Large language models (LLMs)—a simple introduction

Generative artificial intelligence and LLMs

In the context of solutions like ChatGPT, the term generative artificial intelligence (generative AI or GenAI) is frequently used. It’s therefore worth clarifying how this term relates to LLMs and what distinguishes the two.

Generative AI is a broad concept encompassing an entire class of artificial intelligence models capable of creating new content, regardless of the format. This includes systems that generate text, code, images, music, and even video. Thus, one could say that GenAI acts as an umbrella term covering various model types that share a common trait: the ability to create, rather than merely analyze.

Within this broad set of technologies, LLMs represent a specialized subgroup focused specifically on NLP. LLMs are designed and trained specifically to generate text that is coherent, grammatically correct, and semantically meaningful. They learn not only the rules of syntax and punctuation but also style, context, language registers, and subtle nuances of expression. This makes their output fluid, logical, and contextually appropriate, and often indistinguishable from human-written text.

A significant milestone for generative AI was the introduction of the GPT-3 model, which gained widespread recognition due to its capabilities in conversation, essay writing, and translation. Its successor, GPT-4, has advanced even further. This multimodal model can process not only textual data but also images as input. Consequently, GPT-4 can understand sentences as well as pictures, charts, and illustrations, significantly expanding its applications.

In summary, every LLM is a form of generative AI, but not all generative AI models are LLMs.

Training of large language models

The training process for an LLM comprises three main stages: pretraining, fine-tuning, and alignment. Additionally, precise control of a model’s behavior can be achieved through carefully structured prompts—a technique known as prompt engineering.

Pretraining

Pretraining constitutes the foundational stage in the development of an LLM. During this phase, the model learns to understand natural language and how to generate text. It is at this stage that AI absorbs general linguistic patterns, semantic rules, syntax, and basic knowledge about the world, enabling effective performance on tasks like translation, writing, summarizing, or conversation.

Models are trained on massive datasets—counted in trillions of tokens—and can possess hundreds of billions or even trillions of parameters. This training process relies on self-supervised learning (also known as unsupervised learning), where models predict subsequent words or text fragments based on context. Through this, models build internal representations of language and world knowledge, simultaneously developing capabilities for logical reasoning, programming, and understanding complex instructions. The larger the dataset and parameters, the more frequently emergent abilities appear, including unexpected abilities such as step-by-step reasoning, emotional interpretation, or explanation of phenomena.

Since 2023, leading model developers have adopted a multimodal approach, integrating different data types even at the pretraining stage. Modern models simultaneously learn from text, images, audio, and video data. Examples include Google Gemini and GPT-4o, which form common conceptual representations regardless of the input data format. In 2025, nearly all top LLMs are capable of processing or generating multimodal content, significantly expanding their applications, such as visual code analysis, meme interpretation, and synthesizing audiovisual descriptions. Multimodality has thus become the new standard, marking a departure from earlier generations trained exclusively on textual data.

In tandem with increasing model scales and data volumes, pretraining costs have soared into tens of millions of dollars per training run. Consequently, the industry seeks methods that are more efficient, both economically and environmentally. Techniques such as coreset selection, involving choosing smaller, representative subsets of data, have become increasingly popular. These allow comparable quality training at reduced costs. Another alternative is using synthetic data generated by previous LLM models, a method known as self-feeding or knowledge distillation. Such approaches reduce reliance on continually sourcing billions of tokens from the Internet while maintaining training quality.

As data volumes expand, awareness about their qualitative impact on model effectiveness also grows. Experiences with earlier LLM versions, which exhibited hallucinations or biases, have led to stricter data selection criteria. GPT-4, for example, was trained on carefully curated datasets comprising high-quality Internet content and technical documentation. Currently, data quality is becoming as important as quantity. Processes for filtering, validation, and ensuring linguistic diversity are increasingly emphasized, improving the alignment of models with real-world tasks and reducing the risk of generating inaccurate information.

Fine-tuning

Although pretraining allows an LLM to develop a general understanding of language, it isn’t enough to enable the execution of specialized tasks with the required precision. Therefore, the next stage is known as fine-tuning, a process of adapting the model to specific applications. This involves adjusting model parameters using new datasets that reflect scenarios and goals particular to a given application. Examples include datasets for recognizing emotions in texts or generating emails consistent with a specified communication style.

In fine-tuning, the quality of data is crucial. Training datasets must not only be extensive—often encompassing thousands of carefully prepared examples—but also appropriately cleaned, formatted, and matched to real-world usage scenarios. In sentiment analysis, it’s important to cover a broad spectrum of emotions and sources. For email generation, the precise representation of tone, linguistic register, and typical sentence structures is vital. The better constructed the dataset, the greater the likelihood the model will perform accurately, consistently, and in the desired style.

Today, fine-tuning has become a standard practice in deploying LLMs, both open-source models and commercial environments. Major providers such as OpenAI offer the ability to fine-tune even proprietary models, like GPT-3.5 Turbo or GPT-4o, via API interfaces. This enables companies to build personalized conversational assistants, supporting customer service, document analysis, or daily operational tasks.

Methods of parameter-efficient fine-tuning (PEFT), such as low-rank adaptation (LoRA), are increasingly popular. These significantly reduce computational costs by training only small, additional matrices, which are then integrated into an existing model. As a result, even 70-billion-parameter models can be fine-tuned on a single GPU, making fine-tuning feasible in local, academic, and resource-limited settings.

Alignment

Following the fine-tuning stage is alignment, the aim of which is to adjust a model’s behavior according to specific ethical norms and human values. In practice, this means teaching the model not only “what” to say but also “how” and “why.” One commonly used method is reinforcement learning from human feedback (RLHF), where human evaluators provide positive signals to reinforce desired responses and shape preferred behavioral patterns.

The alignment process also uses more formalized techniques, such as direct policy optimization (DPO) and knowledge transfer optimization (KTO), which precisely define the ethical framework within which the model operates. Increasingly, the mechanism of self-play, where the model engages in dialogue with itself, is employed to further enhance decision-making capabilities and consistency in complex contexts.

Such rigorous alignment is particularly important in sensitive applications like medicine, legal advice, or customer service, where incorrect or unethical responses can have tangible consequences. Thus, technical safety alone is insufficient—predictability, transparency, and alignment with societal values become crucial.

In recent years, constitutional AI has gained prominence. This is a method where the model does not solely rely on human judgment “case by case” but learns from a set of explicit principles. Such rules might include that the model ” should not offend,” “should justify its refusals,” or “should operate transparently.” Based on these principles, an AI generates self-criticism, evaluates its own outputs, and modifies them to better align with the set objectives. This entire process is supported by other models that assess which version of an answer aligns best with the accepted “constitution.” This approach, known as reinforcement learning from AI feedback (RLAIF), enables alignment to scale with minimal human involvement, reducing toxicity, logical errors, and unjustified refusals.

A unique challenge for modern LLMs involves very long conversational contexts, characteristic of models such as GPT-4o or Claude 3. In extended dialogues, models may exhibit sycophancy, blindly agreeing with the user even if this contradicts previously established guidelines. To prevent this, systems are trained to maintain a consistent “character,” for example that of a helpful, honest, and harmless assistant, even during multi-stage conversations involving tens of thousands of tokens.

In parallel, mechanisms enhancing operational transparency are advancing. Modern models can not only refuse to fulfill a request but also explain the reason behind the refusal, for example, by stating: “I cannot fulfill this request because it violates principle X.” This introduces an element of accountability and strengthens user trust. Techniques like interpretable chain-of-thought are emerging, where the LLM generates an intermediate reasoning process that is subsequently verified by a separate “critical” component, improving the factual and ethical alignment of responses.

Industry standards and regulations increasingly influence alignment as well. Major model providers, such as OpenAI, Anthropic, Google, and Meta, commit to safety testing their systems before public release. Implementations include marking AI-generated content, an “off-switch” for autonomous agents, and systems for auditing generated responses. Alignment thus evolves from purely technical to a response to rising social expectations and regulatory requirements, particularly in regulated sectors.