Ideas on how to use large language models

Never copy data into an LLM

Ethical considerations: are the data yours to share?
Practical considerations: uses up unnecessary LLM memory and context (taking away from a better answer)
Learning considerations: you’ll learn better if you think more critically about what your data are and what questions you really need an LLM’s help with

On that note…

Use hypotheticals

For example, your prompt could be

I have a `data.frame` with columns `x` and `y`,
how do I….

Most LLMs will easily work with that prompt to first generate a toy data object like

my_data <- data.frame(x = rnorm(10), y = rnorm(10))

And then address your question

Use hypotheticals

This is useful because

it makes you think about your data and what you need to figure out
it keeps the real data private
having the LLM make up toy data gives you a chance to evaluate the LLM to see if it’s really understanding your context and question

Break your problem into smaller parts

Rather than prompting

I want to model how `y` changes with `x` based on group `g`, how?

Try breaking it up with your human brain first and figure out what steps you actually want to ask about

Prep data for modeling (suppose you already know how)
Identify the model formula: y ~ x * g (suppose you already know how)
Use an R function to fit the model (perhaps this is where you are unsure)

Break your problem into smaller parts

Rather than prompting

I want to model how `y` changes with `x` based on group `g`, how?

Ask specifically about what you don’t know!

I have data with columns `y`, `x`, and `g`. In R I want to model `y ~ x * g` but not sure which function to use, how could I figure that out?

Push back on the LLM

if the answer is too broad, further refine your prompt to focus the LLM, think back to breaking up the problem into smaller steps and identifying which you need help with

I have data with columns `y`, `x`, and `g` where `x` and `y` are numerical and `g` represents different groups. In R I want to model `y ~ x * g` but not sure which function to use, how could I figure that out?

Push back on the LLM

if the answer seems unnecessarily complex, ask it, “is there a simpler way to do this?”
LLMs can really over think things using the tidyverse so ask about non-tidyverse solutions
ask the LLM to explain steps that seem opaque or that you simply don’t understand

Check the LLM’s answers

run the code yourself, if there are errors, tell the LLM (paste the error message in) and make it fix the mistake
try asking questions about steps you are fairly sure you already know, if the LLM agrees this will help you know it is headed in the right direction

Never copy paste code from the LLM and call it done

make sure you understand every line
re-write the comments in your own words
you should have used hypotheticals and not the real data, so at a minimum you will need to change variable names

Ask for explanations

show the LLM code or ideas you don’t understand and ask it to explain
topics in this course have a lot of material written about them, so LLMs should be well trained
try questions like these

Ask for explanations

what is this code doing

```
# insert code here
```

Ask for explanations

why would I need a random effect for data like _blank_ and a model like _blank_?

Ask for explanations

what is different about the lm function versus the glm function?

The “what if…”

What if you don’t follow these guides?

First off, we’re all learning LLMs together, so you might have your own guides that work just fine
But second, if you don’t follow these guides nothing will blow up…not right away anyway
The concern is that without understanding the code and concepts you could introduce errors into your analyses
Sometimes such errors could have serious consequences