Large Language Models like ChatGPT, Claude and Gemini are programs trained on massive text datasets—books, articles, code repositories, forums—totaling billions or trillions of words. At their core they perform “next-token prediction,” a process where during training they hide the next word in a sentence (“The cat sat on the ___”), predict it, then adjust internal weights based on the error. Repeating that billions of times builds a system that can string words into coherent paragraphs, essays or even code snippets. Under the hood, every word you type is first broken into tokens—units that might represent whole words or sub-word pieces—each mapped to a unique number. The model converts those numbers into vectors, feeds them through dozens of attention-based neural layers that compare tokens to each other, and outputs a probability distribution over possible next tokens. Those output numbers then get decoded back into text. Because the model never sees raw letters or words directly, it struggles with fine details like letter-counts or arithmetic, handling language only through patterns learned during training. Token limits—called context windows—define how much you can feed an LLM at once. A model with a 200,000-token window holds roughly 150,000 words in its “working memory.” Cutting-edge systems like Anthropic’s Claude Opus 4.6 and Google’s Gemini handle up to 1 million tokens (about 750,000 words), while Meta’s Llama 4 Scout can chew through 10 million tokens. Bigger windows let the model consider more context, but they don’t automatically fix issues like factual errors or logical gaps. Training LLMs involves two phases: base training for general language patterns, and fine-tuning on specialized data to sharpen performance on tasks like summarization or coding. Model quality depends on data diversity, model size (parameter count) and compute resources used during training. Small models can run on laptops but lack sophistication. Large models need data centers but deliver far richer, more coherent outputs—so you pick the right tool based on your task and resource constraints.