Artificial intelligence knows how to write computer programming code

When Tom Smith - an experienced programmer first encountered the Codex - artificial intelligence (AI) capable of writing computer software, he immediately conducted a 'job interview' for the program. this. Tom asked Codex to perform coding challenges that human programmers often have to overcome if they want to get high-paying jobs at technology companies in Silicon Valley (USA) like Google or Facebook. .

The Codex completely passed those requests in a split second, and even completed other requests later. 'Those challenges are not easy for humans, including me, but that AI handles it in seconds. It feels so weird to watch this process,' Smith said.

Tom Smith went on to assign other tests to the Codex and found that the AI ​​was not only skilled at answering questions that were boxed in interviews, but could also translate any programming language on its own. period to another type. Codex seems to be a technology that can replace human programming in the near future.

But after several weeks of 'working' with the new technology, Smith believes the Codex poses no threat to professional programmers. Like many other experts before him, he recognized this as a tool to help boost human productivity. The Codex can help shape a whole new generation of computer learners, by showing them how to write simple programming code line by line, like a 'tutor'.

'This will be a tool that makes life much easier for programmers,' Smith shared on NYT. In fact, despite more than a decade of development with many leaps and bounds, even the most impressive AI systems to date are only intended to serve people, not replace them.

Codex is a program developed by OpenAI. This is one of the most ambitious research labs in the world, dedicated to bringing insights into the world of artificial intelligence. OpenAI created an AI called GPT-3 that can write Twitter posts, write speeches, write poems or write articles. The program is based on a mathematical system commonly known as a 'neural network' that is capable of learning skills through the analysis of large amounts of input data.

GPT-3 also surprised the researchers who wrote this system when it demonstrated the ability to write computer programs on their own even at a short and simple level. The AI ​​is said to have learned this skill from programs posted on the internet. Therefore, OpenAI decided to train a new system to be proficient in both prose and programming. Finally, the Codex was born.

The system can understand both textual content and code (to a certain extent). For example, if you write a request (in English) to make a program to make snow fall on a black background, the Codex will return the code to 'make' the request. Or when asked to create a program containing a blue ball capable of bouncing in the screen, this technology will respond quickly. 'If Codex is wrong to write the code, the system will do exactly as requested,' said Ania Kubow - a professional programmer who has experienced Codex.

According to the NYT, AI can write software in 12 different computer programming languages ​​and translate between them. Even so, the Codex can still make mistakes and can't match a human. Sometimes the codex-written program doesn't run, or there's a security flaw, sometimes it just gets close to what's required. OpenAI estimates that the Codex programmed correctly only about 37% of the total execution times.

In other words, Codex is only really useful to experienced programmers because they will know where to edit in the answers that AI gives so that the program can run completely.

Jeremy Howard - founder of artificial intelligence lab Fast.ai thinks this is 'a way to code without having to write too much. Although the code returned is not always correct, it is enough for the programmer'. Howard also believes that the Codex can be of help to newbies learning how to code because it is possible to create simple programs from written requirements descriptions.

Joel Hellermark, an entrepreneur in Sweden is trying to perfect to turn this system into a tool to teach coding.

Kareem Winters

Update 10 September 2021