ChatGPT流量为何下滑？AI竞争与用户期望

Introduction

The rise of ChatGPT represents a pivotal moment in the history of artificial intelligence. Developed by OpenAI, this conversational AI model has rapidly evolved from a research prototype into a global phenomenon, reshaping industries and redefining human-computer interaction. This blog post chronicles the technical development, core architecture, and far-reaching impact of ChatGPT and its underlying GPT models.

ChatGPT 的崛起是人工智能发展史上的一个关键时刻。由 OpenAI 开发的这个对话式 AI 模型，已迅速从一个研究原型演变为全球现象，重塑了行业格局并重新定义了人机交互。本文旨在梳理 ChatGPT 及其底层 GPT 模型的技术发展历程、核心架构和深远影响。

Development History of OpenAI and GPT Models

Founding of OpenAI and Early Research (2015-2017)

OpenAI is an artificial intelligence research laboratory based in San Francisco, USA, consisting of the for-profit company OpenAI LP and its non-profit parent, OpenAI Inc. Originally established as a non-profit organization in late 2015 by Elon Musk, Sam Altman, Ilya Sutskever, and other investors, the company was dedicated to developing artificial intelligence and natural language tools.

OpenAI 是一家位于美国旧金山的人工智能研究实验室，由营利性公司 OpenAI LP 及其非营利性母公司 OpenAI Inc 组成。该公司最初于 2015 年底由埃隆·马斯克、萨姆·奥尔特曼、伊尔亚·苏茨克维及其他投资者作为非营利组织创立，致力于开发人工智能和自然语言工具。

In 2016, Microsoft's Azure cloud service provided OpenAI with the necessary computational power, enabling the large-scale deep learning and neural network rendering required for future models like ChatGPT. That same year, OpenAI released its first products: OpenAI Gym and Universe, open-source toolkits for reinforcement learning.

2016 年，微软 Azure 云服务为 OpenAI 提供了必要的算力条件，使得日后 ChatGPT 所需的大规模深度学习和神经网络渲染成为可能。同年，OpenAI 发布了其首个产品：用于强化学习的开源工具包——OpenAI Gym 和 Universe。

A pivotal breakthrough came in 2017 with the "Attention is All You Need" paper by Google researchers, which introduced the Transformer architecture. This architecture became the foundational core for nearly all mainstream generative AI models. Concurrently, OpenAI's work on projects like "Sentiment Neuron" and OpenAI Five shifted its focus towards Large Language Models (LLMs) and scaling parameters, strengthening its belief in achieving Artificial General Intelligence (AGI).

一个关键的突破发生在 2017 年，谷歌研究人员发表的论文《注意力就是你所需要的一切》引入了 Transformer 架构。该架构已成为几乎所有主流生成式人工智能模型的核心基础。与此同时，OpenAI 在“情绪神经元”和 OpenAI Five 等项目上的工作，使其开始关注大型语言模型和参数规模，增强了其实现通用人工智能的信念。

The GPT Era: Iterative Scaling

GPT-1 (2018)

OpenAI released the first-generation large-scale model, GPT-1. It was trained on a corpus of billions of text documents, with 117 million parameters. GPT-1 combined the Transformer architecture with unsupervised learning, a method for training machine learning models on unannotated data.

OpenAI 发布了最早的一代大型模型 GPT-1。它基于数十亿文本档案的语言资料库进行训练，模型参数量为 1.17 亿个。GPT-1 将 Transformer 架构与无监督学习相结合，这是一种根据事先未标注的数据训练机器学习模型的方法。

GPT-2 (2019)

GPT-2 was released with its parameters scaled up to 1.5 billion. Its architecture principle remained the same as GPT-1, with the main difference being its significantly larger scale (10x). This period also saw a major structural shift at OpenAI. Following Elon Musk's departure and the unexpectedly high funding demands of large models, OpenAI transitioned from a non-profit to a "capped" for-profit structure in March 2019, creating the limited-profit entity OpenAI LP under its non-profit parent. In July, Microsoft announced a multi-year partnership with OpenAI, investing $1 billion and co-developing new Azure AI supercomputing technologies.

GPT-2 发布，模型参数量提高到 15 亿个。其模型架构原理与 GPT-1 相同，主要区别在于规模更大（10 倍）。这一时期，OpenAI 的内部结构也发生了重大变化。随着埃隆·马斯克的退出以及大模型对资金的超预期需求，OpenAI 于 2019 年 3 月从非营利性转变为“封顶”的营利性结构，在非营利母公司下创建了限制性营利实体 OpenAI LP。同年 7 月，微软宣布与 OpenAI 开展为期多年的合作，注资 10 亿美元，并共同研发新的 Azure AI 超算技术。

GPT-3 (2020)

GPT-3 emerged with a staggering 175 billion parameters, over 10 times more than GPT-2. Technically, it removed the fine-tuning step of the first-generation GPT, directly using natural language prompts as instructions. This trained GPT-3 to continue text and answer questions based on prior context, covering a broader range of topics. GPT-3 achieved a monumental leap in generating human-like text, capable of answering questions, summarizing documents, generating stories in various styles, and translating between languages like English, French, Spanish, and Japanese. In September 2020, Microsoft secured an exclusive license to the GPT-3 model, granting it sole access to the underlying code.

GPT-3 诞生，参数量达到了 1750 亿个，是 GPT-2 的 10 倍以上。在技术路线上，它去掉了初代 GPT 的微调步骤，直接输入自然语言当作指示进行训练，使模型能够根据读过的文字接续问题，同时涵盖了更为广泛的主题。GPT-3 实现了生成类人文本能力的巨大飞跃，可以回答问题、总结文档、生成不同风格的故事，并在多种语言间进行翻译。2020 年 9 月，微软公司获得了 GPT-3 模型的独占许可，意味着微软可以独家接触到 GPT-3 的源代码。

From InstructGPT to ChatGPT and Beyond

InstructGPT (2022)

In January 2022, OpenAI fine-tuned GPT-3 using supervised training and iterative refinement, releasing InstructGPT. This model was better at following human instructions and produced less offensive language, fewer errors, and reduced misinformation.

2022 年 1 月，OpenAI 对 GPT-3 进行了监督式训练的微调和迭代，最终发布了 InstructGPT。InstructGPT 更善于遵循人的指示，并且产生的冒犯性语言、错误信息和整体错误更少。

The Launch of ChatGPT (2022)

On November 30, 2022, OpenAI officially launched ChatGPT. A sibling model to InstructGPT, it was essentially a conversational version of GPT, based on the GPT-3.5 iteration. It could not only answer questions but also compose articles, write code, and mimic human conversational styles. Its seemingly omnipotent answering capabilities led to a全新认识 of the general capabilities of large language models. ChatGPT went viral on social media, amassing over 1 million registered users within just five days.

2022 年 11 月 30 日，OpenAI 正式发布了 ChatGPT。它是 InstructGPT 的姐妹模型，本质上是在 GPT-3.5 版本基础上开发的能够对话的 GPT 版本。它不仅能够回答问题，还能创作文章、编程，甚至模仿人类的对话风格，其几乎无所不能的回答能力使得人们对大语言模型的通用能力有了全新的认识。ChatGPT 迅速在社交媒体上走红，短短 5 天，注册用户数就超过 100 万。

Rapid Integration and GPT-4 (2023)

2023 witnessed explosive growth and integration:

February 2023: Microsoft announced the integration of ChatGPT across its product line (Bing, Office, Azure). OpenAI launched the paid subscription plan ChatGPT Plus.
March 2023: Microsoft integrated the rumored GPT-4 model into Bing and Edge. OpenAI officially launched GPT-4, a multimodal model capable of processing both text and image inputs. ChatGPT gained support for third-party plugins,解除 its inability to access the internet.
May 2023: OpenAI launched the iOS ChatGPT app.
Throughout 2023: Widespread adoption by companies (Snapchat, etc.) and significant regulatory attention began, notably from the European Data Protection Board (EDPB).

2023 年见证了爆炸性的增长和整合：

2023年2月： 微软宣布在其全线产品中整合 ChatGPT。OpenAI 发布了付费订阅计划 ChatGPT Plus。

2023年3月： 微软将 GPT-4 模型集成到必应及 Edge 浏览器中。OpenAI 正式推出多模态模型An AI model capable of processing and generating multiple types of data such as text, images, and audio. GPT-4。ChatGPT 宣布支持第三方插件，解除了其无法联网的限制。

2023年5月： OpenAI 推出 iOS 版 ChatGPT 应用。

整个2023年： 该模型被众多公司广泛采用，并开始受到监管机构（如欧洲数据保护委员会）的显著关注。

Continued Evolution: GPT-4o to GPT-5 (2024-2025)

The innovation pace continued to accelerate:

May 2024: OpenAI launched GPT-4o ("o" for omni), a model accepting any combination of text, audio, and image as input and output, aiming for more natural human-computer interaction. The ChatGPT desktop app was released.
2024-2025: Key developments included the launch of ChatGPT Enterprise, advanced voice modes, memory functions, Canvas for collaborative work, and deep integration with operating systems (Apple).
Mid-2025: OpenAI announced the upcoming GPT-5 model, signaling the next major leap. Subsequent releases included GPT-4.5 and the GPT-5.1 series, focusing on enhanced reasoning and conversational ability.
August 2025: OpenAI launched the more powerful GPT-5 model for coding and writing, making it available to free and paid users.
Late 2025: Milestones included ChatGPT reaching 800 million weekly active users, the introduction of an Apps SDK for building interactive apps within ChatGPT, and the rollout of group chat features.

创新步伐持续加速：

2024年5月： OpenAI 推出 GPT-4o（"o"代表全知），这是一个能接受文本、音频和图像任意组合输入并生成任意组合输出的模型，旨在实现更自然的人机交互。ChatGPT 桌面版应用程序发布。

2024-2025年： 关键进展包括发布 ChatGPT 企业版、高级语音模式、记忆功能、用于协作的 Canvas 工具，以及与操作系统（苹果）的深度集成。

2025年中： OpenAI 宣布即将推出 GPT-5 模型，预示着下一次重大飞跃。随后发布了 GPT-4.5 和 GPT-5.1 系列，专注于增强推理和对话能力。

2025年8月： OpenAI 推出了更强大的、适用于编码和写作的 GPT-5 模型，并向免费和付费用户开放。

2025年底： 里程碑包括 ChatGPT 周活用户突破 8 亿，推出了用于在 ChatGPT 内构建交互式应用的 Apps SDK，以及群组聊天功能的上线。

Core Technical Architecture and Operational Theory

Theoretical Foundation: The Rise of Large Language Models

Large models typically refer to machine learning models with a massive number of parameters. Based on a pre-training approach, they use Natural Language Processing (NLP) to understand and learn human language, completing content generation tasks such as information retrieval, machine translation, text summarization, and code writing through human-computer dialogue. The origin of large models can be traced back to the early days of AI research in the 20th century. With the advent of machine learning, deep learning technologies, and improvements in hardware capabilities, training on large-scale datasets and complex neural network models became possible, ushering in the era of large models. The Transformer architecture, introduced by Google in 2017, significantly enhanced sequence modeling capabilities through its self-attention mechanism. Subsequently, the concept of Pre-trained Language Models (PLMs) became mainstream. PLMs are pre-trained on massive text datasets to capture general language patterns and then fine-tuned for specific downstream tasks. OpenAI's GPT series is a paradigm of generative pre-trained models.

大模型通常指的是拥有巨大参数量的机器学习模型。这些模型基于预训练方式，通过自然语言处理来理解和学习人类语言，以人机对话方式完成信息检索、机器翻译、文本摘要、代码编写等内容生成任务。大模型的由来可以追溯到 20 世纪的 AI 研究初期。随着机器学习、深度学习技术的出现和硬件能力的提升，大规模数据集和复杂神经网络模型的训练成为可能，从而催生了大模型的时代。2017 年谷歌推出的 Transformer 模型结构通过引入自注意力机制，极大地提升了序列建模的能力。此后，预训练语言模型的理念逐渐成为主流。PLM 在大规模文本数据集上进行预训练以捕捉语言的通用模式，然后针对特定任务进行微调。其中，OpenAI 的 GPT 系列模型是生成式预训练模型的典范。

Operational Mechanism of ChatGPT

ChatGPT is built upon the GPT series of large models. OpenAI employed Reinforcement Learning from Human Feedback (RLHF) to train ChatGPT. During the initial model training, human trainers played both sides of a conversation to provide dialogue data as learning material, enabling the model to form logical reasoning chains that align with human language. When humans played the role of the chatbot, the model also generated suggestions to help trainers craft their responses. Additionally, ChatGPT incorporated training focused on ethical standards, refusing to provide effective answers to malicious prompts containing violence, discrimination, criminal intent, etc., based on pre-designed ethical guidelines.

ChatGPT 是基于 GPT 系列大模型构建的。OpenAI 采用“从人类反馈中强化学习”的训练方式对 ChatGPT 进行了训练。在训练原始模型时，人类训练师扮演对话的双方提供对话作为学习资料，使模型形成符合人类语言的逻辑依据和理解链条。在人类扮演聊天机器人时，模型也会生成建议来帮助训练师撰写回复。除此之外，ChatGPT 还采用了注重道德水平的训练方式，按照预先设计的道德准则，对含有恶意（包括暴力、歧视、犯罪等意图）的提问和请求拒绝提供有效答案。

Prerequisites for Operation

Massive Computing Power: Powerful computing capability, i.e., the ability to process and calculate vast amounts of data, is a crucial prerequisite supporting ChatGPT's training and operation. Microsoft built a supercomputing platform with tens of thousands of NVIDIA A100 chips to provide superior computing power for ChatGPT and the new Bing.
Training Foundation - Vast Data: ChatGPT's iterative progress relies on training with massive datasets. Its training utilized a vast corpus of text data, including nearly one trillion words collected from sources like books, magazines, Wikipedia, and forums, constructing an extensive language database.
Algorithmic Logic - Pre-training + Fine-tuning: In terms of algorithmic architecture, ChatGPT adopts the fundamental "pre-training + fine-tuning" paradigm. The most critical components are the Transformer-based generative pre-training model and the fine-tuning algorithm using RLHF. The Transformer architecture can better capture contextual relationships, generating content with logic and coherence, and supports parallel processing of multiple words, efficiently handling long text sequences.

运行前提

大算力： 强大的算力，即对海量大数据的计算和处理能力，是支撑 ChatGPT 训练和运行的重要前提。微软耗费上万张英伟达 A100 芯片打造超算平台为 ChatGPT 和新版必应提供更好的算力。

训练基础 - 海量数据： ChatGPT 的进步迭代离不开海量的数据训练。其训练使用了来自人类书籍、杂志、维基百科、论坛等渠道的庞大文本数据，构建了庞大的语料数据库。

算法逻辑 - 预训练+微调： 在算法架构方面，ChatGPT 采用"预训练+微调"的基本模式，其中最为关键的是基于 Transformer 的生成式预训练模型和基于人类反馈强化学习的微调算法。Transformer 架构可以更好地捕捉上下文联系，生成具有逻辑性和连贯性的内容，并支持多词并行处理，能有效生成长文本序列。

(Due to the substantial length of the original content, this post focuses on rewriting the Introduction, Development History, and Core Technical Architecture sections. The original content also includes extensive details on Main Functions, Social Impact, and Related Controversies, which follow a similar chronological and thematic structure.)