The lifecycle has seven stages, and the honest thing to say about them is that you will only ever touch the last five.
Data selection
Choosing what the model learns from. At foundation model scale this means trillions of tokens, and the decisions are about filtering, deduplication, quality and licensing rather than collection.
It is the stage with the longest shadow. Bias in the corpus becomes bias in the model, gaps become blind spots, and copyrighted material in training data becomes a legal question that surfaces years later. When you pick a model you inherit all of it, which is why provenance is a reasonable thing to ask a vendor about.
Model selection
Choosing an architecture and a size. For almost everybody this is not designing a model, it is choosing among published ones on capability, cost, licence, context length, modality and where it can be run.
For most teams this is where the lifecycle actually starts.
Pre training
The expensive one. The model learns language, and a great deal about the world, by predicting the next token across an enormous corpus. It costs millions of dollars and weeks of compute at scale, and it produces a model that is knowledgeable and not yet useful, since predicting text is not the same as following an instruction.
Almost nobody does this. Continued pre training is the affordable relative, where an existing model is trained further on a large domain corpus so it absorbs the vocabulary of medicine or law.
Fine tuning
Adapting a pre trained model to behave a particular way, on a much smaller labelled dataset. Instruction tuning teaches it to follow instructions rather than continue text. Reinforcement learning from human feedback uses human rankings of output to train a reward model that shapes responses towards what people prefer.
Fine tuning changes behaviour, tone, format and domain fluency. It is a poor mechanism for facts, and the exam expects you to know the difference.
Evaluation
Benchmarks, human judgement and task specific measures. The awkward part is that a foundation model is general, so no single score describes it, and benchmark performance often fails to predict how it does on your actual work. Evaluating against your own task is the only thing that answers the question you have.
Deployment
Serving the model. Hosted behind a managed API, or self hosted where you control latency, cost at scale and data residency. On AWS the exam names Amazon Bedrock for managed access to foundation models, SageMaker JumpStart for deploying pre trained models into your own account, Amazon Q for an assistant built on top, and PartyRock as a playground for trying prompts without writing code.
Feedback
What comes back from real use and feeds the next round. Explicit ratings, logged failures, human review of sampled output, and the drift in what people are asking for.
This is the stage teams most often leave out, and leaving it out is what turns a launch into a plateau. Without a route from a bad answer back into the system, nothing improves except by changing model.