Large Language Models

These are all large language models
Predict the new token (or word) based on preceeding words
They work by being trained on the entire internet
Since there is lots of code written by software developers on the internet they are good at generating code
And since there are lots of lessons on how to learn to code on the internet they can also be good at generating text that explains code
There are a variety of meaningful ethical concerns about using LLMs
Environmental, Intellectual Property, and Labor are the big 3
Writing code with them when lacking the background to evaluate that code can also be dangerous
The worst code in science is code that is wrong, but runs, and models are good at producing this kind of code
But they can be useful
Starting in late 2025 frontier models became very good
My lab uses these tools to support our software engineering

Growing evidence that LLM usage can have significant negative effects on learning
We learn through a combination of repetition and struggling with material
If we take that away we don’t learn
So, both in class and more broadly if we need to think about when this use is helpful vs harmful on a personal level
LLMs can be useful for learning and you are allowed to use them for this purpose
Using them to directly answer the exercises won’t help you learn, because humans need practice to learn
That’s why we have exercises
Using LLMs to solve exercises has been compared to taking a forklift to the gym to lift weights
The goal of lifting weights is to get stronger, so having the forklift move the weights for you doesn’t help
In this class practicing the concepts and the specific implementation goes hand in hand
To put that somewhat more enjoyably here’s quick short from Hero of Coding with Strangers

There is also a growing body of scholarship backing this idea
I’m here to teach and you’re (hopefully) all here to learn, so I ask that you listen to Hero and not use LLMs to directly answer the exercises
So what are useful ways to use them?
You can prompt them to explain things you don’t understand
If we’re in class I definitely recommend asking me or the TA, but if it’s midnight and you’re stuck then they are certainly useful
You can also use them to help debug your code, which we’ll talk about more once we start coding
But trying to fix bugs yourself is also important to learning, so start by working the problem yourself and only ask a model for help at the point where you would ask a one of us for help during class

Due to security and privacy risks and Florida and federal law the only approved way to use LLMs at UF is through NaviGator
It also has the nice benefit of being free
The chat-based interface is available at:
Go to https://chat.ai.it.ufl.edu
Sign in
Through this system you have access to a bunch of different models
Hover over them to see what they are approved for
For this class any of these models is fine
For research, it depending on your work you might be restricted to a subset of them
I tend to use Claude, but any of the newer Claude, Gemini, or OpenAI/GPT models will be sufficient for anything we’re doing

Copy-paste Repeating Things 2 Challenge
Result is likely in Python
Is this what we want?
Ask to rewrite in R (using loops and conditionals)
Run the resulting code
What do you think went wrong here?
Answer may be wrong due to misunderstanding data structure
Download files
Open in text editor
Does this match our idea of tidy data?
Some federal data providers include metadata at the top of the csv
There are ways to handle this, but only if you know what the data looks like, which the LLM doesn’t
We could tell model to look at data by using web search
But we’d run into something called a context limit because the data file is too large
So show it the top of the file
Copy first 40 lines and have model update code
Rerun
Result may still be wrong and model may assume space delimited due to copy-paste

Chat-only interfaces can be limited
Real AI coding assistant like Claude Code or Codex running locally will easily solve this problem
We’ll talk about them later in the semester
Without enough context the LLM may do things we don’t want
Use Python or packages in R we aren’t familiar with
Just give us the answer when we want help learning

Last year the models still weren’t that great
So we have a shorter version of this discussion at the start of class
And then didn’t explore AI until briefly at the end of the semester
But since then there has been a step change in these models for writing code
Do you want to spend a short amount of time each week talking about how to effectively use AI for supporting your coding?
Respond contingent on answer, notes below focused on if the answer is Yes
So, this year we’re going try to engage with this a little bit each week
Doing so will be optional
I’m still figuring out what this looks like
In fact everyone who teaches coding is figuring what this looks like
Generally we’re going to stick to focusing on concepts first
Understanding the basic idea of how computational approaches work remains central to coding and working as a scientist
We’re going to learn how to implement things ourselves in R
This helps us learn how to think computationally
And LLM output for science is typically code and you’ll be responsible for the output of that code
Which means you need to understand it
What I’m going to try to add (briefly) at the end of each week is some introduction to how to engage with LLMs for coding effectively
This is going to very much be an experiment
So feedback on what is working or not working for you will be very helpful

Notes