Week 1 — LLMs, Models & Harnesses + Setup

The model landscape and getting your two harnesses working

Published

June 1, 2026

Week 1 — LLMs, Models & Harnesses

The model landscape, what a harness is, and getting Claude Code + VS Code/Copilot working

This is the opening week of the course — the one front-facing, lecture-heavier session. By the end you understand the model landscape (including open-weights models), what a harness is and why it matters, and you have both course harnesses working on your machine.

Before you come to class (30–60 min)

✅ Pre-class checklist

That’s enough. We finish setup together in class.

Learning objectives

By the end of this unit you will be able to:

  • Explain what an LLM is, what a context window is, and why both matter for data work.
  • Distinguish closed/frontier models from open-weights models and say when each is the right choice.
  • Explain what a harness is and why it matters as much as the model.
  • Run an identical small task across Claude Code (terminal + desktop) and VS Code + Copilot, and articulate the differences.

Session shape (200 min · 50·50·40·40·20)

Chunk Focus Mode
1 (50) LLMs & the model landscape Talk
2 (50) Harnesses, open-code, open-weights Talk
3 (40) Hands-on: Claude Code (terminal + desktop) Individual
4 (40) Hands-on: VS Code + Copilot + the comparison Individual
Discussion (20) Surfaces, trust, the jagged frontier Group

Chunk 1 — LLMs & the model landscape (50 min · talk)

📊 Core LLM concepts

Slideshow: LLM Concepts and Applications

  • What an LLM is — a statistical model predicting the next token from massive training data.
  • The transformer revolution — how 2017’s “Attention is All You Need” made today’s models possible.
  • Context windows — what they are and why they shape every data-analysis workflow.
  • Training & feedback — pre-training, resources, and human feedback (RLHF); why models differ.
  • The jagged frontier — where LLMs excel and where they quietly fail. See Mollick on the shape of AI jaggedness.
  • Cyborg vs Centaur — two ways to split work with AI (Mollick): the Centaur hands off whole sub-tasks; the Cyborg interleaves human and AI move-by-move. You’ll feel both this term.

The model landscape (see Which AI model):

  • Frontier closed models — Claude, GPT, Gemini. State of the art, rented per token.
  • Open-weights models — Llama, Qwen, Mistral, DeepSeek, Kimi. Download and run yourself; roughly SOTA-minus-a-few-months. Why they matter: privacy, cost at scale, reproducibility. Full orientation in Open-weights & local models.

Chunk 2 — Harnesses, open-code & open-weights (50 min · talk)

🔧 The model is the engine; the harness is the car

A harness wraps a model and lets it actually do things — read your files, run code, edit files, iterate. The harness matters as much as the model.

This course uses exactly two harnesses:

Harness Surfaces Notes
Claude Code Terminal and desktop app Terminal-native agent that sees your project, runs code, iterates.
VS Code + GitHub Copilot IDE In-editor assistance; you can point it at different backing models.

That’s it — no other tools to learn. Everything transfers conceptually to alternatives.

Two awareness topics (not required to install):

  • Open-code tools — open-source agentic harnesses you run yourself and point at any model.
  • Open-weights / local models — running a private, frozen model locally (e.g. via Ollama), and pointing a harness at it.

IDE assistance vs terminal-native execution — the jump this course makes is not “AI vs no AI.” It’s about where the AI runs and how tight the loop is. Copilot keeps you in the editor; a CLI agent compresses the ask→run→inspect→fix loop into one place.

Chunk 3 — Hands-on: Claude Code (40 min · individual)

🖥️ Same task, two surfaces — terminal then desktop

  1. Install & authenticate Claude Code in the terminal (finish Installing AI CLI Tools if needed). Open a terminal, make a folder, and launch:
mkdir my-first-agent && cd my-first-agent
claude
  1. A first real task. Try a tiny data-search prompt and verify the result against reality:
Get me an income dataset by planning region (county) in Connecticut for 2023.
Present the results as a table I can copy and edit.

Then find the actual data and compare. Try it more than once; tweak the prompt to be more specific.

  1. Now do the same thing in the Claude desktop app. Notice what changes between terminal and desktop: file access, permissions, how you see the work.

Chunk 4 — Hands-on: VS Code + Copilot + the comparison (40 min · individual)

🧩 The IDE harness

  1. Set up VS Code + Copilot (guide). Open Copilot Chat (Ctrl+Shift+I / Cmd+Shift+I) and run the same Connecticut data task. Then switch the backing model in Copilot (e.g. a GPT model vs a Claude model) and rerun — notice whether the answer or the workflow changes.

🎯 The task: one small analysis, three surfaces

Don’t just look up data — make something. Pick one small, checkable analysis and run it on all three surfaces (Claude Code terminal, Claude desktop, VS Code + Copilot):

“Get Connecticut median household income by county for 2023, load it into a dataframe, and make one labelled bar chart sorted high to low. Save the chart as a PNG.”

For each surface, note:

  • Did it actually run the code and produce the PNG, or just hand you a snippet?
  • Was the data right? Open the numbers and compare to the Census/ACS source. Flag anything invented.
  • How many turns did it take to get a correct chart?

Fill in a tiny comparison table — this is the raw material for your delivery note:

Surface Ran the code? Data correct? Turns to a good chart Felt best for…
Claude Code (terminal)
Claude desktop
VS Code + Copilot

🦙 Stretch (optional) — run a model locally

Read Open-weights & local models. If you’re curious, install Ollama, pull a small model (e.g. ollama run llama3.2), and run a tiny task locally. Reflect on the privacy/cost/capability trade-off versus the frontier models you used above.

Discussion (20 min)

  • Terminal vs desktop vs IDE — which felt right for which kind of task?
  • Closed vs open-weights — when would privacy or cost push you to a local model?
  • The jagged frontier — what did the AI nail, and where did it quietly get the Connecticut data wrong? How did you catch it?
  • Switching the backing model in Copilot — did it change the answer, the style, or nothing you could see?

Delivery

📦 What to hand in (Sunday 23:55)

  • Working environment — evidence that Claude Code (terminal + desktop) and VS Code + Copilot all run (a screenshot of each completing the Connecticut task is fine).
  • A short note (½ page): which setup you’ll use for the rest of the course and why; one thing the AI got wrong on the Connecticut task and how you caught it.

Academic integrity & AI use

  • AI as assistant — use AI to enhance your capabilities, not replace your thinking.
  • Maintain authority — you remain responsible for all outputs and interpretations.
  • Verify everything — always validate AI suggestions, especially statistical claims.
  • Document usage — keep track of how AI helped, to learn and for transparency.

Red lines: never submit unverified AI output as your own; always understand the analysis you present; don’t outsource critical thinking.

Knowledge Base

Further reading (optional)