New OpenAI Offering Can Generate Code From Spoken Words 

By John P. Desmond, AI Trends Editor 

A new machine learning tool from OpenAI that can translate spoken words into code, an offshoot of the GPT-3 large language model released by OpenAI last year, is seen by development experts as a potential assist to programmers and not a threat to their existence.  

In recent demos of the new product, called Codex, experts from OpenAI showed how the software can be used to build simple websites and games using natural language. It can also translate between programming languages and handle data science queries. A user can type an English command, such as “create a webpage with a menu on the site and title at top,” and Codex is capable of translating the command into code. 

“We see this as a tool to multiply programmers,” stated OpenAI’s CTO and co-founder Greg Brockman in an account in The Verge. “Programming has two parts to it: you have ‘think hard about a problem and try to understand it,’ and ‘map those small pieces to existing code, whether it’s a library, a function, or an API.’” The second part is tedious, he says, but it’s what Codex is best at. “It takes people who are already programmers and removes the drudge work.”  

OpenAI had built a tool based on GPT-3 called Copilot, to experiment with automated code generation. Developers are allowed to test it via download from GitHub, the code repository now owned by Microsoft, which holds an exclusive license to technology from OpenAI. Codex builds on the functions of Copilot, making it more powerful, by for example not only completing code but creating it. Codex was trained specifically on open-source code repositories it scraped from the web.  

This evolution has led some coders to complain that OpenAI is profiting unfairly from their work. Asked about this by The Verge, Brockman stated, “We do need this debate.” His position is that most in the coding community will benefit from OpenAI’s work. “The net effect is a lot of value for the ecosystem,” he stated.   

Codex Seen as Continuing Evolution of Programming  

Wojciech Zaremba, a cofounder of OpenAI and lead for the Codex projects, sees the tool as the next step in the evolution of programming, which has evolved to be less cryptic and more intuitive with more English words included over time.  

“Each of these stages represents programming languages becoming more high-level,” stated Zaremba. “And we think Codex is bringing computers closer to humans, letting them speak English rather than machine code.” 

Codex itself can speak more than a dozen coding languages, including JavaScript, Go, Perl, PHP, Ruby, Swift, and TypeScript. It’s most proficient, though, in Python.  

To those concerned that Codex could put programmers out of business, Thomas Smith said not to worry. “Codex won’t replace human developers any time soon, though it may make them far more powerful, efficient, and focused,” he stated in a recent account in IEEE Spectrum. Smith is cofounder and CEO of Gado Images of San Francisco, a firm using AI and ML technologies to capture and share visual history, who as an early beta tester, put GPT-3 and Codex to the test.  

The reason is, most software developers spend 20% to 50% of time on software projects writing code, and the rest learning from clients about how they do their work, so that the automated system can do its job well. “Unless someone trains Codex to sit down with clients, win their trust, understand their problems and break those problems down into solvable component parts, the system won’t threaten skilled human developers any time soon,” Smith stated.  

Automating Some of the “Grunt Work” of Programming  

Instead, the creators of Codex suggest that the tool will help to automate the “grunt work” associated with software development, thus opening up the field to a wider range of people, and maybe creating a new specialty: “prompt engineering.” This is the crafting of text prompts to allow the AI system of Codex to do its work.   

Demand for software developers grew by 500,000 in 2020, to reach a total of 24.5 million, according to the 2020 Worldwide Developer Population and Demographic Study reported by Evans Data, in an account in InfoWorld. Growth did slow to 2.4 percent during the pandemic year, compared to a predicted four percent.   

Moreover, Codex is not able to write as well as a human developer. “At the moment, it absolutely cannot,” stated Smith in IEEE Spectrum. “That’s why many in the tech community see Codex less as a generator of new code, and more as a powerful tool to assist humans.”   

OpenAI’s researchers were curious to see how software developers would use GPT-3 for natural language processing applications. The outcome surprised them. “The applications that most captured people’s imaginations, the ones that most inspired people, were the programming applications,” stated Brockman in a recent video demonstration of Codex reported in TechTalks. “Because we didn’t make the model to be good at coding at all. And we knew that if we put in some effort, we could make something happen.”  

The result of that effort was Codex, which has a 37% accuracy on coding tasks, versus GPT-3’s performance of zero pertinent, according to Ben Dickson, the software engineer who founded TechTalks.    

In his account of the Codex demo, Dickson described how the presenters used the tools to create a user interface in JavaScript, place objects on the screen, and make the objects controllable with the keyboard arrow keys. Another video showed Codex generating data science code and generating charts in Python’s matplotlib library.  

“These are not complicated tasks, but they’re tedious and error-prone processes, and they usually require looking up reference manuals, browsing programming forums, and poring over code samples. So, having an AI assistant writing this kind of code for you can save some valuable time,” stated Dickson.  

Dickson cautioned that Codex is a work in progress, with some evident flaws. “Sometimes, the model generates code that is very far off from what the developer intends,” he stated in TechTalks.  

This observation about Codex was similar to that of Jeremy Howard, an AI researcher who founded Fast.ai, a firm that works to make deep learning more accessible. “It’s a way of getting code written without having to write as much code. It is not always correct, but it is just close enough,” Howard stated in a recent account in  The New York Times.  

The consensus among these early testers of Codex is that it works better with a human controlling and overseeing the code it produces.   

“A.I. is not playing out like anyone expected,” stated Brockman, OpenAI’s CTO. “It felt like it was going to do this job and that job, and everyone was trying to figure out which one would go first. Instead, it is replacing no jobs. But it is taking away the drudge work from all of them at once.”  

Read the source articles and information in The Verge, in IEEE Spectrum,  in InfoWorld, in TechTalks and in The New York Times.