AE07 - Working with AI generated Code

Be the boss of your AI helper

Purpose

An LLM can hand you code that runs and is still bloated, built on guesses about your data, or quietly wrong. You are responsible for what you submit. Today you will read real ChatGPT output critically, see how much the code depends on what the model knows about your data, and practice cutting code down to the part that does the job.

By the end you should be able to:

• Spot the assumptions an LLM makes when it can’t see your data
• Cut bloated code down to the minimal tidyverse version and justify each deletion
• Decide what is safe to delegate to an LLM and what you have to understand yourself

Instructions

• Work in groups of 2-3. Talk through each question together, then individually write answers on the front page of your worksheet. Detach the front page from the back page, and turn it in by the end of class. The ChatGPT answers are on the back pages.
• The time estimates next to the suggestions are guidelines

The data (parts 1-3)

All of the code today is about the same two data sets from the Exam 1 take-home.

votes: California county presidential returns, 2008 to 2020. One row per county, year and candidate.

county_name	year	county_fips	candidate	party	candidatevotes	totalvotes
ALAMEDA	2008	6001	BARACK OBAMA	DEMOCRAT	489106	621029
ALAMEDA	2008	6001	JOHN MCCAIN	REPUBLICAN	119555	621029
ALAMEDA	2008	6001	OTHER	OTHER	12368	621029

(There are other columns too: state_po, office and version.)

census: one row per US county (usdata::county_complete). The columns that matter today:

column	meaning
fips	County code. Stored as text, so it has to be converted to a number to match county_fips.
name	County name, such as “Alameda County”
bachelors_2019	Percent of adults with a bachelor’s degree

Part 1: A vague prompt (~10-15 min)

This is how most people actually write a prompt: short, with the details left out. This is what was typed into a new ChatGPT chat:

I have election results for California counties and a census data set. Find the five counties with the highest Democratic vote share in 2020 and their bachelor’s degree rate. Write R code.

WarningYour turn

Refer to ChatGPT’s answer on your worksheet, and compare it to the data above.

1. Could you run this code on our data as written? Yes or no, with 1 sentence justification.
2. Name one assumption the model made about the data.
3. Underline the steps that exist only because the model could not see the data.

Part 2: Trim the function (~10-15 min)

In the same chat, the next thing typed was:

Now turn that into a reusable R function that takes a year and a number n, and returns the top n counties.

The worksheet has ChatGPT’s function. We changed the names in it to match our data, and removed the example runs to fit it on one page.

WarningYour turn

1. Cross out everything that is not needed to do the actual job.
2. Next to each thing you cross out, write one sentence on why it is not needed.
3. Put a check mark next to lines you do understand and could explain to someone else.
4. Star any line you do not understand, have not seen in class, or could not explain to someone else.