BestBlogs.dev Highlights Issue #28

DeepSeek-V3, DeepSeek's latest self-developed MoE model, features 671B parameters with 37B active parameters and was pre-trained on 14.8T tokens. It significantly outperformed other open-source models like Qwen2.5-72B and Llama-3.1-405B in various evaluations, achieving performance comparable to leading closed-source models such as GPT-4o and Claude-3.5-Sonnet. DeepSeek-V3 demonstrates substantial improvements in encyclopedic knowledge, long-form text generation, code, mathematics, and Chinese language processing. Algorithmic and engineering advancements have boosted generation speed from 20 TPS to 60 TPS, enhancing user experience. The API service pricing has been revised, with a 45-day discounted trial period available. DeepSeek-V3's native FP8 weights are open-sourced, supporting inference frameworks including SGLang, LMDeploy, TensorRT-LLM, and MindIE, encouraging community contributions and expanding application scenarios. DeepSeek plans to continue building upon the DeepSeek-V3 base model and share its ongoing research with the community.

DeepSeek V3, a Mixture of Experts (MoE) model with 671 billion parameters and 37 billion active parameters, was pre-trained on 14.8 trillion high-quality tokens. It outperforms open-source models like Llama 3.1 405B on multiple benchmarks and rivals top closed-source models such as GPT-4o and Claude 3.5 Sonnet. Its remarkably low training cost of $5.576 million, significantly less than comparable models, and its competitively priced API make it a game-changer. The article details DeepSeek V3's architectural optimizations, training strategies, and performance, highlighting its efficiency in resource-constrained environments and its innovative approach to distributed inference and load balancing.

OpenAI's twelve-day event unveiled a series of significant AI advancements. The o3 model, while expensive (potentially $3500 per query), demonstrated superior performance to existing models. Day nine's developer update focused on improved structured output, a crucial infrastructure for future AI agent development. Other key announcements included ChatGPT Pro, the impressive Sora video generation technology (despite some technical limitations), and RFT (Reinforcement Fine-Tuning) technology. These releases highlight rapid progress in AI across mathematics, coding, and video generation. The event also explored the evolving definition of AGI, emphasizing the importance of learning. OpenAI's announcements reinvigorated the AI industry, offering a path forward during challenging times.

Anthropic's collaboration with various industry teams over the past year yielded insights into building and deploying large language model (LLM) agents. Contrary to expectations, the most effective agents weren't based on massive, intricate frameworks or specialized libraries, but rather on simple, composable designs. The report details agent definitions, applications, building blocks, and diverse workflows (including prompt chaining, intelligent routing, parallel processing, leader-follower models, and iterative evaluation-optimization). It emphasizes the importance of maintaining simplicity, transparency, and thorough documentation throughout the development process. Anthropic advises developers to start with direct LLM API usage, gradually increasing complexity only when demonstrably improving results. In the realm of AI, as in cooking, the most sophisticated results often come from the simplest approaches.

The article summarizes the state of LLM agents in 2024, based on a presentation by Professor Graham Neubig at Latent Space LIVE! during NeurIPS 2024. It covers the growth of companies like OpenHands (formerly OpenDevin), the increasing importance of agents across various domains, and eight perennial challenges in agent development. These challenges include the agent-computer interface, human-agent interaction, LLM selection, planning, reusable workflows, exploration, search, and evaluation. The article also showcases practical applications of coding agents through live demos, emphasizing their utility in tasks like data analysis and software development. Neubig's presentation highlights the potential of 2025 to be the "year of agents" with advancements by major players like OpenAI, DeepMind, and Anthropic.

This article details the significant progress in RAG technology throughout 2024, highlighting innovative systems and frameworks such as GraphReader, MM-RAG, CRAG, RAPTOR, and T-RAG. These systems leverage graph-based approaches, multimodal integration, and self-correction mechanisms to substantially improve RAG's performance and expand its applications. The article includes numerous links to relevant papers and projects, facilitating deeper learning and research. Furthermore, it explores RAG's applications in diverse fields like medicine, finance, and open-domain question answering, showcasing its innovative capabilities in complex data analysis, long-form text generation, and multi-hop question answering.

Google Engineering Director Addy Osmani's article explores the practical applications and challenges of AI programming in 2024. AI programming tools are categorized into "Bootstrappers," for rapid initial code generation, and "Iterators," for iterative tasks like code completion and refactoring. While these tools enhance development speed, their impact differs significantly between senior and junior engineers. Senior engineers leverage AI effectively, refining and optimizing generated code, while junior engineers risk creating brittle code, prone to failure. The article introduces the "70% Problem," where non-engineers can quickly achieve 70% of a task using AI, but the remaining 30% demands deep expertise, often leading to prolonged debugging cycles. Furthermore, relying heavily on AI tools may hinder learning fundamental programming and debugging skills. The article concludes by forecasting the rise of Agent Software Engineering, where AI agents will autonomously plan, execute, and iterate solutions, acting as powerful collaborators for developers rather than replacements.

Hugging Face has launched smolagents, a lightweight library designed to empower language models (LLMs) with agentic capabilities by allowing them to write actions in code. The library simplifies the creation of AI agents that can interact with the real world through tools like search APIs and other functions. The article explains the concept of agency in AI systems, illustrating how LLM outputs can control workflows, and introduces a spectrum of agency levels from simple processors to multi-step agents. smolagents emphasizes simplicity, first-class support for code agents, and integration with the Hugging Face Hub. It also supports various LLMs, including open-source models, and provides a secure execution environment via E2B. The article includes practical examples, such as building a travel planner agent, and benchmarks showing the effectiveness of open-source models in agentic workflows. Future steps include tutorials, best practices, and advanced use cases like text-to-SQL and multi-agent orchestration.

This article details Li Jigang's experiences using prompts to interact with large language models, delving into their essence, design, and impact on model output. He argues that prompts are a specialized language for human-AI communication, with different prompts producing vastly different results. Based on two years of dialogue with large models, Li Jigang defines prompts not as mere tools, but as a unique entity—a 'Cosmic Language' bridging human cognitive understanding and the AI's internal parameter space. He emphasizes that prompts are not just communication tools, but keys to achieving resonance and surpassing limitations. By carefully crafting the interaction context, large models can be guided beyond default outputs, resonating with human cognition to generate exceptional results. The article also demonstrates how the Johari Window framework can optimize AI dialogue, and discusses the impact of AI evolution on prompt usage for entrepreneurs and individuals. Finally, Li Jigang addresses the evolving nature of human cognition in the AI era, stressing the importance of maintaining a resonant approach to AI interaction and avoiding passive acceptance of AI-driven outputs.

In an insightful article, OpenAI Chairman Bret Taylor examines the emergence of AI-driven autonomous software engineering. Taylor highlights the evolving role of software engineers, transitioning from code writers to operators of code generation machines, fueled by the rapid advancement of AI programming tools. He proposes that future programming systems should be inherently designed for this workflow, discussing the crucial roles of programming language design, formal verification, testing, and development workflows. Taylor emphasizes that AI not only expands software creation opportunities but also enhances software capabilities, urging the software engineering community to actively explore and develop new systems for this autonomous era.

This article explores the emergence of Recraft, a groundbreaking AI image generation tool. In October 2024, Recraft achieved top ranking in a text-to-image competition, earning an ELO rating of 1172 and outperforming prominent models such as Midjourney and Stable Diffusion, solidifying its position as a community favorite. Its V3 model excels in image quality, prompt fidelity, and anatomical accuracy, particularly when generating complex scenes and images with extensive text descriptions. Furthermore, Recraft's naturalistic aesthetic and minimalist interface have driven its rapid adoption on platforms like Xiaohongshu (a popular Chinese social media platform), accumulating over 2.18 million likes—significantly more than Midjourney and Stable Diffusion. The article concludes by highlighting Recraft's potential for future growth, driven by its expanding style library and ongoing feature enhancements, ultimately shaping the future of the AI image generation industry.

Speak, the pioneering AI language learning unicorn, recently secured $78 million in Series C funding, achieving a $1 billion valuation. Its ARR nears $50 million, exhibiting 100% year-over-year growth, and it serves over 10 million users, primarily in South Korea and Japan, with expansion into Taiwan underway. Speak leverages AI to deliver cost-effective, personalized one-on-one language tutoring, revolutionizing traditional language learning. Its speech recognition system adeptly handles accents, providing swift and reliable feedback to enhance user experience. Speak's pricing strategy caters to a broad spectrum: offering mass-market appeal while also providing a premium experience at a higher price point. Rigorous model evaluation is central to Speak's AI development; internal tools and iterative evaluation loops ensure optimal performance of new models. Unlike Duolingo, which focuses on informal language learning, Speak specializes in helping non-native English speakers achieve fluency. Looking ahead, multimodal audio technology holds immense potential in language learning, though currently in its nascent stages.

Bolt.new, a browser-based operating system incorporating AI code generation, dramatically simplifies programming, enabling rapid web application development and deployment. Its success stems from its browser-based environment, advancements in AI code generation, and its focus on non-professional developers. By combining WebContainer technology and AI models, Bolt eliminates the complexities of setting up local development environments, making it ideal for beginners. Seamless integration with deployment platforms like Netlify provides a smooth user experience, allowing quick website launches without intricate configurations. At its core, Bolt utilizes a WebAssembly operating system running within the browser, ensuring a fast development environment without per-minute charges. Mature AI code generation technology, such as Sonnet 3.5, produces high-quality code, empowering even non-technical users to build sophisticated applications. The open-source version, Bolt Local, serves as a crucial tool for testing new code generation models, fostering community evaluation and optimization. Bolt's robust error handling and task decomposition significantly enhance user experience, particularly in complex projects. By providing comprehensive contextual information, Bolt significantly improves the efficiency and usability of AI coding tools. A flexible pricing model, encompassing fixed subscriptions and on-demand token purchases, caters to diverse user needs, fueling rapid revenue growth. Bolt's design and model support excel at finding optimal, streamlined solutions, minimizing file count and maximizing performance. Dominic Elm, the team's lead engineer, brings invaluable cross-domain expertise in prompt engineering and multi-agent systems development. The company's shift from a B2B to a B2C model involved strategic adjustments, prioritizing data analysis and user experience.

This article shares the experience of a senior product manager who implemented a six-step user growth strategy for their AI startup, "AI Researcher." The strategy successfully addressed user acquisition, improved conversion rates, clarified product iterations, and defined the core focus of user growth efforts. By analyzing the monthly paying user formula, the author prioritized acquisition channels, target users, and research scenarios, identifying key growth levers. The strategy focused on platforms like Xiaohongshu (a popular Chinese social media platform similar to Pinterest), Zhihu, and WeChat Official Accounts, targeting high-value users in finance and product management. The ICE Scoring System helped prioritize Xiaohongshu and improving conversion rates within the industry research scenario. Data analysis revealed a correlation between the number of user-generated research records and paid conversion rates, identifying a "key metric" of 3 records. This finding informed user growth strategies and product experience optimization. Improvements included simultaneous generation and output to reduce wait times, WeChat subscription notifications, automated reference retrieval, and enhanced user experience during the report generation process. The Hooked Model guided product optimization, including features for managing research materials, reward mechanisms, and user action guidance. A/B testing validated growth strategies, such as optimizing report generation, visual improvements, and trigger channels, significantly improving conversion rates and retention. Future growth plans include mobile expansion, enhanced search and reading functionalities, and collaboration with Key Opinion Leaders (KOLs).

This article examines the user interaction experience design of large language models (LLMs) within input, analysis, and output modules, drawing upon the Black Box Theory. While the complexity of LLMs can be daunting, the author argues that improvements in user experience are achievable through optimized input prompts, reduced latency, interruption tools, and enhanced objectivity in generated content. The importance of multimodal output and user feedback within AI platforms is also discussed, along with a forward-looking perspective on future AI application development. The author emphasizes that despite rapid technological advancements, user experience design remains paramount for the success of AI products.

The article, authored by Andrew Ng and featuring insights from various AI leaders, delves into the transformative potential of AI in 2025. It highlights the ease of building software prototypes using AI, which lowers development costs and accelerates the creation of simple applications. The article also discusses the role of generative AI in enhancing creativity, particularly in fields like art and video production. Key themes include the importance of safety, accessibility, and customization in AI development, as well as the potential for AI to revolutionize learning and problem-solving. The article concludes with a call to action for readers to embrace AI by learning and building with it, highlighting the fun and educational value of prototyping.

Andrew Ng's article on DeepLearning.ai reviews the significant AI advancements in 2024. While underlying AI technology has progressed, the acceleration of applications is even more remarkable. Agentic systems, capable of reasoning, tool use, and desktop application control, have become prominent. Fierce competition has dramatically reduced model prices, with smaller, efficient models proliferating. This progress creates a widening gap between those at the cutting edge and those yet to engage with AI. Ng expresses optimism for 2025, conditional on avoiding restrictive regulations, and encourages continuous learning to keep pace with AI's transformative impact.

The article provides a comprehensive review of the AI landscape in 2024, focusing on the evolution of AI engineering, the competitive dynamics among major AI players, and key debates within the field. It highlights the growth of AI engineering, marked by events like the AI Engineer Summit and the rise of platforms like GitHub Models. The discussion also covers the transition from research-heavy to engineering-heavy approaches in AI/ML, the challenges of scaling large pre-trained models, and the importance of inference time compute. The competitive landscape is analyzed, with OpenAI, Gemini, and Anthropic leading the race, and the article notes significant market share shifts and the impact of price wars. The evolving definition of 'small models' and the role of private test sets in evaluating large language models are also discussed.

This report summarizes major 2024 advancements in artificial intelligence, focusing on breakthroughs in large language models, image generation, and speech recognition. Companies like OpenAI, Google, and Anthropic competed fiercely on model capabilities, speed, and pricing, accelerating AI's iteration and adoption. The performance gap between open-source and proprietary models narrowed significantly. Inference costs dropped substantially, with smaller models achieving the intelligence levels previously exclusive to larger ones. Image generation saw improvements in photorealism, prompt fidelity, and text rendering. Speech recognition reached new milestones. The report concludes by looking ahead to 2025 and the potential for AGI (Artificial General Intelligence).

In a year-end interview, Microsoft CEO Satya Nadella detailed Microsoft's strategic shifts in AI and cloud computing. He reflected on past successes and failures in search, mobile, and cloud, emphasizing the crucial role of a growth mindset in Microsoft's resurgence. He attributed Microsoft's past decline to arrogance, but noted that a growth-oriented mindset and strategic focus enabled the company to establish a unique position in the cloud market. Regarding AI, Nadella acknowledged missed early investment opportunities but highlighted the transformative partnership with OpenAI, particularly in Transformer models and natural language processing. He described the intense competition in AI infrastructure and models, but emphasized Microsoft's regained market share through Azure and Copilot. Nadella predicted significant network effects in AI applications, varying across different scenarios. He also discussed the complexities of AI agent interaction across ecosystems, focusing on OS-level security and user authorization. He outlined the future of AI assistants, emphasizing memory, tool usage, and permission management as key elements, mentioning Microsoft's deployment of autonomous AI assistants in enterprise settings. Nadella revealed that Microsoft's AI revenue primarily stems from API services, especially the inference costs of applications like ChatGPT and Copilot, and underscored Microsoft's two-year lead with OpenAI. He suggested that future businesses might favor asset-light operations, building upon existing models. Finally, he discussed Microsoft's strategies for capital expenditure, AI accelerator development, and model scaling, emphasizing software-intensive, multi-layered product structures and the importance of scaling laws (the principle that model performance improves with scale). He reiterated Microsoft's focus on post-training work rather than reinventing the wheel, emphasizing collaboration with OpenAI.

This article analyzes the 2024 AI landscape. While overall progress lagged behind expectations, significant strides were made in voice, video, and agent technologies. OpenAI's advancements, especially the logical reasoning capabilities of its o1 and o3 models (OpenAI's new models, versions 1 and 3 respectively), are noteworthy. The article emphasizes the need for startups to concentrate on specific application scenarios to compete effectively. Domestic large language model companies are urged to adopt a more focused approach than Anthropic, specializing in niche areas. Furthermore, the article contrasts the domestic and international AI startup environments, highlighting the shortcomings in domestic toolchains and infrastructure, and the burgeoning application ecosystems overseas.

In the Radical AI Founder Masterclass, Fei-Fei Li discussed key aspects of AI entrepreneurship. She stressed the critical role of data and computational resources, emphasizing the need for reliable data sources and a robust data flywheel. She also addressed team building challenges, particularly the need for mindset shifts among technically-minded founders to focus on user needs and maintain transparent communication. Fei-Fei Li highlighted the importance of corporate culture as the foundation for long-term stability. She explored AI applications in entertainment, healthcare, and manufacturing, sharing insights on product development and female leadership. Finally, she likened entrepreneurship to a personal journey, urging the pursuit of a 'North Star' goal, authenticity, and resilience in the face of adversity.

In a conversation with Tencent Technology, Li Xiang, CEO of Li Auto, shared his insights on AI and autonomous driving. He views cars evolving from industrial-age transportation to intelligent mobile spaces in the AI era, with AI as the key to future competitiveness. Li Auto heavily invests in AI, dedicating half its R&D budget to it, and leads globally in foundation models, end-to-end technology, and vision-language model (VLM) technology. Li Xiang stressed Li Auto's ambition to become an AI company, not just an electric vehicle manufacturer, believing AI will define the future. He also discussed OpenAI's advancements in chatbots, reasoners, and agents, highlighting the importance of foundation models. Li Xiang's ultimate vision is an 'AI family member,' a goal he hopes to achieve in his lifetime.

In a conversation with financial author Zhang Xiaojun, Shixiang Technology CEO Li Guangmi discussed the current state and future trends of the large language model (LLM) competition. He highlighted the parallel development of computing power, models, and applications as key to success, with companies like OpenAI, Anthropic, and Perplexity competing for dominance. Li Guangmi forecasts that 2025 will be defined by coding and agent technologies, where AI companies will innovate through dynamic software orchestration and productivity task reorganization. The discussion also analyzed ChatGPT's business model and challenges, emphasizing the crucial role of context in AI development and the emergence of new software generation paradigms. Ultimately, the goal is for AI companies to surpass Google by becoming the ultimate distributor of information, content, and tasks, evolving into the next generation of super assistants.

The article presents a curated reading list of approximately 50 essential papers for AI engineers, designed to be practical and relevant for 2025. The list is organized into ten sections, each focusing on a critical area of AI engineering: Frontier LLMs, Benchmarks and Evals, Prompting, Retrieval Augmented Generation (RAG), Agents, Code Generation, Vision, Voice, Image/Video Diffusion, and Finetuning. The selection criteria emphasize practical relevance, avoiding overly theoretical or widely known papers like 'Attention is All You Need.' Each section highlights five key papers, providing context on why they matter and their practical applications. The article also includes honorable mentions and additional resources, making it a valuable guide for both newcomers and experienced practitioners in the AI field.