Rich color comes from the temple of Wofei
Qubits | Official account QbitAI
What is the strongest and most useful feature of ChatGPT so far?
It may be the Code Interpreter.
This feature will not be officially available to all Plus users until next week, but many netizens who tested in advance found the same:
Its potential is too deep.
Whether it's code writing, data analysis, or graph generation, it's like a "real warrior" who can use sophisticated enough methods to output very high-quality results.
However, many people are delayed by its name, thinking that it is a programmer's exclusive tool, in fact, it is a programmer who can work for any little novice.
Now, Twitter user @Ethan Mollick has compiled a user guide specifically for it, and it is a hit when it is released.
Such a dry goods, how can you miss it?
Solve the top five problems with "old ChatGPT"
Before we get started, let's get a basic look at the code interpreter feature, which Ethan explains in his blog about why it's important.
It should be noted that Ethan has obtained the right to test from the alpha version of the feature and has been playing it for many months, so it is very experienced.
In his view, this code interpreter is a general-purpose toolbox for AI to help humans solve problems by writing Python code.
It can upload files up to 100MB, including compressed form.
Most importantly, it solves five major problems with "old ChatGPT":
First, by writing Python code, it addresses LLM's natural weaknesses in math and language. With it, it is possible to do very complex math problems and process words more accurately (such as counting the exact number of words in a paragraph, as shown in the figure below).
Secondly, reduce the incidence of "hallucinations" and nonsense.
When AI does tasks directly with Python code, the code helps keep it "honest" because if the code is incorrect, an error occurs in the program. And because the code manipulates the data rather than the LLM itself, AI doesn't insert errors into the data.
Third, make the model more versatile. Because a lot of problems can be solved by code, GPT-4 is very good at determining when to use the code interpreter in a novel and interesting way. For example, we can ask it to prove that the Earth is round.
Fourth, the whole process doesn't 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, like authors. With the code interpreter, the model automatically corrects its own mistakes.
Finally, unlock more "AI moments", that is, moments when we feel as if there is really a soul behind the model, and the moment has become perfect.
How to use it correctly?
Using the example of a dataset dealing with "superheroes and their superpowers" (see reference link [3]), Ethan shows you how to properly use the code interpreter for data analysis.
First of all, after turning on this function, click the + sign to upload the data file, if you have a data dictionary, you can also paste it in.
Then, in a dialog box, GPT-4 will look at the file and tell you what you found.
It should be noted that the code interpreter is used to talk to GPT-4 as a real person, and there is no need to elaborate the prompt words as we usually do. Except, of course, the following two cases:
(1) The model sometimes forgets that it can do certain things, such as making GIFs or 3D drawings, at which point 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 can let GPT-4 complete the most troublesome data merging and cleaning work in the data analysis. It automatically corrects errors and focuses on the author's way of asking questions (again, giving orders as if they were human beings).
Then, the analysis begins. The code interpreter is very good at a variety of complex analysis methods.
For example, when the author hinted that he was interested in doing some predictive modeling to predict the superpowers that a hero might have based on other factors, it directly built a random forest classifier, which is very cool.
And when you don't agree with its approach to estimating missing data, you can call it another way instead of giving it up.
It's really a lot of ideas, and you just have to ask it, "Can you do another very complex and interesting analysis?" It produces the following results (superhero network diagram):
The most impressive thing is that it can give some very "human" reasoning results, such as "superheroes' superpowers can generally give people a very strong visual effect, because it can be easily presented on their carriers, that is, comic books and movies..."
Finally, a variety of visualizations are at your fingertips.
Overall, Ethan says that with the code interpreter, GPT-4 is well positioned to act as a data scientist.
There are another 20 best examples
The above example is not enjoyable?
Good news, another user summarized the 20 most interesting examples of code interpreters on the whole network - in addition to data analysis, it can do more tasks than you would expect, such as converting images into videos, generating QR codes, analyzing the stock market, and so on.
Here are all examples or effects:
1. Convert pictures to videos
2. Extract text from images
3. Correctly use Monte Carlo, ARIMA, PCA, random forest and other data analysis methods
4. Make up science fiction stories with real physical phenomena
5. Extract the color palette from the image
6. Generate a QR code
7. Convert GIF to video
8. Analyze Apple option data to determine the best protocol
9. Data analysis on Spotify playlists
10. Real estate data analysis
11. Generate and plot random data
12. Generate a matrix rain of GIFs
13. Clean data, perform data analysis, and generate visualizations
14. Create a gif of each lighthouse location in Beautiful Country
15. Converting data to websites
16.Copy the STATA code in Python
17. Download and analyze the Bitcoin dataset
18. Visualization of crime data trends
19. Generate gene co-expression heat map
20. Exploratory analysis of the Titanic dataset
For more details and methods, please check the original address.
What other powerful features have you unlocked, welcome to share~
Original address:
https://www.oneusefulthing.org/p/what-ai-can-do-with-a-toolbox-getting
https://twitter.com/chaseleantj/status/1677679654680035328
Reference Links:
[1]https://twitter.com/karpathy/status/1677512911953231874
[2]https://twitter.com/emollick/status/1677313733071085569
[3]https://www.kaggle.com/datasets/claudiodavi/superhero-set
— End —
Qubits QbitAI · Headline number signed
Follow us and be the first to know the latest scientific and technological trends