Data Carpentry for Biologists

AI Assistance and Assistants for Coding

Notes

Have students install Positron before class

Problems with LLMs

  • There are a variety of meaningful ethical concerns about using LLMs
  • The use a lot of energy to train and run and therefore put a lot of CO2 in the atmosphere
  • They use millions of peoples work without credit or payment, arguably in violation of copyright and licenses
  • Writing code with them when lacking the background to evaluate that code can also be dangerous
    • They are good at giving you an answer, even if it’s the wrong one & code that is wrong but runs is the scariest kind of code in science
    • Bad code can accidentally destroy your work
    • Malicious hackers are finding ways to get models to generate code that lead to compromised computers
  • Cost
    • Start out with generous free plans
    • Most companies already increasing prices (and still losing money)
    • Expect costs to continue to rise
  • Privacy/confidentiality
  • Reduced learning making it difficult to go from beginner to expert
  • But they can be useful tools and we’ll spend today exploring how

What are you doing now?

  • First - How are you using LLMs for coding now?

Chat

  • UF provides access to a number of different model’s chat systems for free
  • UF Navigator
  • Navigate to: https://chat.ai.it.ufl.edu/

  • Sign in

  • So, many of you have already used chat interfaces
  • “How do you drop NA’s from a table in R”
  • Copy the resulting code into RStudio
  • Give it a quick read for anything bad
  • Run it

Improving output from LLMs

  • Provide details (value of knowing things)
  • Ask model to write code not return results
  • Provide model with context
    • Better prompts
    • Attaching files
    • Including relevant urls
  • If the resulting code doesn’t run, tell model what happened and ask for fixes
  • This set of practices is so common they are now integrated into many IDEs

Assistants

  • “AI Coding Assistants” provide a combination of chat, autocomplete, and local code execution

Activating the assistant in Positron

  • Click on ⚙️
  • Choose Settings
  • Type “Assistant”
  • Scroll down and select Enable
  • Restart Positron
  • Welcome -> New Folder -> R Project

Integrated Chat

  • Click on the 🤖 (bottom of sidebar)
  • Add a Chat Provider -> GitHub Copilot -> Sign in
  • On Authorize your device page paste code -> Continue
  • Authorize GitHub Copilot Plugin -> Close
  • Return to Positron
  • Type “How do I remove rows with na using dplyr”
  • Show at top of code block
  • From the top of code blocks you can
    • Run code in the console
    • Move code to the editor either in the current file or a new one
    • Copy it
  • Can then keep or undo changes
  • More useful if we give it context
library(dplyr)
library(readr)

surveys <- read_csv("https://ndownloader.figshare.com/files/2292172")
  • Save file
  • Show that file is in the context in chat
  • “Add a dplyr pipeline that removes rows with NA in the weight column from the surveys table”
  • We could copy this code over, or we can change the settings for the assistant to let it edit our files
  • Switch from “Ask” to “Edit”
  • Rerun query
  • Read and accept the edits

Inline assistant

  • In the text editor Ctrl-i
  • “Only keep the year, month, day, and weight columns”
  • Accept changes

Autocomplete

  • Hit enter after the last line
  • Look at the autocomplete
  • Interpret output
  • We typically don’t want an LLM guessing at what analysis we want
  • Start it with a comment
# calculate the average weight and number of individuals in each year in each month

Debugging

  • Let autocomplete do as much of this as possible
# Write a dplyr pipeline that returns only species starting with the letter "D" and
# where weights are greater than 46 then calculate the average weight of each species
  • Run the pipeline and debug. If the model doesn’t error introduce an error
  • Highlight code and select Review
  • There will soon be Fix and Explain options that pop up in the error message itself to further simplify this process

Vibe coding & Agent-based approaches

  • Who’s heard of vibe coding before?
  • Instead of writing code yourself with assitance from an LLM just have the LLM write a full first draft
  • Provide feedback to the LLM to make improvements to the code

  • Anyone have experience with this yet?

  • Change to Agent model
  • This will allow the model to actually run code for you
  • By default it will write the code, show it to you, and ask you if you want to run it

  • Be careful, I relatively recently asked a model to perform a git operation I didn’t remember how to do and the resulting code would have deleted by work from the last hour

  • “Using R create species distribution models for Great Egrets, White Ibis, and Roseatte Spoonbills. Use these models to make predictions for the distribution of those species in 26 years. Create a website displaying the current model predictions and the future predictions.”
  • Explore the code, emphasize need to read it
  • “Download the relevant data from gbif, update the code to use that data, and check that it works”