在人类历史的大部分时间里，每一项认知任务——每一次计算、每一次翻译、每一次模式识别、每一个决策——都由人类大脑独立完成。 事实

革命 0 — 农业

大约一万年前，人类从狩猎采集社会转向定居农业。这从根本上改变了能量获取方式、人口密度和社会组织方式。它促成了城市、文字和专业化劳动的形成 ^[1]. 但它没有改变一个基本事实：所有智能仍然是生物性的。

第一次工业革命, beginning in the late 18th century, marked humanity's first large-scale substitution of mechanical power for human and animal muscle. James Watt's improved steam engine ^[2] became the symbol of an era that transformed production, transportation, and society.

第二次工业革命 brought electricity, the internal combustion engine, chemicals, and steel. It enabled energy to be distributed anywhere, and production to be scaled through the assembly line ^[3]. Henry Ford's moving assembly line dropped the time to build a Model T from 12 hours to 93 minutes.

The Third Industrial Revolution — the Digital Revolution — began with the transistor (1947), integrated circuit (1958), and microprocessor (1971). It made computation cheap, information instant, and communication global. 事实

For most of computing history, software followed a fundamental pattern:

This pattern — explicit rules written by humans — was extraordinarily powerful. But it meant that computers could only do what humans could precisely describe. 解读

In 1943, Warren McCulloch, a neurophysiologist, and Walter Pitts, a logician, published "A Logical Calculus of the Ideas Immanent in Nervous Activity" ^[4]. They proposed a mathematical model of a neuron: a binary threshold unit that could perform logical operations.

这是第一个正式描述：一个由简单计算单元组成的网络如何在原则上执行任何逻辑或算术运算。它是现代神经网络理论的种子。

Alan Turing's 1950 paper, "Computing Machinery and Intelligence" ^[5], opened with a direct question:

Rather than defining "thinking," Turing proposed the Imitation Game — later called the Turing Test — in which a machine would be judged intelligent if it could converse with a human interrogator indistinguishably from another human. 事实

值得注意的是，图灵预测到 2000 年，机器有 70% 的概率通过测试。这个预测没有实现——而且测试本身也被批评为衡量欺骗而非智能。 解读

In the summer of 1956, John McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester organized the Dartmouth Summer Research Project on Artificial Intelligence ^[6]. 这是第一个致力于机器智能研究的会议——它赋予了该领域名称。

最初的提案包含了一个极其大胆的假设：

事实 达特茅斯提案写于 1955 年，会议于 1956 年夏天在新罕布什尔州汉诺威的达特茅斯学院举行。它被广泛认为是 AI 作为正式研究领域的奠基事件。

Frank Rosenblatt introduced the Perceptron at the Cornell Aeronautical Laboratory in 1957 ^[7]. 这是一个可以通过调整连接权重来学习分类模式的系统——是 McCulloch-Pitts 神经元模型具备学习能力的直接实现。

感知机 created enormous excitement. The New York Times reported that the machine was "the first device to think as a human." Rosenblatt himself predicted that Perceptrons would eventually "be able to walk, talk, see, write, reproduce itself and be conscious of its existence." 解读

20 世纪 60 年代和 70 年代初见证了非凡的早期成就: Logic Theorist (often called the first AI program), ELIZA (a surprisingly convincing conversation simulator), SHRDLU (a blocks-world reasoning system), and the foundations of expert systems.

但乐观超出了现实。AI 系统在玩具领域有效，但在真实世界的复杂性中失败了。 The Lighthill Report (1973) in the UK was damning, leading to steep funding cuts. The first AI Winter had begun. 事实

The 1980s saw a resurgence of AI through expert systems — programs that encoded human expertise as explicit rules. 像 MYCIN（医学诊断）和 DENDRAL（化学分析）这样的系统展示了商业价值。 事实

But expert systems had a fundamental limitation: knowledge still had to be extracted from humans and encoded by hand. They worked well in narrow domains, but couldn't learn from data, couldn't generalize, and broke down at the boundaries of their encoded rules. 解读

In 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams published "Learning representations by back-propagating errors" ^[8]. They demonstrated a method by which neural networks could learn internal representations by propagating error signals backward through the network.

Backpropagation — combined with gradient descent — became the foundation of virtually all modern deep learning. While earlier work had explored the idea (Paul Werbos, 1974), the 1986 paper made it practical and influential. 事实

In May 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov in a six-game match ^[9]. It was a powerful demonstration of machine capability in a domain considered a pinnacle of human intellect.

However, Deep Blue was not machine learning in the modern sense. It used brute-force search — evaluating up to 200 million positions per second — with specialized hardware and hand-tuned evaluation functions. It was a triumph of engineering, but the paradigm was fundamentally different from what would later define modern AI. 解读

The 1990s and 2000s saw a fundamental shift in AI. Instead of hand-coding rules, researchers began building systems that learned patterns from data. This was the rise of statistical machine learning. 事实

关键发展包括支持向量机（SVM）、贝叶斯网络、决策树和集成方法（随机森林、梯度提升），以及自然语言处理和语音识别的统计方法。 事实

In 2006, Geoffrey Hinton and Ruslan Salakhutdinov published "Reducing the Dimensionality of Data with Neural Networks" ^[10], 证明了深度神经网络可以通过逐层预训练学习有用的表示。这重新点燃了对神经网络的兴趣。

三股汇聚的力量推动了深度学习革命：

The 2012 ImageNet Large Scale Visual Recognition Challenge (ILSVRC) was a watershed moment. Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton submitted AlexNet ^[11] — a deep convolutional neural network trained on GPUs.

AlexNet achieved a top-5 error rate of 15.3%, compared to the second-best entry at 26.2%. The gap was so large that it signaled a paradigm shift in computer vision — and in AI more broadly. 事实

AlexNet used two NVIDIA GTX 580 GPUs, had 62 million parameters, and was trained for 5-6 days. The use of GPUs for general-purpose computing (CUDA, 2006) was critical to this breakthrough. 事实

In 2014, Ian Goodfellow and colleagues introduced Generative Adversarial Networks (GANs) ^[12]. The idea was elegantly simple: two neural networks — a generator and a discriminator — compete in a game, with the generator learning to produce increasingly realistic outputs.

That same year, Kingma and Welling introduced Variational Autoencoders (VAEs) ^[13], providing a different approach to generative modeling. Together, these works began the shift toward generative AI — systems that don't just classify or predict, but create. 解读

In March 2016, DeepMind's AlphaGo defeated Lee Sedol, one of the greatest Go players in history, 4-1 in a five-game match ^[14]. Go had long been considered a grand challenge for AI due to its enormous branching factor (~250 possible moves per turn, vs. ~35 for chess).

第 37 手 — played in Game 2 — was a move that no human professional would likely have chosen. It appeared to be a mistake, but it turned out to be a brilliant strategic play. This moment became a cultural symbol of machines doing something that looks like creativity. 解读

Whether this constitutes "creativity" is a philosophical question. What is factually clear: AlphaGo made a move that expert human players found surprising and later recognized as strategically valuable. 该系统结合了策略网络、价值网络和蒙特卡洛树搜索。 事实

In June 2017, Vaswani et al. from Google Research published "Attention Is All You Need" ^[15]. It proposed the Transformer architecture — a neural network design based entirely on attention mechanisms, without recurrence or convolution.

This paper is one of the most influential in AI history. As of 2024, it has been cited over 100,000 times. Transformer is the foundation of virtually every major modern AI system: GPT, BERT, Claude, Gemini, Llama, Stable Diffusion, 等。 事实

自注意力如何工作

Each word in a sentence is transformed into three vectors: Query, Key, and Value. The model computes attention scores by comparing each word's Query with every word's Key. These scores determine how much each word should "attend to" every other word. The weighted sum of Values produces the output.

Before a language model can process text, it must first convert words into numbers. This is done through tokenization — splitting text into tokens (words, subwords, or characters) and mapping each to an ID. 事实

Common tokenization methods include Byte-Pair Encoding (BPE), WordPiece, and SentencePiece. The exact method varies by model. GPT models use BPE, while BERT used WordPiece. 事实

In 2020, Kaplan et al. published "Scaling Laws for Neural Language Models" ^[16], 表明语言模型的性能与模型大小、数据集大小和计算量之间存在平滑的幂律关系。 更大的模型，在更多数据上训练，使用更多计算，始终表现更好。

这一发现具有深远影响：它表明规模扩展没有根本性的天花板——至少在所衡量的指标上是如此。 结果是竞相建造越来越大的模型。 解读

事实 Kaplan 等人发现，性能（以交叉熵损失衡量）与模型大小、数据大小和计算量呈幂律关系。 论文建议，要最优地改进模型，模型大小和数据大小应该一起扩展——而不是单独扩展其中一个。

OpenAI's GPT (Generative Pre-trained Transformer) series has been one of the most influential lines of AI models:

但 AI 历史不仅仅是 OpenAI 的历史。其他重要的模型家族包括： BERT (Google, 2018), PaLM (Google, 2022), Claude (Anthropic, 2023), Llama (Meta, 2023), Gemini (Google DeepMind, 2023), DeepSeek (2024), 等。 事实

On November 30, 2022, OpenAI released ChatGPT ^[21] — a dialogue-optimized language model trained with reinforcement learning from human feedback (RLHF). 它是历史上增长最快的消费应用，据报道在发布两个月内达到了 1 亿用户。

ChatGPT 标志着 AI 从专业研究领域走向主流文化现象的时刻。 It was not the most capable model (GPT-4 was released four months later), but it was the first to make AI capabilities directly accessible to the general public. 解读

A large language model (LLM) processes text through a sequence of transformations:

At each step, the model predicts the next token — the most probable continuation of the text it has seen so far. 这本质上是一个基于数万亿 token 文本训练的下一个 token 预测机器。 事实

The model does not "understand" in the human sense. It computes statistical patterns in a high-dimensional space. Whether these patterns constitute "understanding" is a matter of active philosophical and scientific debate. 解读

Pretraining

模型在大量文本语料（互联网、书籍、代码）上训练，以预测下一个 token。 这是模型学习语法、事实、推理模式和世界知识的地方。 The resulting compressed knowledge is stored in the model's weights.

Post-Training

预训练后，模型进行额外训练以与人类偏好对齐：

现代 AI 远不止于对话式智能体。相同的基础架构——Transformer——已被迁移到一系列令人瞩目的领域：

In 2021, DeepMind's AlphaFold2 ^[23] achieved near-experimental accuracy in protein structure prediction — a problem that had challenged scientists for 50 years. The system achieved a median GDT score of 92.4 on the CASP14 targets, approaching the accuracy of experimental methods like X-ray crystallography.

AlphaFold has been used by over 2 million researchers worldwide and has been recognized as a transformative contribution to biology. The 2024 Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper of DeepMind for their work on AlphaFold. 事实

AI is increasingly used as a scientific instrument across domains:

AI training compute has grown at an extraordinary rate. According to Sevilla et al. (2022) ^[24], the compute used to train notable AI systems doubled approximately every 3.4 months from 2010 to 2022 — far outpacing Moore's Law's doubling time of ~2 years.

This growth has been enabled by a combination of: more powerful GPUs, distributed training across thousands of accelerators, improved networking, algorithmic efficiencies, and massive investment. 解读

神经网络的复兴是由硬件革命推动的。 Graphics Processing Units (GPUs), originally designed for rendering 3D graphics, proved remarkably efficient for the matrix operations at the core of deep learning.

NVIDIA's CUDA platform (2006) allowed GPUs to be programmed for general-purpose computing. The company's V100 (2017), A100 (2020), H100 (2022), and B200 (2024) GPUs have become the backbone of AI training and inference. Google developed its own Tensor Processing Units (TPUs) in 2016. 事实

The cost of AI inference has fallen dramatically. The price per token for state-of-the-art language models has decreased by orders of magnitude since GPT-3 was introduced in 2020.

Training costs have also evolved. GPT-3's training cost was estimated at approximately $4.6 million. More recent large models are estimated to cost tens to hundreds of millions of dollars to train. However, algorithmic improvements and architectural innovations (like Mixture-of-Experts) have improved efficiency. 解读

Benchmarks are essential tools for tracking AI progress — but they are not measures of general intelligence. Each benchmark measures performance on a specific set of tasks, under specific conditions, with specific metrics. 事实

Known limitations of benchmarks include:

解读 A model scoring 90% on MMLU does not mean it is "90% of human intelligence." It means it answered 90% of the questions in that specific benchmark correctly. The gap between benchmark performance and real-world reliability remains significant.

The most recent paradigm shift in AI is the move from simple question-answering to agentic behavior — systems that can take multi-step actions toward a goal, using tools and adapting to feedback.

Agent capabilities include: tool use (calculators, search, APIs), code execution, browser manipulation, file system operations, memory management, and multi-step planning. However, current agents remain unreliable for long-horizon tasks, and suffer from error accumulation, hallucination, and security vulnerabilities. 解读

The convergence of AI with robotics brings intelligence into the physical world. Vision-Language-Action models combine perception, reasoning, and motor control. Research in autonomous vehicles, robot manipulation, and humanoid robots is advancing rapidly.

However, physical world capabilities lag significantly behind digital capabilities. Household robots, fully autonomous driving, and general-purpose manipulation remain research challenges rather than solved problems. 事实

AI appears to be software, but it rests on a massive physical infrastructure: data centers, networking, semiconductors, electricity, water for cooling, rare earth minerals, and global supply chains. 事实

训练大型 AI 模型消耗大量能源。 GPT-3's training is estimated to have consumed approximately 1,300 MWh ^[25]. 数据中心能耗正在快速增长，国际能源署的预测显示未来几年将大幅增加。 事实

AI's economic impact is the subject of intensive research. Key findings include:

AI 系统带来了一系列需要仔细考虑的风险。 这些风险从直接的技术问题到长期的社会挑战。

解读 AI 风险的严重程度和概率正在积极讨论中。 本网站呈现了各种观点，不认可任何单一立场。 灾难性风险立场在 AI 研究者中是少数观点，但被一些领先的研究者和组织认真对待。

AI 发展是一个全球现象，不限于美国。 世界各地都存在主要的研发中心：

事实 根据斯坦福 AI 指数（2024 年），美国在 AI 模型、投资和出版物方面领先，但中国在 AI 论文和专利数量上领先。 全球 AI 格局正日益多极化。

人类历史可以被视为人类能力逐步外部化到机器和系统中的过程。 论点

最后五行——感知、语言、模式识别、编程、推理——代表了 AI 开始外部化的能力。 问号表示这个过程是不完整的、不均衡的，且尚未完全可靠。 解读

Andrej Karpathy popularized the concept of Software 2.0 — software that learns its behavior from data rather than being explicitly programmed. 这种演进可以被视为人类与机器关系的一种进步。 解读

Machines used to execute our instructions.

Now they increasingly infer what the instructions require.

And sometimes generate the intermediate steps themselves.

Throughout history, technology has been a tool — a passive instrument that amplifies human capability. AI represents a shift toward technology becoming a cognitive collaborator — an active participant in knowledge work. 论点

论点 The "Future" column is labeled as a scenario/hypothesis. 真正自主的认知系统——能够独立追求长期目标的系统——尚未以可靠、通用的形式存在。 当前的 AI 系统仍然依赖于人类基础设施、人类提供的目标和人类监督。

一个有用的框架：每次技术革命都使以前稀缺的资源变得更加丰富。 论点

前三行——体力工作、信息获取、预测——已经在被 AI 改变。 后三行——分析、推理、专业知识——部分受到影响，但仍然是人类认知保持显著优势的领域。 The "?" indicates that the trajectory is not yet clear. 解读

没有单一、普遍接受的智能定义。不同领域对其定义不同：

For the purposes of this site, we use an operational framework:

解读 这个定义是一个工作框架，而非哲学主张。 选择它是为了使分析可行，而非解决关于智能本质的深层哲学问题。

本网站收集的证据表明：

It is essential to distinguish between the claim that AI is beginning to mechanize cognition and the claim that AI has achieved general intelligence. 当前的 AI 系统有严重的局限性：

论点 The "智能革命" thesis does not require that AI has achieved human-level general intelligence. It only requires that machines are beginning to perform cognitive tasks that were previously exclusively human — and that this represents a categorical shift in what machines can do.

This is why AI may be more than
the Fourth Industrial Revolution.

It may be
第一次 智能革命.

AI 是否真正成为 "第一次 智能革命" 不仅仅取决于模型参数或基准测试分数。它取决于：

查看完整参考书目 →      改变 AI 的论文 →      AI 人物 →