Problem Set 0

Getting set up, and three conversations

Everything in this problem set happens here, on this site. Nothing to download, nothing to screenshot.

intro2ml.com : the course website that has the schedule, weekly notes, problem sets, policies, office hours. Start here when you want to know what is happening.
learn.sematlas.com : this website has the interactive slides for many lectures. Its main purpose is to make it easier for you to give feedback, ask questions, and interact with LLMs in the context of this course. Forgive any bugs that arise, and leave a comment (the button at the top right) if you find one.

Space or →	Next slide. Some slides reveal a line at a time, so keep going.
←	Back.
Esc	Zoom out to see the whole deck at once. Click any slide to jump.
F	Fullscreen.
A	Ask a question about the slide you are on.
C	Leave me a comment on this slide.
R	Readings and references for this lecture.

The same buttons are in the toolbar at the top right, if you would rather click than type.

Press A anywhere in a deck.

The question goes to a model that has been given the slide you are currently on and the slides around it, so you can ask "where did this term come from?" and it knows what "this" is.
Your questions are saved. Open My Questions to reread everything you've asked, with a link back to the slide that prompted it.
You have a budget of 100 questions a week per person. It is generous on purpose: it exists so a runaway loop cannot empty the course account, not to ration you.
Read critically. LLMs are currently very good study partners (if you use them properly) but also very confident liars. So if for nothing else, treat these conversations are a good exercise in critical thinking. If you find any problem, press C (or the comment box on the top right) to tell me.

The following problem set is three conversations; each on its own slide.

Each slide has a minimum number of your turns before it counts as done. That is a floor, not a target so you can keep the conversation going if you're finding it useful.
I read the conversations. Not to grade the model's answers, but to see how you push on an idea. A conversation where you argue back, where you build on the LLM's insight, you probe for a deeper understanding (etc), is worth more than one where you ask generic questions. Basically, show me you're thinking. Please don't use another LLM to generate the conversation -- that would be sad and I can usually tell.
The conversations save as you go. So you can close the tab, and come back to it later.

Revealjs website hosted on Render with Anthropic/OpenAI API keys

This website is an experiment in integrating LLMs into slides and enhancing them with interactive elements that wouldn't be possible on traditional slides.
Your interactions go to OpenAI or Anthropic, who run the model, and are stored on this website's server, which I run, where I can read and review them.
The conversations are private so I'm the only one who can read them. They are kept for the semester and the grading period, and I delete them after that unless you ask me to keep them.
Since this is not the university's official learning system (i.e. Moodle), you can opt out of the exercises that use this website. Just email me and we can agree on another way to do the interactive problems/conversations.

Pick a field you're interesting in. Music, robotics, aerospace, cybersecurity, basketball, Formula 1, medicine, archaeology.

The purpose of this exercise is to think about the applications of ML to a context relevant to you.
I'd like you to write without worrying much about editing sentences. I'm looking curiosity, critical thinking, creativity. In this context: what data, collected by whom, for what models. You can dump ideas on the LLM and react to how it interprets them. Go beyond generic prompts like "Use ML to improve healthcare"; think how, where and for whom.
The conversation starts on the next slide. You have to go through a minimum of six turns. But feel free to go beyond that.

Take the strongest idea from the first conversation and turn it into something a person could actually start on Monday.

An abstract that could describe fifty projects describes none. Name the dataset, the input and output, and how you would know whether it worked.
Check what already exists. Finding that your idea has been done is not failure, it is the literature telling you where the edge is.
This is not your final project. It is practice at the move you will make all semester: vague interest, sharp question.

Build a syllabus for "Machine Learning for X", where X is your field. Then compare it to the one you are actually taking.

Designing a course forces you to put ideas in order, and ordering is where you find out what you do not understand.
Ask what has to come first, and why. Any topic you cannot justify the position of is one to be suspicious about.
Where your syllabus and mine disagree, tell me. Sometimes you will be right.

The rest of Problem Set 0, the calculus and linear algebra review, is on intro2ml.com and goes to a Moodle submission link, as usual.
Stuck, or something is broken? Press C on the slide where it happened. That reaches me with the slide attached.