Session 05¶

Topic 1: Interfaces¶

• terminal
• SSH
• Zsh
• editors
• Git
• AI 👈

Today¶

AI¶

1. AI coding assistants
2. AI agents
3. AI hands-on

AI coding assistants¶

Why use AI for coding?¶

• real-world codebases are large, complex, and interdisciplinary
• data scientists spend more time:
  → debugging / refactoring
  → gluing libraries together / reading unfamiliar code
• AI changes how we interact with code – not just how fast we type

→ Most of your time is not spent inventing algorithms. It’s spent making code work.

LLMs for coding (ChatGPT, Claude)¶

Large Language Models (LLMs) are trained on:

• programming languages
• documentation
• common coding patterns

you chat with them:

• ask questions
• generate code
• explain errors

Example prompts¶

• “Explain this pandas function”
• “Convert this NumPy code to PyTorch”
• “Why does my training loop blow up?”

What AI is great at¶

💬

• understanding unfamiliar code
• generating predictable code (boilerplate)
• translating ideas to code
• debugging error messages

When AI fails¶

💬

• can be wrong without a clue about it 🙃
• invents functions or APIs that don’t exist 👻
• cannot distinguish what’s correct and what’s not 🥴
• follows patterns 🐜 – not logic 🧠
• lacks hidden assumptions 🫣
• gets overwhelmed by long project structure 😬

→ AI is a powerful junior collaborator – not an oracle – it never replaces expertise.

AI agents¶

AI Coding Agents¶

AI agents don’t just answer questions, they:

• read your codebase
• modify your files
• run commands (in the terminal)
• iterate until a task is done

→ AI agent = autonomous junior developer

Example tools:

• Claude Code
• GitHub Copilot

• OpenCode
• Mistral Vibe

What can AI coding agents do?¶

• implement features across files
• refactor entire modules
• add tests
• fix failing tests
• migrate APIs or libraries

Example prompts:

• “Refactor this notebook code into a Python package”
• “Add logging and config files”
• “Turn this prototype into production-ready code”

Why this matters for data scientists¶

New skills¶

• prompting with constraints
• reviewing AI-generated code (👋 git diff)
• designing experiments and tests
• knowing what to ask and what to distrust

Best practices (and survival guide)¶

Do: 🚦

• treat AI output as a draft
• write tests first (or ask the agent to)
• ask why, not just what
• use Git and keep commits small & reviewable

Don’t: ⚠️

• blindly trust generated code
• let agents run unchecked on critical systems
• confuse productivity with correctness

AI amplifies thinking.

AI hands-on¶

AI four ways¶

1. in the browser
2. in the browser with split screen
3. in the terminal
4. in the editor

AI - 1^st way¶

the traditional way is to open a chat in the browser

open https://chatgpt.com

Demo - Example prompt¶

“Implement a to-do list manager in Python with a command-line interface.”

AI - 2^nd way¶

still in the browser: split screen with chat and code

open https://claude.ai

Demo¶

• let’s repeat the same prompt as before
• then request some change to the generated code

AI - 3^rd way¶

in the terminal: now files in current working directory are used as context

opencode
# we'll look into how to install this tool in the bonus material

Demo¶

• in the terminal, go to the root folder of a codebase
• open AI agent CLI; then ask to explain the codebase
• instruct to modify a particular code in a file

AI - 4^th way¶

in the editor: also uses files as context

Demo¶

• in VS Code, open the root folder of a codebase
• ask in the chat window of GitHub Copilot (e.g. to explain the codebase, or modify a file)

Bonus¶

Install OpenCode on Matrix¶

The steps are described in the wiki → OpenCode.

Key takeaways¶

AI as a force multiplier, not a shortcut

your human judgment still defines:

• goals
• correctness
• ethics

the best data scientists will be:

• strong conceptual thinkers
• excellent at reviewing
• effective collaborators with humans and machines

Looking ahead¶

New topic: The Go language¶

Go’s philosophy:

Don’t communicate by sharing memory; share memory by communicating.

Prepare for the next session 🌱¶

sessions/06-todo