Understanding GPT‑4, ChatGPT, and the Foundations of Large Language Models

Artificial intelligence (AI) is often described as the next major revolution after the industrial and information technology revolutions, and the rapid adoption of OpenAI’s products like ChatGPT illustrates how AI is permeating every industry.

Many users enjoy ChatGPT’s capabilities without understanding the underlying technology; the article emphasizes that with the OpenAI API and a little Python knowledge, anyone can start building intelligent applications.

Foundations of NLP and AI – AI is defined as computer systems that perform tasks requiring human intelligence. Machine learning (ML) is a subset of AI that enables systems to learn from examples, while deep learning (DL) is a further subset that uses artificial neural networks inspired by the brain.

Large language models (LLMs) such as GPT‑4 and ChatGPT are built on the Transformer architecture, which processes input tokens in parallel and uses attention mechanisms to capture contextual relationships.

Figure 1‑1: Hierarchical relationship from AI to Transformer

The article lists common NLP tasks that LLMs can perform: text classification (e.g., sentiment analysis), automatic translation (including code translation), question answering, and text generation.

Attention mechanisms – Cross‑attention helps the model focus on relevant parts of the input when generating each output token, while self‑attention evaluates the importance of each token relative to all others, enabling the model to build new concepts from context.

Examples illustrate how cross‑attention selects the English words “sunny” and “weather” to generate the French word “ensoleillé,” and how self‑attention highlights “Alice” and “colleagues” to understand the pronoun “her.”

Figure 1‑2: Cross‑attention focusing on key input tokens

Figure 1‑3: Self‑attention creating the new concept “Alice's colleagues”

Because Transformers can process tokens in parallel, they map well onto GPUs, which excel at parallel computation, allowing large models to be trained on massive datasets.

The original Transformer paper ("Attention Is All You Need", 2017) introduced encoder‑decoder structures; GPT models use only the decoder, relying solely on self‑attention for generation.

Figure 1‑4: Evolution from n‑gram models to modern LLMs

Tokenization and prediction – When a prompt is received, the model first splits it into tokens (words, sub‑words, spaces, punctuation). Using the Transformer’s attention, the model assigns probability scores to possible next tokens and selects the highest‑probability token, repeating the process iteratively until a complete sentence is formed.

Figure 1‑5: Iterative token‑by‑token text completion

Figure 1‑6: Steps to obtain an InstructGPT model (redrawn from Ouyang et al.)

GPT‑4, released in March 2023, is a multimodal model that accepts both text and images, achieving higher scores than previous models on professional exams such as the US bar exam and the International Biology Olympiad.