The AI side — models, agents, and why an "aggregator"

Plain-English, no background needed. The big picture sketched what software and AI are; this lesson zooms into the AI part — what these models actually are, and why this app exists. (The technical plumbing — servers, databases, the cloud — comes a bit later, in Foundations.)

You don't need to build a model to build this app. We never train AI here — we call models other companies run. This lesson is about understanding what you're ordering, not how to cook it.

1. What is an AI model?

A model is a program that has read an enormous amount of text (or images, or video) and learned the patterns in it. You give it some input, it predicts a sensible continuation.

• Analogy: an incredibly well-read assistant who has skimmed much of the internet. Ask a question and they give you their best answer from memory — fast, fluent, and usually right, but not looking anything up unless you give them a tool to.
• It predicts, it doesn't "know." That's why a model can be confident and wrong (people call this a hallucination). You stay the editor.

Models come in different kinds, grouped by what they take in and put out:

• Text (language) — text in, text out: chat, summarize, translate, write code.
• Image — a prompt → a picture (or edit/transform an existing image).
• Audio — speech → text (transcription), text → speech (voices), or generated music and sound.
• Video — a prompt (or a starting image) → a moving clip.
• 3D — a prompt → a 3D shape or object (for games, AR/VR, product design).

This app focuses on text (chat), image, and video — the three in the product spec. Audio and 3D are part of the wider AI world; they're listed here so you see the whole landscape, not because the app does them (today).

2. Different models, different strengths

There isn't one "best" AI — there are many, each with trade-offs: some are sharper at code, some at creative writing, some make better images, some are cheaper, some are faster, some will write things others refuse to (see §6 below).

• Analogy: a kitchen with many specialist chefs. You wouldn't ask the pastry chef to grill a steak.
• In this project: the whole point is to put many models side by side so you can send the same prompt to several and compare — instead of being locked into one app's single model.

3. What is an AI agent?

A plain chatbot answers and stops. An agent is a model that can take actions in a loop to reach a goal: it can use tools (search the web, read a file, run a command), look at the result, and decide the next step — over and over until it's done.

• Analogy: a chatbot is someone answering trivia. An agent is an assistant you hand a task to — "book this trip" — who then makes the calls, checks options, and reports back.
• In this project: the app is built by an agent — Claude Code (in the desktop app) writes the code, runs the commands, and commits the work, while you direct. So you're not just using AI here; you're watching an agent build software. That's a big part of what this repo teaches.

4. Why an "aggregator"?

Think of an API as a menu one program offers another: a fixed list of things you can order, and how to ask for them. (More on this in Foundations.) An aggregator is one app that orders from many providers' menus and brings the results into a single place.

• Without one: you juggle separate apps and subscriptions — one for chat, another for images, another for video.
• With one: a single interface for all of them, one place for your history, and the ability to compare models on the same task.

5. Bring-your-own-key (BYOK)

To order off a provider's menu, you need an API key — a secret password that says "bill this to my account." BYOK means each user brings their own keys.

• Analogy: the app is the waiter that carries your order to many kitchens — but you hand over your own payment card (your key). The app never puts the meal on its own tab.
• Why it's good: you pay only your own usage (no markup), you're not sharing a pool with strangers, and you get access to a far wider range of models than a typical consumer app exposes.
• Because keys are basically passwords to money, storing them safely is a real topic of its own — that's a later lesson, Keeping your keys safe.

6. AI censorship & "uncensored" models

Many mainstream models have guardrails — they'll refuse certain requests. Other, more open models have fewer restrictions and will attempt things the big ones won't.

• Neither is simply "better" — it's a trade-off between safety and freedom, and the right choice depends on what you're doing (and on using them legally and responsibly).
• In this project: because it's BYOK and aggregates many providers, it can surface this difference honestly — letting you see which model will and won't do a given task, rather than hiding it. Understanding that models differ here is itself part of AI literacy.

See it in the real project

Want the product's full "what & why"? Read ../prd/phase-1.md.

Next: Foundations — the concrete pieces (server, frontend/backend, the cloud, database vs. files) and the proper names for them.