AI Large Language Models for Beginners: Core Concepts and Chinese Model Landscape

For beginners entering the AI field with no prior background, the hardest hurdle is often not the technology itself, but the confusion around fundamental concepts. Artificial intelligence, machine learning, deep learning, large language models — these terms are frequently used interchangeably, yet they form a clear hierarchy. This article is based on the first lesson of a beginner-friendly AI LLM application development course, covering the core concepts, historical milestones, and the current landscape of domestic Chinese AI models.

Clarifying Four Key Concepts: AI, Generative AI, Machine Learning, and Deep Learning

To truly get started with large models, you first need to distinguish four easily confused concepts: artificial intelligence, generative AI, machine learning, and deep learning.

Artificial Intelligence (AI) is fundamentally a goal, not a specific technology. It is an interdisciplinary field spanning computer science, statistics, linguistics, neuroscience, philosophy, and psychology — aimed at researching, designing, and building machines capable of learning, reasoning, and acting. Interestingly, why do some people dismiss today's AI as hopelessly dumb while others fear it will replace humanity? The root cause lies in differing expectations — some expect AI to be smarter than Einstein or as omnipotent as Doraemon, while pragmatists are satisfied if AI can solve a specific problem in a professional domain.

Generative AI (GAI) sets a more concrete goal: enabling machines to produce complex, structured content. Before it emerged, AI primarily handled tasks like classification, selection, and judgment — essentially a slightly smarter search engine that couldn't truly create "from scratch," such as independently writing an article.

Underpinning this leap is a fundamental paradigm shift from discriminative models to generative models. A discriminative model only needs to answer "is this a cat or a dog?", while a generative model must learn the underlying probability distribution of data in order to construct new, logically coherent content from scratch. The key architectural breakthrough enabling this was Transformer, proposed by Google in 2017. Its core mechanism — Self-Attention — allows a model to simultaneously attend to the relational weights of all positions in an input sequence, completely overcoming the bottleneck that recurrent neural networks (RNNs) faced with long-text processing, and laying the architectural foundation for all subsequent large language models.

Machine Learning and Deep Learning: Two Approaches to Achieving AI

If AI and generative AI represent the goals, then machine learning and deep learning are the means to achieve them. Understanding the distinction between the two is a crucial step in building a coherent mental model of AI.

Three Learning Paradigms in Machine Learning

Machine learning, as the foundational training approach, comprises three main methods:

• Supervised Learning: Provide labeled data. For example, feeding a model large quantities of images labeled "cat" or "dog" allows it to extract features and establish connections — like a teacher walking a student through every step of every example problem.
• Unsupervised Learning: Think of this as "self-study." The model receives unlabeled or even incomplete data and is trained through "completion" tasks — such as masking part of a sentence for the model to fill in, or adding a mosaic to an image for the model to reconstruct.
• Reinforcement Learning: Built on top of the previous two, this approach uses positive and negative feedback to constrain model behavior. The concept comes from education and psychology — much like how we both reward and correct children — guiding the model to continuously improve on similar problems.

In the era of large models, reinforcement learning has taken a more concrete form: RLHF (Reinforcement Learning from Human Feedback). This is the core mechanism behind ChatGPT's ability to follow instructions and produce meaningful responses. Human annotators first score the quality of different model outputs; these scores train a Reward Model; and then a reinforcement learning algorithm (such as PPO) is used to continuously optimize the language model's output strategy, aligning it more closely with genuine human preferences. DeepSeek R1 further explored a pure reinforcement learning approach (GRPO), significantly reducing reliance on human-annotated data and achieving industry-wide breakthrough results in mathematical reasoning and logical capability.

Deep Learning: What Does "Deep" Actually Mean?

Deep learning is a subfield of machine learning centered on neural networks that simulate the structure of the human brain. It's worth clarifying that "deep" here does not refer to the level of intelligence, but to the number of layers through which information is processed — the same dataset can yield different dimensional features at different layers.

For example: given the number sequence "123456," the first layer abstracts "six digits," the second notices the maximum is 6 and the minimum is 1, the third discovers it's an arithmetic sequence, and the fourth can even infer the next number is 7. This layer-by-layer accumulation is what "depth" truly means.

However, more layers don't always mean better results. Before residual connections were invented, deep networks suffered from a serious vanishing gradient problem: error signals would decay to near zero through multiple layers of back-propagation, making it nearly impossible to effectively update the parameters of earlier layers — causing deep networks to actually perform worse than shallow ones. The solution to this bottleneck was discovered by He Kaiming, introduced below.

From Deep Blue to DeepSeek: Key Milestones in Large Model History

Several unavoidable milestones mark AI's journey into the public consciousness.

Deep Blue defeating chess world champion Garry Kasparov (1997) wasn't technically mysterious — it used brute-force search algorithms to enumerate every possible move by the opponent, combined with programmer-defined evaluation functions for decision-making. Because the combinatorial possibilities in chess are finite, computers of that era could handle it. Soon, however, someone asked: what about Go? The answer was no — and AI entered nearly two decades of stagnation.

AlphaGo's victories over Lee Sedol and Ke Jie (2016–2017) relied instead on artificial neural networks in deep learning. It's worth noting that deep learning once faced a "depth collapse" problem: as layers accumulated, different inputs would converge to the same output after multiple processing stages, losing discriminative meaning. This bottleneck was overcome by Chinese scholar He Kaiming in his 2015 paper Deep Residual Learning for Image Recognition, which introduced Residual Connections. The core idea is to introduce "skip connections" between network layers that add the layer's input directly to its output (i.e., F(x)+x), allowing gradients to bypass deep layers and propagate directly backward — fundamentally solving the vanishing gradient problem and making it possible to scale network depth from dozens of layers to hundreds or even thousands. This directly established the engineering foundation of modern deep learning.

The birth of ChatGPT 3.5 brought generative AI to the general public. OpenAI's iteration from 1.0 to 3.5 validated a critical insight: the more data you feed a model, the more capable it becomes — confirming this was at least the right technical direction.

This phenomenon is backed by a theoretical framework known as Scaling Laws, systematically proposed by OpenAI researcher Kaplan and colleagues in 2020. The theory posits that model performance follows predictable power-law relationships with parameter count, training data volume, and computational resources — when all three scale together, model capability improves in a stable and predictable manner. Even more surprising is that when scale crosses certain thresholds, models can suddenly acquire capabilities they completely lacked before — called Emergent Abilities — such as multi-step logical reasoning and complex mathematical computation, which are entirely absent in smaller models.

The release of DeepSeek R1 signaled the start of a "hundred-model war" in the large model space. More importantly, its open-source strategy allowed developers who were previously locked out by technical barriers and team requirements to "stand on the shoulders of giants" — focusing solely on data feeding and fine-tuning to get a large model running. This greatly accelerated the flourishing of the domestic AI model ecosystem. The core value of open source lies in the fine-tuning ecosystem: developers can apply parameter-efficient fine-tuning techniques like LoRA and QLoRA to base models using their own domain-specific data, obtaining specialized models for vertical domains at minimal compute cost. This is one of the most important technical levers for application developers — a space also represented by Meta's open-source LLaMA model series.

The Chinese Large Model Landscape: A Thriving Competitive Field

The current domestic large model space is flourishing with diversity. Here is an overview of the major players:

• Alibaba Qwen (Qwen Max): Powerful and well-rounded, a popular choice among developers;
• Zhipu ChatGLM: Backed by Tsinghua University, with deep technical foundations;
• Baidu ERNIE Bot (Wenxin Yiyan): Industry critics say it "always gets up early but misses the market" — hitting every major technical milestone but failing to convert timing into a first-mover advantage;
• DeepSeek V3/R1: Developed by DeepSeek, its open-source impact has been far-reaching and it is a key driver of the domestic LLM ecosystem;
• Tencent Hunyuan, iFlytek Spark, and ByteDance Doubao (Volcano Engine) each hold their own ground and continue iterating;
• Moonshot AI Kimi K2 Thinking: A powerful model with comprehensive capabilities on par with Qwen Max at the top of the domestic rankings.

Internationally, Google Gemini, Anthropic Claude, and OpenAI ChatGPT continue to compete fiercely. For everyday users, more models mean lower prices and greater choice — an unambiguous industry dividend.

How Should Application Developers Position Themselves?

Here is an important distinction: developing AI applications is not the same as developing large models. Building large models requires deep mathematical expertise and low-level systems knowledge — the barrier is extremely high. The core work of an application developer is to treat large models as a "black box," focusing on the design and optimization of inputs and outputs, and "steering" the model to solve specific real-world problems. Just as you don't need to understand how an engine works to build something truly valuable with it.

Summary: The Hierarchy of Large Models in One Sentence

To summarize the relationship between these concepts in one sentence: AI and generative AI are the goals; machine learning is the means; deep learning is a more powerful means; and large language models (LLMs) are the practical, deployable product form that this evolutionary path has produced today. For learners riding the AI wave, understanding this hierarchy is the very first step on the path to large model application development.

Key Takeaways