From AI Models to Real Applications
See the difference between an AI model and an AI application, and the layers an app wraps around a model.
Four tracks in, you know what a model is: a huge file of numbers that turns text into a next-token guess, running on GPUs in a data centre. But nobody uses a model. People use apps: a chatbot, a coding assistant, a support bot on a shop’s website, a button in a spreadsheet. This track is about the gap between the two.
On its own, a model knows nothing about you, your data or what day it is. It doesn’t know it’s meant to be a cooking assistant, can’t look anything up, and replies in free text that a program can’t easily use. Everything that makes an AI product useful is added by the app around it:
- Instructions that tell the model its job, tone and rules.
- Context: your data, the conversation so far, facts the model can’t know.
- Tools it may use, such as a search or a database lookup.
- Output handling: asking for a usable format, checking the answer, then showing it nicely.
The striking thing is how much of an AI product’s quality lives in those layers. Two apps built on the very same model can feel completely different, one vague and generic, the other sharp and personal, purely because of what each puts around the model.
The rest of this track walks through those layers in order: how an app talks to a model at all (APIs), which model to pick, how to write instructions and prompts, how to get structured data back and stream answers, and how to feed in the right information without paying for the wrong.
Key takeaways
- A model turns text into text; an app turns that into something useful.
- Apps add instructions, context (your data, memory), tools and output handling around the model.
- Two apps on the same model can behave completely differently.
- Much of an AI product’s quality comes from these layers, not from the model alone.
Check your understanding
Go deeper (more technical)
Engineers sometimes call these layers the harness or scaffolding around a model. In larger systems they include retrieval of documents (Track 6), tool use and agent loops (Track 7), evaluation, logging, caching and fallbacks to another model when one fails.
A useful habit when an AI feature misbehaves: before blaming the model, look at exactly what text the app sent it. Most problems turn out to be missing context or unclear instructions.