GPT reads the comics
I needed to generate some data for a test I was writing for my Intro to Programming class. So I asked GPT (3.5 originally) just to give me a list of fruits and their colors. It provided me with a great list which I started to paste into my code and convert to JSON when I realized, GPT should be able to do that for me. So I asked it to and it did a great job.
Then I wanted to add some more attributes to my fruit data so I asked it this:
Do you know the XKCD comic that ranks fruits on a 2-dimensional graph of how tasty they are vs how much trouble they are to eat?
Yes, I'm familiar with that XKCD comic! It's called "Fruit Collider" and was published on August 10, 2011. In the comic, fruits are plotted on a two-dimensional graph with tastiness on the x-axis and inconvenience (difficulty to eat) on the y-axis. The fruits are then labeled and connected by lines to show which ones are most similar in taste and inconvenience. The comic is a humorous take on the subjectivity of taste and the trade-offs between convenience and flavor when it comes to choosing what to eat.
You can find the comic on the XKCD website at the following URL: https://xkcd.com/937/
Please note that the comic is intended for amusement and shouldn't be taken as a definitive guide on the tastiness or difficulty of eating different fruits.
Amazing. So I asked it to add some scores to my data:
Can you add tastiness and inconvenience rankings to the list of fruits above based on that comic?
Sure, I can certainly give it a shot! Based on the Fruit Collider comic, I would assign each fruit a tastiness score from 1 to 10 (with 10 being the tastiest) and an inconvenience score from 1 to 10 (with 10 being the most inconvenient to eat). Here's the updated list:
And it did a pretty darn good job. Obviously it's just a comic so there are no exact numbers to check but it seemed to at least be directionally correct.
However, then I decided to fact check the first part of the answer. Turns yout GPT was not all that familiar with the comic because “Fruit Collider” is not the name of the comic in question, nor was it published on August 10, 2011.
The actual comic I was asking about, and which GPT did more or less describe, is called “Fuck Grapefruit” and was published on February 25, 2008 as https://xkcd.com/388/. “Fruit Collider” is a different XKCD comic published on January 31, 2018 as https://xkcd.com/1949/. And the comic GPT linked to is titled “TornadoGuard” and was published on August 12, 2011 and has nothing to do with fruit. Also the description wasn’t quite right: the fruits are labeled but they are not “connected by lines to show which ones are most similar in taste and inconvenience”.
GPT-4 does a bit better, getting the number of the comic right, providing the right URL, and not hazarding a publication date.
Do you know the XKCD comic that ranks fruits on a 2-dimensional graph of how tasty they are vs how much trouble they are to eat?
Yes, I'm familiar with the comic you're referring to. It's XKCD comic #388 titled "Fruits." The comic presents a humorous 2-dimensional graph that plots various fruits based on how tasty they are against how much trouble they are to eat. The fruits are scattered across the graph, and some are labeled with funny annotations that provide additional context or humor.
You can view the comic here: https://xkcd.com/388/
Please note that the comic is intended for amusement and shouldn't be taken as a definitive guide on the tastiness or difficulty of eating different fruits.
It did however still get the title wrong, perhaps because it's not supposed to say “fuck” and it halucinated the bit about some fruits being “labeled with funny annotations that provide additional context or humor”.
Overall, I was pretty impressed by it’s ability to do the thing I actually cared about, namely, generate a bunch of fake data loosely based on the XKDC comic. But the fact that almost every factual statement it made was wrong reminded me again how little you can rely on GPT as a source of facts that you are not able to independently verify.