AI, Machine Learning, Deep Learning, and Large Models: Relationships and Evolution

Many people new to AI share a common confusion: what exactly is the relationship between artificial intelligence, machine learning, deep learning, large models, and generative AI? Some assume ChatGPT represents all of AI; others think these are completely separate technologies. In reality, there's a remarkably clear evolutionary thread connecting them — progressing from the broad to the specific, layer by layer. This article, based on content from Bilibili creator Jinze's introductory course on large models, systematically maps the key milestones in AI development to help beginners build a solid conceptual framework.

What Is Artificial Intelligence: A Field, Not a Product

Artificial Intelligence (AI) is the study of enabling computers or machines to learn, understand, reason, and act in ways similar to humans. Two key concepts matter here: machines — the subject of study includes programs, software, systems, or robots; and intelligence — a concept that has been debated at the heart of AI research for decades.

By science-fiction standards, machines would need autonomous consciousness and the ability to think like — or surpass — humans. We're clearly not there yet. But from a more pragmatic engineering perspective, if a machine can complete tasks that previously required human effort, we can reasonably say it has achieved some degree of intelligence. Voice assistants that understand commands, maps that plan routes — these are all real-world AI applications.

AI is therefore not a specific product but an entire technical field, spanning computer science, mathematics, statistics, linguistics, and even neuroscience. The reason people hold such vastly different views on AI comes down to differing expectations of what "intelligence" means. For learners, the more useful question is: Can AI help me solve real problems?

From Deep Blue to Machine Learning: The Limits of Rule-Based AI

Two landmark events define different eras of AI development. In 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov. In 2016, AlphaGo defeated Go champion Lee Sedol. These events represent fundamentally different stages in AI.

Deep Blue worked in a way completely unlike today's large models. It relied primarily on search algorithms: analyzing vast numbers of possible board positions in advance and selecting the move with the highest winning probability based on predefined evaluation rules. This worked for chess because the rules are explicit — every piece's movement and its consequences can be defined in advance.

But many real-world tasks don't have clear rules. Judging whether an image contains a cat, or analyzing the emotional tone of an article — these are nearly impossible to solve by enumerating rules ahead of time. The traditional AI that Deep Blue represents follows a core logic: "humans tell the machine the rules, and the machine computes accordingly." As problems grew more complex and we could no longer even define the right rules, AI entered a new phase — machine learning.

The Three Core Paradigms of Machine Learning

The central idea of machine learning is this: instead of telling machines all the answers, let them discover hidden patterns from large amounts of data. It encompasses three main learning approaches:

• Supervised Learning: Provide the machine with "questions and answers." For example, labeling images as "cat" or "dog" so the model learns the relationship between images and labels, and can then predict new images. Spam filtering and financial risk control both use this approach.
• Unsupervised Learning: Provide data without labels and let the machine find patterns on its own. For instance, feeding in large amounts of user behavior data and letting the model automatically cluster groups like "tech product enthusiasts" or "travel lovers."
• Reinforcement Learning: Adjust behavior through trial, error, and feedback. Like a child learning to walk — stumble and correct, succeed and reinforce — machines gradually learn goal-aligned behavior through reward mechanisms.

It's worth noting that large model training draws heavily on these ideas. For example, after initial training, ChatGPT is further refined through RLHF (Reinforcement Learning from Human Feedback) — teaching the model not just what the answer is, but how to respond in a better and safer way.

Deep Learning and Transformer: The Technical Foundation of Modern Large Models

As the information demands of image recognition, speech understanding, and natural language processing surged, traditional machine learning hit its limits. This gave rise to an important branch — deep learning — whose core technology is the neural network.

Neural networks are computational models inspired by how neurons connect in the human brain. They link large numbers of computational nodes together, processing information through multiple layers: when recognizing an image, the first layer learns edges and colors, the second learns shapes and structures, the third learns object composition. The more layers, the richer the levels of information the model can learn. To address information loss in deep networks, researchers introduced residual connections, allowing information to skip certain layers and pass directly forward — a foundational technique in modern deep learning.

Attention Mechanisms: The Breakthrough That Changed Everything

In 2017, a research team at Google published the paper Attention Is All You Need, introducing the Transformer architecture. GPT, Claude, and many other large models we know today are all built on Transformer or its variants.

The core problem Transformer solves is enabling a model to flexibly focus on the most relevant parts of its input — this is the attention mechanism. Consider the sentence: "Apple released a new phone, and it has been well received by many users." Humans easily understand that "it" refers to the new phone, but traditional computers treat it as just another character. Attention mechanisms help models attend to surrounding context when processing any given word, establishing accurate semantic connections.

Equally important: unlike traditional Recurrent Neural Networks (RNNs), which tend to "forget" earlier content when processing long texts, Transformer can analyze multiple parts of a text simultaneously, making it possible to handle massive, longer inputs. Researchers then discovered that simply scaling up model size, training data, and compute consistently improved model capability — and this became the foundation for the large model explosion.

Generative AI and Large Language Models: From Judgment to Creation

Large Language Models (LLMs) are AI systems trained on massive text datasets that can understand and generate natural language. Looking back at AI history, there's a clear shift: earlier AI was mostly doing classification and judgment (is this email spam? is this image a cat?), while the generative AI that has risen in recent years can create content from scratch.

Getting a computer to write an essay used to be extraordinarily difficult — it requires understanding the topic, organizing structure, and choosing appropriate expression. Today's large models can generate complete articles, code, and images from simple prompts. The explosion of ChatGPT 3.5 in November 2022 was the result of three forces converging: data accumulation, compute breakthroughs, and architectural innovation.

How Should Everyday Learners Position Themselves

For those looking to enter the AI application space, there's no need to fixate on mastering the underlying mathematics and model training — that's the domain of foundational researchers. A more practical path is to leverage the general capabilities of existing large models and apply them to specific business contexts: building enterprise knowledge bases, developing automated Agent systems, and freeing people from repetitive work so they can focus on goal-setting, decision-making, and creative tasks.

Just as the invention of the car didn't require most people to study engine mechanics — it just required learning to drive to get around more efficiently.

Comparing Leading Large Models and Future Trends

Today's leading large models each have their strengths:

• OpenAI GPT series: The most widely influential, covering text generation, code assistance, Q&A, data analysis, and more.
• Anthropic Claude: Known for long-document analysis and logical reasoning; particularly popular among developers who write code.
• Google Gemini: A multimodal model capable of handling text, images, audio, and video.
• Chinese large models: Alibaba's Qwen (open-source), Baidu's ERNIE series, and the widely discussed DeepSeek, among others.

There's no absolute standard for choosing a model — it depends on your task type, usage environment, and goals. For writing practice, mainstream general-purpose models work well. For software development, Claude or Gemini are strong choices. For enterprise scenarios, factors like data security, private deployment, cost control, and system integration all need to be weighed.

Two clear directions are shaping the future: first, continued improvement in model capabilities — including reasoning, code generation, and multimodal processing; second, specialization by industry, with purpose-built models for finance, law, medicine, and other professional domains. Agents (AI agents) will likely become the most important application paradigm in the coming years — capable not just of answering questions, but of understanding goals, making plans, calling tools, and executing tasks in a complete intelligent loop. For example, ask an Agent to analyze company sales performance, and it can automatically pull data, analyze trends, generate a report, and offer recommendations.

Tracing this evolutionary arc — from rules to machine learning to deep learning to Transformer to generative AI — is the essential first step for every large model learner. Once you understand the hierarchical relationships between these concepts, everything that follows — whether prompt engineering, fine-tuning techniques, or Agent development — becomes far easier to navigate.