Getting everything running

TL;DR#

We covered what this course is for, why an agent is good at writing code, and what a prompt actually costs. The setup block took the whole of the last hour, so most people leave with a repository and a hello.py rather than a notebook. Every step of that setup is written down at /guides/getting-started/, and working through it at your own pace before next Wednesday is the whole of what you owe.

What we covered#

A chatbot answers, an agent finishes. Both halves of that definition came out of the room. The part worth keeping is the loop: a model calls a tool, observes what came back, and reasons again until the goal is met. Observation is the step that makes it an agent rather than a text box.

Agents are good at code because code can be verified. You run it and find out, and when it fails the traceback says where. There is also a great deal of well organised code on GitHub for a model to have learned from. The same test explains why mathematics is moving fast right now, and why work whose results are hard to check is harder to hand to an agent.

A prompt is much bigger than what you typed. Four words went into a live agent and over a thousand tokens went out. The rest is the system prompt that defines the tool, plus everything already said in that session. Two consequences follow. The Babson API budget is spent faster by re-running a failing cell twenty times than by asking twenty good questions. And a long conversation carries its whole history with every new message, so starting a fresh one when you change topic is free and worth doing.

What the course grades. You may use an agent. You are accountable for what you hand in, at the level of a block or a function rather than a line: what does this do, and how do the pieces fit together. Every assignment says which mode applies. There are no exams.

Git, in three words. A repository is a folder for one project. A commit records a set of changes with a message describing them. A push sends those commits to GitHub, which is the copy other people can see. Your GitHub profile is your second resume, which is why the work goes there.

uv installs and manages Python for you. You do not download Python from a website. uv keeps several versions side by side and picks the right one, which is why uv run python rather than python is the habit worth building.

marimo instead of Jupyter. Somebody asked why, since Jupyter is still what most workplaces use. In a Jupyter notebook you can change a value in one cell and leave stale results sitting in another, with nothing to tell you. marimo recomputes everything that depended on what you changed, so what is on screen always matches the code. You will get both formats for anything you need to hand to someone else.

The setup steps#

Everything the setup block covered is now a numbered page you can walk alone: /guides/getting-started/.

These stopped people today, and none of them means anything is wrong with your machine.

• uv is not recognised right after you install it. A terminal reads the list of available programs once, when it opens, so a terminal that was already open has an old list. Close it and open a new one. This was the single most common problem in the room.
• A terminal asking for your password shows nothing at all while you type it. No dots, no moving cursor, nothing. Your keyboard is not broken and the machine is not frozen. Type the password and press Enter. This is the one that stops people, because every other password box in your life shows dots. The installer should not need a password at all, because it installs into your own home folder. If it asked, sudo was in the command. Take it out, paste the line exactly as the guide has it, and it will not ask again.
• You edited a file and GitHub Desktop shows no change. The file was not saved. Press Ctrl+S, or Cmd+S on a Mac. A dot next to the filename means unsaved.
• You have no README. The checkbox for it is on the create-repository screen and is easy to miss. You do not need to start over. Make a new file in VS Code and name it README.md.
• Nothing works in the terminal and the errors look strange. Check that VS Code is opened on your repository folder, not on some other folder. The quickest way is GitHub Desktop, then Repository, then Open in Visual Studio Code.
• Repository names take no spaces. Use lowercase, and use the course code.

If something still fails after all that, send the error itself rather than a description of it. Select the text, copy it, paste it into an email. A screenshot is fine too.

What we did not get to#

The Northwind example and the notebook were both named and then set aside. We pick up both next session, and nothing in them is assumed knowledge.

Public links#

• Getting Started, the numbered version of the last hour
• What employers actually asked for, the job-posting analysis from this morning, which gets updated through the term
• The MIT report on AI in teaching, learning and research
• uv installation
• The tokenizer playground, where you can paste a sentence and see it broken into the tokens you are billed for
• Two things on how the models themselves work. Neither is required for this course: Neural Networks: Zero to Hero, a video series, and Build a Large Language Model (From Scratch), a book

Before next session#

1. Work through /guides/getting-started/ to the end. Stop wherever it breaks and bring the error with you.
2. Your repository needs two things: a README.md with something of yours in it, and a hello.py that runs. Both committed and pushed.
3. Finish the pre-course survey if you have not, and register your GitHub account with your Babson address.
4. Watch for your Babson AI key. Updated 9/14: there is nothing to apply for. Babson's AI team emails you the key, and Your Babson AI Key has the setup.

Checkpoint 1 closes Sunday 9/20 and covers the first two sessions. Each of its items is simply present or absent. Today gets you two of them: a repository that exists, and more than one commit in it. The notebook and the agent setup come from next Wednesday, so there is nothing to chase on those yet.