In the past few days, the big trick that has been hidden by OpenAI for several months has finally been released, that is, "Code interpreter", code interpreter.
Friends may feel a little scoffed when they look at this name, isn't it a thing that helps write code, similar plugins are everywhere, what is the use of you coming out now?
Huh! Although its name is Code Interpreter, it doesn't do it write code for you.
Simply put, it is equivalent to a translator for you and the AI, which can translate your needs into specific solutions that can be solved by programs through natural language. It also gives you a 100MB file space to upload the files you need to work on.
A few months ago, Code interpreter was opened to some developers to try, including bloggers after a period of trial, found that after using Code interpreter, ChatGPT has made a qualitative improvement, such as ChatGPT can now solve very complex mathematical problems, the incidence of gibberish has also decreased, and it is better at solving practical applications and so on.
So Shichao also can't wait to go on the OpenAI official website that has not been on for almost a month to see how powerful this code interpreter is.
The first step in trying out the code interpreter is to open the code interpreter in "Settings" - "Beta features".
Then under the GPT-4 selection box, select Code Interpreter, so that you can try it smoothly.
Shichao usually also cuts videos, so the first moment I use this function is to ask it to cut a video for me.
Shichao first uploaded a 115-second video, and then told ChatGPT to help me cut out the middle ten seconds.
As a result, ChatGPT first scolded me, saying that my description was unclear and that more information was needed.
Well, that's really wrong with me, so I told it was the middle ten seconds.
It is clear that after a lot of code processing, ChatGPT outputs an "overwatching_ subclip.mp4" file.
After downloading it, it is indeed a clip of the original video at 50-60s, and it is exactly 10s.
Next, I uploaded a table stored in the warehouse, asked it to list items with more than 50 remaining inventory, consolidated them into a new table, and made an animated histogram GIF.
But the code is given, the problem is understood, and it lets us run the code ourselves.
He also said that his own runtime environment does not support direct generation of animated GIFs.
But didn't I use you because I didn't want to write code? What kind of trouble are you making me run by myself?
So Shichao gave ChatGPT a little encouragement, telling it that it can generate GIFs and bravely try it.
Unexpectedly, it actually responded, directly generated a GIF file for us, and gave us a download link, which shows that encouraging education is still very useful.
It's just that the generated GIF data image may be because the animation description is not very clear, it has been moving around, and the text encoding of the Chinese part is also garbled, but overall, the task is completed quite well, and the rest is some details.
Later, I asked ChatGPT to analyze the characteristics of objects with more than 50 inventory quantities, and then made a pie chart representation.
This time, ChatGPT first listed the form, and then drew a pie chart, and the analysis said that "DBTW DuPont paper bags" accounted for the largest part.
Shichao continued to ask what are the characteristics of these items, and ChatGPT was also well summarized based on the data.
I have to say that when I tried it here, Shichao was already a little convinced, and when I could directly help me cut the film before, Shichao felt a little surprised, and this report statistics really improved work efficiency.
In the follow-up, Shichao tried other functions of the code interpreter, and I first asked if it could help me think about using code to show something that seemed impossible but was actually not.
This time ChatGPT gives a code to draw a 3D cube, and after running, it generates a 3D cube projected on a 2D plane.
And ChatGPT also told Shichao that this is a basic concept in computer graphics, which is widely used in a variety of applications, including games, movies and virtual reality.
Later I also tried many other problems, such as drawing a fractal pattern.
There are also plot functions, and a scatterplot of the dataset is drawn.
It can be said that with this code interpreter, many things in the future work are basically just talking.
In my opinion, the code interpreter released by ChatGPT this time is very similar to the open source project AutoGPT that caught fire some time ago.
But in contrast, the code interpreter can not be networked, the ability is much stronger than AutoGPT, not only can understand our needs, find solutions to solve, but also automatically help us draw the data chart we need.
It can be said that it is truly a multimodal realization of large language models.
Finally, Shichao also asked ChatGPT about a mathematical problem that has not yet been proven - the Riemann hypothesis.
Unfortunately, no matter how much Shichao encourages it, this time, ChatGPT says nothing and doesn't answer anything...
It seems that it is still not possible to let AI solve some problems that humans cannot solve.
But at the current pace of AI development, no one can say.
Perhaps after 925 iterations, even the ultimate question of the universe, AI can give an answer in a few seconds.