Xiao Xiao originated from the Cave Fei Temple
Throw ChatGPT a requirement, it can only give you a bunch of code that runs in unknown effect?
Now, with just one small change, ChatGPT can not only directly turn your text requirements into code, but also help you run through and return the output to you!
The secret is the code interpreter plugin.
Writer Andrew Mayne (now joined OpenAI) qualified for the closed beta and tried it, and the effect made him call "magic" -
Ask ChatGPT to write an AI face detection program, and Andrew Mayne uploads his own photos and is quickly detected:
Try it again to write a maze generation algorithm and make the path into a Pac-Man animation, which is completely fine:
Some netizens sighed after seeing this closed beta article:
It's crazy and it's going to change everything.
Let's see what magical "magic" Andrew Mayne found in the closed beta.
What magical "magic" has been discovered?
As one of the most high-profile ChatGPT plugins out there, Code Interpreter is a Python interpreter that works in a sandbox, firewall execution environment and contains some temporary disk space.
Simply put, you only need to enter a sentence into ChatGPT, which can not only write the code, but also run through the code with the help of the interpreter, give an explanation of each line of code, and output the result to you:
The content generated is also diverse, from text, image and sound processing, to board game engines and simple AI algorithms, ChatGPT can handle it all.
Text image sound processing
Let's start with the image generation and processing effect.
For example, convert a Cthulhu image to ASCII from ChatGPT:
Spawn a cat with a hat and a pipe (a bit abstract):
Generate a website QR code:
It's okay to make a relatively simple animation, such as "Blizzard":
Let's look at the sound.
For example, generate Shepard tones (a tone that sounds like it's rising or falling, but is really just an infinite loop of tones):
Not only is the explanation given, but a tone is generated, which sounds like it really means that.
Finally, word processing, taking the OCR algorithm as an example, which is an algorithm that recognizes images as a piece of text:
After uploading an old photo, the AI quickly recognized the corresponding text and converted it into a .txt file, which also looks good:
So, take it a step further and try writing algorithms in ChatGPT?
Write AI algorithms
In addition to the face recognition algorithm mentioned at the beginning, which was generated by ChatGPT based on OpenCV, Andrew Mayne also tried NLP algorithms this time.
Specifically, given half of the sentence, let the AI predict the next word, ChatGPT quickly gives the code:
Looks good, how does the build work?
Andrew Mayne tried:
They went to the...... (They went...) ）
Usually a place noun should be taken here, such as beach or church. However, the AI algorithm written by ChatGPT took "shoulders" after thinking about it???
They went to the shoulders
Andrew Mayne teased that there was no need to worry about ChatGPT writing a GPT-4.
Build the game engine
Finally, there's the generation of the game engine, which Andrew Mayne says is "totally fine," but doesn't give specific steps for ChatGPT.
This is the effect of generating a checkers engine (in addition, a chess engine):
It can also be used to generate a Game of Life:
The program is not written 100% correct
Some netizens said after reading that the code interpreter plugin well fills the loophole of ChatGPT's "poor math":
ChatGPT has few math skills, but it can easily do it by generating Python code.
But there are also many netizens who found a lot of bugs in this demonstration.
For example, although they are all Python programs that can run, there are basically obvious small errors when you look closely, including:
The method of generating Shepard tones is wrong, not only the amplitude must be tuned, but also the frequency;
The resulting Shepard tone itself is also problematic, normally the last tone should be the same as the first, so that the loop can continue indefinitely;
Using a cellular automata to generate a QR code image is actually the reverse of using a QR code image to push out the cellular automata (however, here the author also admits that he "taught" ChatGPT to cheat)
The orbits in the planetary orbit simulation are all circular orbits that should be clearly elliptical from a scientific point of view.
Graphics generation actually has minor bugs, so I won't list them all.
However, "these obvious bugs don't need to be too much of a concern." ”
Some netizens feel that what really needs to worry about is the problem of ChatGPT "overlay error".
As soon as something ChatGPT writes is buggy, a new session must be restarted immediately, otherwise its errors will start stacking exponentially:
It feels like it's trying to hide something, very strange.
Finally, although ChatGPT can already write chess engines, sharp-eyed netizens have found that it still can't play chess.
For example, suddenly eating his own soldiers with a horse. (Manual dog head)