Core Viewpoint - The development of AI is fundamentally changing the software programming paradigm, leading to the emergence of Software 3.0, where natural language prompts are replacing traditional programming code, and large language models (LLMs) are becoming the new programming interface [2][3]. Group 1: Software Evolution - Software 1.0 was characterized by human-written code, while Software 2.0 shifted to neural networks, requiring data preparation and parameter training [2]. - Software 3.0 represents a significant transformation in software development, driven by the rise of large language models [2]. - The transition to Software 3.0 necessitates advancements in hardware, referred to as Hardware 3.0, to support new computational demands [2][3]. Group 2: Hardware Requirements - The dominance of CPUs in Software 1.0 has shifted to GPUs in Software 2.0 due to the need for parallel processing capabilities [3]. - The rapid development of transformer-based models in Software 3.0 has led to the increased adoption of Domain-Specific Architectures (DSA) [5]. - A balance between specialized efficiency and programming generality is crucial for the development of Hardware 3.0 [5][8]. Group 3: Challenges in AI Processor Design - Key challenges in designing AI processors include the lengthy time required to construct AI computing architectures, the prolonged development of instruction systems, and the long cycles for compiling software [9]. - Achieving widespread ecosystem support for self-built instruction systems presents significant hurdles [9]. Group 4: RISC-V and EVAS Architecture - RISC-V's open and modular design allows for the customization of AI acceleration instruction sets, making it a suitable foundation for DSA [8]. - The introduction of the Virtual Instruction Set Architecture (VISA) aims to bridge the gap between AI compilers and backend compilation, enhancing performance optimization [10][11]. - The EVAS architecture integrates VISA with RISC-V microinstructions, ensuring efficient execution of AI computations while improving user programming experience [12][16]. Group 5: Upcoming Innovations - The upcoming chip from the company will support various data types, including INT4, INT8, FP8, FP16, and BF16, with a focus on mixed-precision computing [17]. - The new architecture aims to provide advanced computing solutions for applications in autonomous driving, embodied intelligence, and other edge-cloud industry applications, contributing to the progress towards AGI [17].

明敏 整理自 凹非寺 量子位 | 公众号 QbitAI 大神Karpathy提出"软件3.0"才两周，"软件3.5"已经诞生了？ 交互即智能。 指AI不再是黑盒工具，而是透明的思维伙伴。用户可以在AI思考的任何节点进行干预，提供战略指导或纠正方向。 也就是说，智能是在人类与AI的不断交互合作中涌现。 Software 3.0作为一个概念，在2024年9月之后已经显得有些过时了 。为什么这么说？ Software 3.0的核心困境源于它诞生时的技术背景。2022年ChatGPT发布时，AI的主要能力还集中在文本生成和简单推理，"自然语言编 程"确实是那个时代的最佳解决方案。但2024年9月之后，我们见证了AI能力的代际跃迁：从GPT-4的生成能力到o1系列的深度推理，从简单 的指令执行到具备元认知意识的思考能力。 最关键的变化是，人类首次能够与AI进行真正的思维层面交流 ——AI不仅理解我们说什么，更能理解我们为什么这么说，甚至能主动寻求认 知层面的协作。这种质的飞跃让传统的"输入prompt→等待处理→接收结果"模式显得笨拙而低效，就像用电报方式进行现代通信一样不合时 宜。 他们认为，2024年9月之后，随着 ...

Core Insights - The article discusses the evolution of software paradigms from Software 1.0 (traditional coding) to Software 2.0 (neural network weights) and now to Software 3.0 (prompts), emphasizing the significance of natural language as a programming language [3][8][11] - It highlights the emergence of Large Language Models (LLMs) as a new type of operating system (LLM OS), reshaping the computing ecosystem and enabling new forms of interaction with AI [5][8] - The article identifies the greatest opportunity in developing "partially autonomous" AI applications, which enhance human capabilities rather than aiming for full automation [10][11] Software Paradigms - Software 1.0 involves traditional coding with specific programming languages, while Software 2.0 utilizes neural networks where data sets are prepared to optimize parameters [3] - Software 3.0 introduces prompts as the programming language, allowing for a more accessible and intuitive way to interact with AI [3][8] LLM as an Operating System - LLMs are compared to a new operating system, where they act as the CPU, with their expanding context window serving as memory, and external tools functioning as peripherals [5][8] - The current state of LLMs is likened to the 1960s computing era, where they are primarily cloud-based and accessed through thin clients [6][8] Opportunities in AI Development - The article emphasizes the need to understand the "mental model" of LLMs, which exhibit human-like characteristics but also have limitations such as hallucinations and memory issues [7][10] - Successful AI applications should focus on creating a feedback loop where AI-generated content is quickly verified by humans, enhancing efficiency [10] Accessibility of Software Development - Software 3.0 lowers the barrier to entry for programming, allowing individuals without formal training to create software through natural language [11] - The future of software design must cater not only to humans but also to intelligent agents, necessitating new standards and tools for better interaction [11][12]