So if you’re coming to this video from the Illumine Insights newsletter, we had just left off having had the chat CPT code interpreter create a report in a PDF format. Honestly, if you looked at the PDF, the results were a bit underwhelming. We’re going to try to take a different approach.
One side note, if you come across this video on YouTube, not from the newsletter, I’d advise you to subscribe to the Illumine Insights newsletter. It’s free, and there’s a link to subscribe in the description of the video.
Here we go. We’re already logged into ChatGPT. I’m a Chat DPT Plus subscriber, so I do have access to code interpreter. We’re going to click on GPT4 and then click on code interpreter, which is still in beta.
What we’re going to do is give the code interpreter a pretty descriptive message about what we’re looking for. We’re going to upload the same insurance dataset that we looked at in the newsletter. We’re not going to give any other context. This is a totally new chat, so it doesn’t have any of the context that we provided or as we walked through the production of the newsletter.
First, we’ll go ahead and upload the file. Here is the insurance dataset from our newsletter. Then we’re going to use the prompt we’ve already written. For this dataset, create a downloadable PDF report with a comprehensive descriptive analysis of the data, appropriate data visualizations, and the development of a predictive model. The analysis should include an in-depth discussion of the practical implications of this analysis.
I’m a professor, and this is similar to what I would ask of my students on a small project or a homework assignment. So, I have not seen the results of this. We’re going to see this together for the first time. You’re not going to hear me say much of anything as it does its work. It’ll probably take a couple of minutes to work through this process, and I think it’ll be interesting to see how this goes.
So let’s go ahead and feed the dataset and the prompt to ChatGPT’s code interpreter and see what we get. So far, so good. I’m pretty impressed with what we’re seeing so far. It does give us a little caveat at the end about this being a simplified analysis. That’s fair.
Where I’ve seen a code interpreter struggle a little bit in the past is the creation of a PDF. Let’s see how it does with this step. Alright, it was successful there. Let’s download this thing and see what it looks like.
In the newsletter, I posted the PDF it generated before, and it was not great. Let’s see if this is better. So, we’re going to open this up. Yeah, that’s not really great at all. It seems like we didn’t get the visualizations. This is definitely not what I would call a report. Maybe we can tweak the prompt a little bit.
You know, it’s one of the nice things. Let’s go back and take a look at this and see what we can do. Let’s just come up with a new prompt on the fly. Thank you, but this report does not appear to be in the format of an actual report and does not include discussion of the problem, the predictive model, etc. Let’s see what happens when we do this.
So, a little caveat thrown in there by ChatGPT about the difficulty of placing images inside PDFs. It basically throws up its hands at this point and says, ‘Here’s what a report would look like,’ but doesn’t actually generate the report itself.
Okay, let’s try one more thing. Again, we’re just doing this on the fly. I don’t know why I feel the need to always be polite to ChatGPT, but it seems to be the right thing to do. It’s kind of like you are talking to an ascension being here. So, saying ‘Okay’ and ‘Thank you’ and stuff like that seems to just come naturally.
Interesting. This is kind of strange because I was able to generate a PDF ‘report’ literally yesterday. It comes in and tells us, ‘Okay, unfortunately, the current system only supports the generation of PDFs from figures. It doesn’t support the generation of text-based PDFs.’
So, that’s kind of odd. Yeah, it’s definitely falling a little bit short of what I would do next. You know, as a teaching tool, demonstrating some of the ways to do some basic reporting of results, some appropriate visualizations, it’s pretty darn good. So, I’m looking forward to what the future holds on this type of thing. Deeply impressed. This is really good stuff. However, you know, is it ready for anybody to just sit down and develop a full-fledged analysis, throwing a dataset at it and say, ‘You know, do the work’? I’m not sure that it is. I think we’ve demonstrated, at least from this example, that it is not ready to do that. It still requires some intuition, some guidance from the user. So, data scientists, I don’t think your job is safe for the moment. But as a teaching tool, demonstrating some of the ways to do some basic reporting, it’s pretty darn good. I’m looking forward to what the future holds on this type of thing.
