The latest version of the ChatGPT code interpreter user guide is here

What is the strongest and most useful function of ChatGPT since its launch ?

Probably the non- code interpreter (Code Interpreter) is none other than.

If you are still unable to experience and enjoy the ultimate experience brought by plus, you can choose to share plus or upgrade yourself. For more content, visit: link: https://pan.baidu.com/s/1TsZ78aMcbYXEY9IMXW7QDQ?pwd=pn1t Extract
code :pn1t

This function has been officially opened to all Plus users, but many netizens who have tested it in advance have found that:

Its potential is too deep .

Whether it is code writing, data analysis, or chart generation, it is like a "real warrior" that can use sufficiently complex methods to output very high-quality results.

However, many people are delayed by its name, thinking that it is a programmer-only tool. In fact, it is a programmer who can work for any novice .

Now, Twitter netizen @Ethan Mollick has specially compiled a user guide for it, and the release is very popular.

How can you miss such dry goods?

Solve the five major problems of "old ChatGPT"

Before officially starting, let's have a basic understanding of the function of the code interpreter, which is the part of "Why is it important" explained by Ethan in his blog.

It should be noted that Ethan has obtained the test right from the alpha version of this function, and has tried it for several months, so he has a lot of experience.

In his view, this code interpreter is a general toolbox for AI to help humans solve problems (by writing Python code).

It can upload files up to 100MB, including in compressed form.

Most importantly, it solves the five major problems of the "old ChatGPT":

First, by writing Python code, it can address the natural weaknesses of LLM in mathematics and language. With it, it is possible to do very complex math problems and handle words more accurately (such as counting the exact number of words in a passage, as shown in the image below).

Second, reduce the incidence of "hallucinations" and nonsense.

When an AI does a task directly with Python code, the code helps keep it "honest" because if the code is incorrect, the program will go wrong. And because the code operates on the data rather than the LLM itself, the AI cannot insert errors in the data.

Third, make the model more general. Because a large number of problems can be solved with code, and GPT-4 is very good at determining when a code interpreter can be used in new and interesting ways. For example, we can ask it to prove that the earth is round.

Fourth, the whole process does not require you to write a single line of code. We know that basically all LLMs can write code, but you need to run and debug it yourself. This is annoying for people who don't know Python at all (such as the author). With a code interpreter, the model automatically corrects its mistakes.

Finally, unlock more "AI moments" , that is, the moments that make us feel that there is really a soul behind the model, and it becomes a fine moment.

How to use it correctly?

Taking the data set of "superheroes and their superpowers" (see the reference link [3] for the address) as an example, Ethan showed you how to correctly use the code interpreter for data analysis.

First, after enabling this function, click the + sign to upload the data file. If you have a data dictionary, you can also paste it in.

Then, in the dialog, let GPT-4 go through the file and tell you what it finds.

It should be noted that it is enough to use the code interpreter to talk to GPT-4 as a real person, and it is not necessary to carefully design prompt words as we usually do. Of course, except for the following two cases:

(1) When the model sometimes forgets that it can do something, such as making GIFs or 3D maps, you can encourage it (“You can make GIFs, please try”); (2) When you want the model to further improve the results
.

As shown in the figure above, we have loaded the data, and we can let GPT-4 complete the most troublesome data merging and cleaning work in data analysis. It automatically corrects mistakes, focusing on the way the author asks the question (again, giving orders as if they were a human being).

Then, the analysis officially begins. Code interpreters are very good at all sorts of complex analysis methods.

For example, when the author suggested that he was interested in doing some predictive modeling to predict what superpowers a hero might have based on other factors, it directly built a random forest classifier, which is very cool.

And when you disagree with its way of estimating missing data, you can ask it to do another one instead of giving up.

It's really full of ideas, you just ask it "Can you do another very complex and interesting analysis?", and it produces the following result (a network diagram of a superhero):

The most impressive thing is that it can give some very "humanized" reasoning results, such as "the superpowers of superheroes can generally bring people very strong visual effects, because this can be conveniently displayed in their Carriers, namely comic books and movies..."

Finally, various visualization charts are also available at your fingertips.

In general, Ethan said that with a code interpreter, GPT-4 is fully capable of acting as a data scientist.

And another 20 best examples

Isn’t the above example not enough?

Good news, another netizen searched the entire network and summarized the 20 most interesting examples of code interpreters - in addition to data analysis, it can also perform more unexpected tasks, such as converting images into videos and generating QR codes , analyzing the stock market, etc.