Large Language Models: How They Actually Work

What is a Large Language Model?

A Large Language Model (LLM) is a Transformer-based neural network trained on enormous quantities of text to predict the next token in a sequence. That is the entire training objective. From this deceptively simple task, these models develop a remarkable range of capabilities: they write, summarise, translate, reason, code, and converse.

"Large" refers to two things simultaneously: the number of parameters in the model (ranging from billions to trillions) and the volume of training data (ranging from hundreds of billions to trillions of tokens of text). Both scale dimensions matter significantly to performance, and the relationship between scale and capability has been studied carefully.

What makes LLMs distinct from earlier language models is the combination of Transformer architecture, massive scale, and a training approach that requires no hand-labelled data at all. The supervision signal comes entirely from the natural structure of text: given the previous words in a sentence, the next word is the label. This means the entire written record of humanity, once digitised, becomes a training dataset.

Tokens: The Language of LLMs

LLMs do not process text character by character or word by word. They process tokens: subword units produced by a tokenisation algorithm. A token is roughly 4 characters of English text on average, though this varies significantly by language and content type.

The dominant tokenisation method is Byte-Pair Encoding (BPE), introduced by Sennrich, Haddow, and Birch (2016) for machine translation. BPE works by starting with individual characters and iteratively merging the most frequent adjacent pairs until reaching the desired vocabulary size. Common words become single tokens. Rare words are split into recognisable subword pieces. This allows models to handle any text including invented words or misspellings, since any string can be decomposed into the byte-level characters that form the base of the vocabulary.

Tokenisation also explains why LLMs can struggle with certain tasks that seem easy to humans. Counting letters in a word is hard because the model sees token-level units, not characters. Arithmetic is hard because numbers are symbol sequences, not magnitudes. Rhyming is hard because the model does not hear sounds, it sees token identities.

Pre-Training at Scale: The Next-Token Prediction Game

The pre-training objective is called causal language modelling: given all previous tokens in a sequence, predict the next one. This is done for every position in every sequence in the training data, simultaneously, using the Transformer decoder architecture with causal masking (as covered in Lesson 4.4).

GPT-3 was trained on approximately 300 billion tokens drawn from Common Crawl (web text), WebText2 (curated web), Books1, Books2, and Wikipedia. GPT-4's training data composition was not publicly disclosed. The key insight of pre-training is that predicting natural language is an enormously information-rich task. To predict the next word in a sentence accurately, you need some model of syntax, semantics, world knowledge, common sense, cause and effect, and much more. The training signal pushes the model to encode all of this implicitly in its parameters.

Scaling Laws: Why Bigger (Often) Means Better

In 2020, a team at OpenAI (Kaplan et al.) published a landmark paper showing that LLM performance on language modelling tasks follows remarkably clean power law scaling with model size (parameters), dataset size (tokens), and compute budget. Doubling the model size on the same data improves performance in a predictable, measurable way. This allowed researchers to forecast, before training, how well a model would perform.

In 2022, DeepMind researchers (Hoffmann et al., "Training Compute-Optimal Large Language Models," commonly called the Chinchilla paper) refined these laws. They found that many existing large models, including GPT-3, were significantly undertrained: they had many parameters but not enough training tokens for those parameters. The optimal strategy, they argued, is to scale parameters and training tokens roughly equally. Chinchilla (70 billion parameters, 1.4 trillion tokens) outperformed GPT-3 (175 billion parameters, 300 billion tokens) on many benchmarks, despite being four times smaller.

RLHF: From Raw Base Model to Helpful Assistant

A pre-trained language model is extraordinarily knowledgeable, but it is also unruly. Ask it a question and it may respond by generating more questions, writing fiction, continuing the sentence in unexpected directions, or outputting harmful content. It has learned to model all of human text, including text that is unhelpful, dishonest, or dangerous.

Converting a base model into a useful, safe assistant requires additional training. The dominant technique is Reinforcement Learning from Human Feedback (RLHF), applied at scale by the InstructGPT paper (Ouyang et al., OpenAI, 2022) and the technology that powers ChatGPT.

A simpler alternative to RLHF called Direct Preference Optimisation (DPO) was proposed by Rafailov et al. (2023). DPO reframes the RLHF objective as a supervised classification problem, eliminating the need for a separate reward model training stage and the complexity of RL. Many open-source models released since 2023 have used DPO or variants of it.

Anthropic developed a complementary approach called Constitutional AI (Bai et al., 2022), which uses a set of principles (the "constitution") to guide the model to critique and revise its own outputs. This uses AI feedback (RLAIF) in place of some human feedback, making the process more scalable.

Hallucination: Why LLMs Confidently Say False Things

LLMs sometimes state things that are completely fabricated, with the same tone and apparent confidence they use for accurate information. They cite papers that do not exist, give wrong dates for real events, invent biographical details for real people, and produce code that looks correct but does not run. This phenomenon is called hallucination.

Understanding why it happens requires revisiting what LLMs actually do. At every step, the model computes a probability distribution over the vocabulary and samples a token. It does not "consult" a database of facts. It generates text that is consistent with the patterns in its training data. When asked about something it learned well, those patterns produce accurate-looking and accurate text. When asked about something rare, absent, or contradictory in its training data, those same patterns produce fluent but potentially incorrect text.

Mitigation: Retrieval-Augmented Generation (RAG)

The most widely deployed technical solution is Retrieval-Augmented Generation (RAG), introduced by Lewis et al. (Facebook AI Research, 2020). Instead of generating answers purely from the model's parametric memory, RAG first retrieves relevant documents from an external knowledge base (a database, the web, a document collection) and includes them in the model's context. The model then generates its answer grounded in the retrieved evidence.

This does not eliminate hallucination entirely, but it gives the model accurate, up-to-date source material to work from, and it makes verification easier because the sources can be cited. The search systems used by ChatGPT, Bing Chat (Copilot), and Perplexity AI all use versions of this approach.

Context Windows and "Memory"

A context window is the maximum number of tokens the model can process in a single forward pass, including both the input (your prompt, any retrieved documents, conversation history) and the output (the response). Anything outside the context window is simply not seen by the model. LLMs have no persistent memory between separate conversations unless it is explicitly engineered.

The computational cost of the Transformer's attention mechanism scales quadratically with sequence length (O(n²)), which is why very long context windows are technically challenging and expensive. Research into more efficient attention mechanisms (sparse attention, linear attention, state-space models like Mamba) is active precisely because context length is an important capability bottleneck.

The LLM Landscape

As of mid-2025, the LLM landscape consists of a small number of frontier closed-source models and a growing ecosystem of high-quality open-weight models.

Key Takeaways

• LLMs are Transformer decoder networks trained to predict the next token in a sequence. The entire training objective is self-supervised: no human labels are needed, only large quantities of text.
• Tokens are subword units produced by Byte-Pair Encoding. One token is roughly 4 English characters. "ChatGPT" is 3 tokens. This tokenisation explains why LLMs struggle with character-level tasks and arithmetic.
• Pre-training on next-token prediction forces models to learn grammar, world knowledge, reasoning patterns, and common sense implicitly, because all of these are needed to predict language well.
• Scaling laws (Kaplan et al., 2020; Chinchilla / Hoffmann et al., 2022) show that LLM performance follows power laws with model size, data size, and compute. Optimal training scales parameters and tokens roughly equally.
• RLHF (Ouyang et al., 2022) converts a capable but unruly base model into a helpful, harmless assistant through three stages: SFT on human demonstrations, reward model training on preference comparisons, and RL fine-tuning to maximise reward.
• Hallucination occurs because LLMs generate statistically likely text, not because they "know" facts. They do not consult a database. When patterns from training are absent or contradictory, fluent but false text results. RAG mitigates this by grounding generation in retrieved documents.
• Context windows range from 4K tokens (early GPT-3) to 1 million tokens (Gemini 1.5 Pro). Anything outside the context window is invisible to the model. LLMs have no persistent memory between conversations by default.
• The LLM landscape divides into closed-source frontier models (GPT-4, Claude 3.5, Gemini) and open-weight models (LLaMA 3, Mistral). Open models enable research, auditing, and fine-tuning at no cost.