🛠️📊 Talking to an Excel: The Birth of ChatGPT-4.5

If you haven’t been closely following the evolution of language models, you might feel that ChatGPT “hasn’t changed” since its launch last November.

It’s the same: a chat box where you can ask questions. And that’s it.

But the truth is that behind the scenes, there have been incremental changes, similar to what happens with iPhones year after year, when the iPhone 13 is replaced by the iPhone 14.

In the same way, OpenAI’s models have been evolving:

And last week, what I consider to be version 4.5 was released (even though no one has called it that). OpenAI gave all paying users access to a feature called “Code Interpreter.”

Beyond its terrible name, this release gives ChatGPT a powerful “toolbox” that allows it to do very advanced things.

Following the tradition of names that increment by 0.5, this new version represents such a significant leap in functionality that I dare to call it GPT-4.5, even though there have been no changes to the underlying model (although

some

say there have).

As the name suggests, it allows ChatGPT to write code. Don’t worry if you don’t know how to code; that’s the revolutionary aspect of this tool: you don’t need to know programming to do amazing things with code.

Let’s go with an example.

Talking to an Excel

First, something important (and perhaps disappointing for some): this feature is only for ChatGPT Plus users. To use it, you have to pay the USD 20 (!) per month that the service costs.

When you do, you will be able to choose “Code Interpreter” from the dropdown list, and you will see the option to upload a file.

For this example, I’m going to work with this Excel:

A database of 1,000 movies from

IMDB.com

.

First, let’s see if ChatGPT can read it.

The magic is in that “Finished Working” box at the beginning. If you open it, you can see the code that ChatGPT ran to read the file and understand the data.

Again, the interesting part is that you don’t have to understand that code. It’s just a tool that ChatGPT will use to help us review the data.

I can ask some simple questions:

Or some a bit more complex:

If you look at the original file, you will notice two things that make this question difficult to answer directly using Excel formulas: the movie genre column can contain more than one genre separated by commas, and the duration column is of text type and has the word “min.” For Code Interpreter, that’s not a problem.

Since it’s ChatGPT, I can ask for things in any format. Like a table:

That’s considered :)

In that case, it’s better to export that information to work on it elsewhere:

And if you’re wondering, yes. It works.

It’s also excellent at making simple graphs.

Or some a bit more advanced.

Or ask for crazy correlations like this:

Then I asked for a word cloud using the titles, and something very interesting happened:

It tried to create the word cloud but encountered errors in the data. Without me saying anything, it correctly identified and resolved the error 🤯.

It’s so “intelligent” that it suggests generating a new word cloud by removing “stop words” (articles and connectors that don’t add value to the cloud). The result is this:

And with Code Interpreter, there’s no limit to the type of analysis that can be done.

I thought of doing something silly, which perhaps data scientists reading this will find blasphemous, but it illustrates the power of the tool.

I asked it to build a predictive classification model using Random Forest (a commonly used Machine Learning algorithm to try to predict a variable based on other features).

ChatGPT used the variables it had (movie duration, year, genre) and trained a model capable of predicting (probably very poorly) a movie's rating based on those variables.

And I passed it a movie that hasn’t been released yet.

High expectations for the new Nolan movie! (

Here

you can see the complete interaction with ChatGPT).

Beyond the obvious conclusion that such an analysis makes no sense (why would the duration or genre of a movie determine if it is good or bad?), the important thing is that ChatGPT managed to train a complex predictive model, which would take a data scientist a couple of hours, without me having to write a line of code.

This is just an initial interaction I had with the tool just days after its launch, but its potential is much greater.

I’ve seen on Twitter how some manage to:

• Read text in photos of tickets.

• Generate QR codes for links.

• Create animated GIFs of the matrix.

• Visualize data on maps.

• Edit videos.

• Extract the color palette from an image.

• Create an asteroid-type game.

ChatGPT-4.5 has arrived, and I’m sure that in the coming days we will find more and more unique, extravagant, and useful use cases.