Module 1 · Decide and prepare · Lesson 1
Prompt, optimise or train?
When changing a model's weights beats prompts and optimisers and when it doesn't, what fine-tuning one agent step really costs, where you can still train in late 2026, and whose outputs provider terms let you train on.
By the end of this lesson you can
- Place prompting, prompt optimisation, routing, supervised fine-tuning, preference optimisation and reinforcement learning on one ladder, and cite the evidence for climbing each rung
- Separate failures that fine-tuning fixes (format, tool discipline, narrow decisions, cost per call) from ones it can't (missing knowledge, a broken eval)
- Build a cost model that keeps the cost of an experiment apart from the cost of owning a fine-tuned model, and compute its break-even volume
- Choose where to train in late 2026, and decide which models' outputs you may use as training data under provider terms and licences