO.