Generative AI has become the most visible face of artificial intelligence in the last few years — tools that write essays, generate images, compose code, and hold open-ended conversations. At the center of the text-based side of this revolution are large language models (LLMs). This article explains what they are and how they actually work, in plain terms.
What "Generative" Means
Most earlier AI systems were built for classification or prediction: is this email spam or not, what grade does this essay deserve, what object appears in this photo. Generative AI systems are built to produce new content — new text, new images, new audio — rather than simply labeling or scoring existing content.
A large language model is a generative AI system specifically trained on text. Given some input (a question, an instruction, the beginning of a sentence), it generates new text as output, one small piece at a time.
The Core Mechanic: Predicting the Next Token
At the most fundamental level, a language model is trained to do one thing: given a sequence of text, predict what token (roughly, a word or word-fragment) is most likely to come next. That sounds almost too simple to explain the fluency of modern AI writing — but the key is scale.
These models are trained on enormous quantities of text — a meaningful fraction of publicly available human writing — using the neural network architectures described in earlier articles in this series (specifically, the Transformer architecture and its attention mechanism). Through this training, the model doesn't just learn surface-level word patterns; it develops rich internal representations of grammar, factual associations, reasoning patterns, and stylistic conventions, all in service of getting better and better at the deceptively simple task of predicting the next piece of text.
Once trained, generating a full response is just this next-token prediction repeated over and over: predict the most likely next token, add it to the growing text, then predict the next one after that, and so on, until the response is complete.
Why Scale Changed Everything
Language models existed in simpler forms for years before the current wave of generative AI. What changed was scale along three dimensions simultaneously: the amount of training text, the size of the model (measured in the number of adjustable parameters, now often in the hundreds of billions), and the amount of computing power used for training.
Researchers observed something notable as these dimensions scaled up: models didn't just get marginally better — they developed qualitatively new capabilities that smaller models simply didn't have, from following complex multi-step instructions to a kind of emergent reasoning ability. This is part of why progress in the field has moved so quickly in a short span of time.
Fine-Tuning and Alignment
A model trained purely to predict the next word from raw internet text produces something interesting but not necessarily helpful or safe — it will happily continue text in whatever direction its training data suggests, without any particular concern for being accurate, appropriate, or genuinely useful to the person asking.
This is why modern AI assistants go through additional training stages after the initial, massive pretraining phase. Techniques like instruction fine-tuning (training the model specifically on examples of helpful, well-formed responses to instructions) and reinforcement learning from human feedback (where human reviewers rate model outputs, and the model is further trained to produce outputs more like the ones people rated highly) shape a raw, next-token-predicting model into something that behaves like a genuinely helpful assistant — one that tries to be accurate, follows instructions, and avoids harmful outputs.
Capabilities and Genuine Limitations
Large language models can write coherently across an enormous range of topics and styles, hold multi-turn conversations, explain complex concepts at different levels of depth, translate between languages, write and debug code, and assist with tasks that would previously have required a human expert's time.
They also have real, well-documented limitations. They can generate confident-sounding text that is factually incorrect — a phenomenon often called "hallucination," because the model is, at a mechanical level, generating plausible-sounding continuations rather than looking up verified facts. Their knowledge is bounded by what was in their training data, and their training data has whatever biases and gaps existed in the text they learned from. They do not have persistent memory of conversations beyond what is explicitly provided to them in context, unless a system is specifically built to add that.
Why This Matters for Education
Understanding language models at this level of detail changes how you use them. A model that predicts plausible next tokens, rather than retrieving verified facts, is enormously useful as a writing partner, an explainer of concepts, a tutor that can meet a student where they are — and it is genuinely risky to treat as an unquestionable source of truth, especially for younger learners still developing the judgment to separate a confident tone from actual accuracy.
The most effective use of generative AI in education leans into what these systems are genuinely good at — explanation, practice, feedback, personalization — while pairing that capability with the kind of human oversight and critical-thinking instruction that keeps students in the driver's seat of their own learning, rather than passive consumers of whatever the model produces.