---
title: "How the shapes differ"
lesson_id: "04"
---

# How the shapes differ

- **Lesson ID:** 04
- **Goal:** Compare an image classifier and a chat language model by input, output, one step, mistakes, and shipped app work.
- **Human lesson:** [04-how-they-differ.html](04-how-they-differ.html)

## Prerequisites

- Lessons 01 to 03, or familiarity with image and language model jobs.
- A basic app feature in mind.

## Inputs, outputs, and artifacts

- **Inputs:** ClothingCNN as the image example and a chat language model as the text example.
- **Outputs:** A plain side-by-side comparison and a small multimodal note.
- **Artifacts:** Two comparison `.flow` diagrams and the practical comparison list.

## Agent build steps

1. Set up ClothingCNN and a chat model as the two running examples.
2. Compare pixel-grid input with token-sequence input.
3. Compare one image classification call with repeated next-token work.
4. Compare label mistakes with fluent but incorrect text.
5. Explain what an app ships around each model.
6. Add a short note that multimodal systems can accept image and text together.
7. Link [the forward course](../pytorch-core-ai-xcode/) and [From the metal to the model](../core-ai-from-metal-back/), including the need to profile actual silicon.

## Constraints

Keep the comparison concrete and easy to read. No transformer math, papers, CUDA, quizzes, or em dashes. Keep both prior-course links relative. Use HTML `.flow` diagrams, not Mermaid.

## Key concepts

- ClothingCNN: pixel grid in, one look, one clothing label out.
- A chat model: token sequence in, next token out, repeated until the reply is done.
- Compare input, output, step count, visible mistakes, and what the app ships.
- Classifier mistakes look like a wrong label. Chat mistakes can look like smooth wrong facts.

## Takeaways

- You can compare a classifier and a chat model without using brand names.
- You can list what a clothing app ships versus what a chat app ships.
- You can ask what enters, what one step does, what leaves, and where mistakes show.

## Acceptance checks

- The comparison covers input, output, one step, mistakes, and what ships.
- ClothingCNN and a chat model remain the running examples.
- The optional multimodal point stays short and jargon-light.
- Both prior courses are linked and the Agent brief link is present.
- [ ] Proceed to [lesson 05 brief](05-picking-a-shape.llms.md).