7.2.1 LLM Overview Roadmap: Capability, Cost, Product Fit

LLM overview is not a model-name list. It helps you decide what a large model can do, what it costs, and when prompting, RAG, Agent, or fine-tuning is a better route.

Look at the Capability Stack First

Route	Use when
prompt	the model already knows enough and task is simple
RAG	private or changing knowledge must be cited
Agent	the model must use tools or take steps
fine-tuning	behavior/style/format needs repeated adaptation

Run One Route Decision

request = {
    "needs_private_docs": True,
    "needs_tool_action": False,
    "needs_repeated_style": False,
}

if request["needs_tool_action"]:
    route = "Agent"
elif request["needs_private_docs"]:
    route = "RAG"
elif request["needs_repeated_style"]:
    route = "fine-tuning"
else:
    route = "prompt"

print("recommended_route:", route)

Expected output:

recommended_route: RAG

This is not a full architecture decision. It is the habit: choose the smallest route that solves the actual product need.

Learn in This Order

Order	Read	What to keep
1	7.2.2 Development History	why scaling and instruction tuning mattered
2	7.2.3 Core Concepts	context, tokens, temperature, latency, cost
3	7.2.4 Industry Landscape	model/provider selection notes
4	7.2.5 LLM Call Workbench	one request/response record

Evidence to Keep

Keep this page’s proof of learning as a small evidence card:

Capability Stack

tokens, context, pretraining, instruction, alignment

Cost Check

context length and output length affect cost/latency

Product Fit

choose model behavior by task need, not hype

Evaluation Loop

fixed cases, score, failure note

Next Action

connect overview to prompt testing in 7.5

Pass Check

You pass this roadmap when you can explain one model choice in terms of capability, context, cost, latency, data privacy, and route fit.

Check reasoning and explanation

1. A passing answer explains how tokens, context, attention, prompts, and generation behavior connect in one request-response path.
2. The evidence should include at least one reproducible prompt or structured-output test, plus notes on why the output passed or failed.
3. A good self-check separates prompt design, RAG, fine-tuning, and alignment: use the lightest method that fixes the observed problem.