Skip to content
AEO Canon · the reference for answer-engine optimization

How Are AI Models Trained?

AI models are trained in stages — large-scale pretraining on text to learn language, then fine-tuning and reinforcement learning from human feedback (RLHF) to make them helpful, honest, and safe to use.

BBurke Atkerson3 min read

AI models are trained in stages: massive pretraining on text to learn language, then fine-tuning and reinforcement learning from human feedback (RLHF) to make them helpful and safe. That pipeline is where a model's abilities — and its blind spots, biases, and knowledge cutoff — all come from.

What are the stages of training an LLM?

Training an LLM happens in three broad stages, each doing a different job: pretraining builds raw language ability, fine-tuning teaches the model to follow instructions, and RLHF aligns it with what people actually want.

  1. 1

    Pretraining

    The model is shown trillions of tokens of text and learns to predict the next one. This is where it absorbs grammar, facts, reasoning patterns, and style — at enormous compute cost.

  2. 2

    Supervised fine-tuning (SFT)

    The pretrained model is trained on curated prompt–response examples written or vetted by humans, teaching it to follow instructions rather than just continue text.

  3. 3

    Reinforcement learning from human feedback (RLHF)

    Humans rank multiple model outputs; a reward model learns those preferences, and the LLM is optimized to produce answers people rate as helpful, honest, and harmless.

The raw output of stage one is a base model — fluent but unaligned. Stages two and three are what turn it into the polite, instruction-following assistant you actually talk to.

How much data and compute does training take?

Training a frontier LLM takes trillions of tokens of text and millions of dollars of compute. The most influential guidance on the right balance is DeepMind's 2022 Chinchilla study, "Training Compute-Optimal Large Language Models" (arXiv 2203.15556), which trained over 400 models and found that for a fixed compute budget, model size and training data should scale together — roughly 20 tokens of data per parameter. To prove it, they trained a 70-billion-parameter model (Chinchilla) on about 1.4 trillion tokens and beat a much larger model trained on less data.

Why the data-to-size ratio matters

Chinchilla's lesson — that many models were under-trained on data for their size — reshaped the field toward training on far more text. It's also why the quality, breadth, and recency of training data became a central competitive question, not an afterthought.

What does RLHF actually change?

RLHF changes a model from a raw text-predictor into a usable assistant by tuning it toward human-preferred behavior. A pretrained model will happily continue a prompt in unhelpful, unsafe, or rambling ways; RLHF teaches it to answer the question, decline harmful requests, admit uncertainty, and match the tone people expect. Most of what feels like an assistant's "personality" and judgment is a product of this stage. Labs document their training and alignment choices in model cards and system documentation — the kind of lab documentation worth reading when you evaluate a model.

Why is a trained model frozen in time?

A trained model is frozen because training produces a fixed set of weights — once learning stops, the model's built-in knowledge does not change until it is retrained. Everything it absorbed has a cutoff date, after which it knows nothing unless told. This is the knowledge cutoff, and it is the reason models can be confidently out of date and why retrieval-augmented generation exists: to feed the model current information at answer time rather than waiting for the next training run.

Why does training matter for AEO?

Training matters for AEO because it determines what a model knows on its own versus what it must look up — and only the "look up" path can cite your content. A model's parametric knowledge is frozen and unattributable; your opportunity to be cited lives in the retrieval layer, where current, trustworthy, well-structured sources are pulled in at query time. That's the link from training to what is AEO: you can't get baked into the model, but you can become the source it retrieves and trusts — which is what authority and the rest of The AEO Canon are about.

Next: what is training data for where a model's knowledge comes from, or why AI models hallucinate for what happens when that knowledge runs out.

Frequently asked questions

How is an LLM trained, step by step?
In three broad stages. (1) Pretraining: the model learns language by predicting the next token across trillions of words of text. (2) Supervised fine-tuning: it learns to follow instructions from curated example prompts and responses. (3) Reinforcement learning from human feedback (RLHF): humans rank outputs and the model is tuned toward the preferred ones, making it more helpful and safe.
How much data does it take to train an LLM?
Trillions of tokens. DeepMind's 2022 Chinchilla study found that for compute-optimal training, models should see roughly 20 tokens of data per parameter — so a 70-billion-parameter model was trained on about 1.4 trillion tokens. Frontier models since have used even more data and compute.
What is RLHF?
Reinforcement learning from human feedback (RLHF) is the stage where humans rate or rank a model's responses, and the model is optimized to produce the kinds of answers people prefer. It's the main reason a raw next-token predictor becomes a helpful, well-behaved assistant rather than just an autocomplete engine.
Does training give a model live knowledge of the world?
No. Training is a snapshot. Once a model is trained, its built-in knowledge is frozen at its knowledge cutoff date and does not update until it is retrained or fine-tuned. To answer about anything newer, the model needs retrieval or a search tool.

Last updated .

Part of

Related reading

AI-detection tools are unreliable and not what answer engines use to decide citations — engines judge content on quality, originality, and accuracy, not on whether a machine wrote it. Stop chasing detection and make content genuinely original, because generic content fails no matter who wrote it.

2 min read

AI-generated content can get cited, but only when it's made genuinely original, accurate, and useful — raw model output tends to be generic, unsourced, and interchangeable, which is exactly what engines skip. The deciding factor is the substance and originality you add, not whether a model helped write it.

2 min read

It depends on the engine — web-grounded engines like Perplexity and Google AI can surface new content within days once it's crawled, while a model's built-in training knowledge lags months behind its cutoff. So fresh content reaches retrieval-based answers quickly but base-model knowledge slowly.

2 min read