Editor: Peach Layan
After waiting for more than 3 months, the ChatGPT "code interpreter" was finally fully opened, and a wave of magic struck.
Family, it's finally here!
Just now, the beta version of ChatGPT "Code Interpreter" was officially opened to all Plus users.
Karpathy, a scientist at OpenAI, also publicized a wave of this, and the code interpreter beta is quite powerful.
It's your personal data analyst: it can read uploaded files, execute code, generate charts, perform statistical analysis, and much more. I anticipate that it will take some time for the community to fully demonstrate its potential.
Some time ago, the decline in GPT4 ability caused an uproar on the Internet, and some people even felt that it was only a little stronger than GPT3.5.
And now, the code interpreter makes people feel once again that the familiar and powerful GPT-4 is back!
As soon as the plug-in was launched, many netizens got started and tested it.
Feature unbanning
So, what exactly can a "code interpreter" do?
In 30 seconds, the picture becomes video
No, some netizens actually used ChatGPT to turn pictures into videos. It's even more explosive than Midjourney and Runway.
How is this done? Netizens gave a full set of tutorials.
Step 1: Open the Code Interpreter Model. To enable it, go to Settings -> Beta Features -> Code Interpreter.
Step 2:
Select the Code Interpreter option from the drop-down menu. Then, upload the image you want to animate.
Step 3:
I would like to show some food from a luxury restaurant. Therefore, I asked ChatGPT to animate this food image from left to right. By modifying my GPT cues (see below), you can create your own animations.
The specific prompts are as follows:
Step 4:
Press Enter. After 30 seconds, ChatGPT will give you a download link for the video.
Co-founder Greg Brockman retweeted the code interpreter used for image conversion in cases where it is impossible to remember ffmpeg calls.
It turns out that the "code interpreter" is really good, and now take a screenshot of the video and you can do it in just one sentence.
"There is a certain memory limit, but it is no problem to handle small files. Just throw a 4-second gen-2 video to TA, capture the first two seconds, no problem, get it done quickly.
It will automatically experiment with the appropriate code, and if you want to compile the program, you can first throw it to a certain file to make a request, and then look at the code provided by ta."
PDF to picture
"Baoyu", a well-known domestic Internet technology blogger, used the "code interpreter" to convert PDFs into pictures.
In the past, you needed to merge each page into a single image, but now you can do it with one prompt.
GIF to video
In addition to changing pictures to videos, you can also change GIFs to videos.
The tips netizens gave to ChatGPT were: I upload a GIF, you give me a 5s MP4, and slowly zoom in with a story shot.
Upload GIFs
Finally, ChatGPT immediately generated a video, and ChatGPT can also make a "matrix" GIF.
Generate QR codes
Two days ago, OpenAI announced that the ChatGPT "code interpreter" will be fully available.
It shows an example of being able to generate QR codes, and in just one sentence, you can simply do the job with your mouth.
"Pocket" data analyst
The biggest use case for code interpreters is data analysis.
For example, let it analyze life expectancy data from countries, the world's largest cities, (can upload CSV, word, PDF or video files)
Then, talk to the data and ask any questions you might want to ask.
Or ask to extract unique random trends from your data to get surprises.
For example, identify unique trends from two pieces of data and display them in a table.
The final step is visualization.
Turn your data into any chart or graph you want. Best of all, it's just a click away.
For example, generate a scatter plot that illustrates the relationship between a country's GDP per capita and its average life expectancy.
Seeing such a powerful "data analyst", netizens began R.I.P again.
User Guide
How does the Code Interpreter work? Ethan Mollick, a professor from Wharton, wrote a guide specifically for this.
First, Mollick points out that the name Code Interpreter continues OpenAI's long tradition of giving their own products bad names.
Because — although it's called a "code interpreter," it's actually most useful for people who don't know programming at all.
Essentially, the "code interpreter" lets GPT-4 upload and download information and write and execute programs for the user, which allows AI to do all sorts of things that it couldn't do before.
Specifically, it provides AI with a general toolbox for solving problems (by writing code in Python), large memory (up to 100MB, which can be compressed), and users can integrate this toolbox into AI to take advantage of LLM.
This helps solve many of the problems with previous versions of ChatGPT. Say
It allows AI to process math (even if complex) and more accurately (e.g. actually counting words in paragraphs), for the underlying reason that it can write Python code to address LLM's natural math and language weaknesses.
For example, like the following figure.
As you can see, the "code interpreter" can accurately calculate the number of words in a paragraph.
In addition, the "code interpreter" reduces the chance of hallucinations and fabrications.
When AI uses Python directly, the code helps make it "honest." The reason is also very simple, if the code is incorrect, Python will report an error. Since the code manipulates the data, the AI does not insert error information into the data.
Of course, this is not perfect. AI still has the probability of hallucinating, but this is not common and is unlikely to affect the code or the data itself.
In this way, AI is closer to general artificial intelligence - a large number of problems can be solved by code.
Moreover, GPT-4 is very good at deciding when to use the "code interpreter" in a new and interesting way.
Mollick gave the example of how if he had GPT-4 prove to the user that the Earth was round, he would use a "code interpreter" that provided multiple arguments that combined text with code and images.
Previously, people who needed programming in various majors had to write a little code, but they had to debug it themselves after writing. Even if the previous AI can provide some help, it is still quite difficult for people who do not have a Python learning foundation, and they have to repeatedly modify the code with the help of AI.
Now, with the "code interpreter", AI can correct and debug its own!
Then, Mollick talked about how to use the "code interpreter" to process data, which is a real poke at a large number of user pain points.
According to Mollick, the Code Interpreter operates at a very advanced level, automating many complex quantitative analyses and being able to take very complex data methods.
Users can easily upload data, even compressed data like ZIP files, just by clicking the plus sign.
At the same time, when providing data, it is best if the user can give a prompt. The AI can then figure out the meaning and structure of the data from the context!
In addition, Mollick pointed out that when using the "code interpreter", users are more in a "conversation" with the AI, and do not need to rack their brains to write a rigorous and condensed prompt.
It's like chatting with a data analyst, how you talk to people, how you talk to a "code interpreter".
But there are two special cases where something like prompt needs to be written.
One is that AI sometimes can't remember what it can do, and it needs a prompt to wake up the AI.
For example: "You can make GIFs, please try it."
Another situation is that when the user is not satisfied with the result, you can use the prompt to let the "code interpreter" optimize the result.
For example, "Please test the results further" or, "Can you make this chart a little better?"
After loading the data, GPT can perform data merging and cleaning operations. It's quite complicated, but like mentioned above, just think of the "code interpreter" as a data analyst who meets you face-to-face.
At the same time, Mollick reminds users that it is best for AI to check the results after self-correction and correction, in case of error.
Immediately after Mollick got hands-on, he asked the "code interpreter" to do a predictive modeling to predict what superpowers the superhero might have.
As you can see, the "code interpreter" builds a random forest distributor.
Of course, in the actual operation process, users still have to "communicate more" with GPT. Mollick doesn't want AI to use numerical data to estimate missing data, Mollick wants to give up what is not.
But he pointed out that it is also possible to talk to the GPT and let it be treated individually.
Mollick then typed GPT a "Can you do another complex and interesting analysis?" prompt. He said he was impressed by some of the analysis methods chosen by the Code Interpreter.
It's like the image below.
The "code interpreter" cleverly analyzes the network situation of the superpower. As you can see, the methods are diverse and very interesting.
Perhaps the most impressive thing about the "code interpreter" is its reasoning ability. Mollick sums it up as very "human."
In addition, the level of interactivity of visualizations remains online and constantly improving. Users can constantly let the AI improve or change the output.
Mollick had the AI create an interactive table with 6 charts, one of which was 3D.
To make the chart more aesthetically pleasing, Mollick kept typing prompts to adjust the results.
The end result looks like the one on the right of the figure below.
Top 10 amazing magic
Netizen @GPTDAOCN recently summarized the top ten shocking use cases of ChatGPT:
Visualization of the code interpreter
1. Convert GIF to MP4: The ability to convert moving pictures (GIF) to video files (MP4) makes it easier for you to share and use.
2. Create various graphs: You can use this function to generate various graphs such as bar charts, line charts, pie charts, etc., to visually present and analyze your data.
3. Extract colors from images to create PNG color palette: This feature can extract the main color from any image and then generate a PNG format color palette.
4. Clean your data to generate in-depth visualizations: If you have some messy data, this feature can help you clean and organize it and turn it into clear visualizations.
5. Convert CSV file to GIF: This function can display the data in the CSV (comma separated values) file in a dynamic way to generate a GIF image.
6. Analyze playlists to summarize your musical tastes: You can use this feature to analyze your music playlists to summarize your musical tastes.
7. Convert your dataset into a fully functional HTML website: This feature quickly converts your data into a fully functional website for you to display and share.
8. Create downloadable stock datasets: This feature can help you create a stock dataset, and this dataset is downloadable for more in-depth analysis and research.
9. Convert image to text file: If you need to extract text from an image, this feature can help you convert the text in the image into a text file.
10. Detailed analysis of unformatted data: For raw data that is not in a prescribed format, this feature can help you perform in-depth analysis to obtain valuable information.
In fact, there are many more use cases for code interpreters.
Netizen comments
I'm trying the ChatGPT code interpreter. So far, I can only suggest that "every developer starts working on it now".
How many people have been cut off by Code Interpreter? Too many administrations process data inefficiently on a daily basis, and now only natural language is needed to automate the office.
Resources:
https://twitter.com/karpathy/status/1677512911953231874