AI can now write its own computer code. This is good news for humans.
As soon as Tom Smith got his hands on Codex – a new artificial intelligence technology that writes its own computer programs – he granted him a job interview.
He asked if this might address the “coding challenges” programmers often face when interviewing for big budget jobs at Silicon Valley companies like Google and Facebook. Could he write a program that replaces all spaces in a sentence with dashes? Better yet, could he write one that identifies invalid zip codes?
He did both instantly, before completing several other tasks. “These are issues that would be difficult for a lot of humans to solve, myself included, and that would grab the answer in two seconds,” said Mr. Smith, a seasoned programmer who oversees an AI start-up called Gado. Images. “It was scary to watch.”
Codex seemed like a technology that would soon replace human workers. As Mr. Smith continued to test the system, he realized his skills went far beyond a knack for answering pre-set interview questions. It could even translate from one programming language to another.
Yet after several weeks of working with this new technology, Smith believes it poses no threat to professional coders. In fact, like many other experts, he sees it as a tool that will ultimately boost human productivity. It can even help a whole new generation of people learn the art of computing, showing them how to write simple pieces of code, almost like a personal tutor.
“It is a tool that can make the life of a coder much easier,” said Mr. Smith.
About four years ago, researchers from labs like OpenAI began designing neural networks that analyzed huge amounts of prose, including thousands of eBooks, Wikipedia articles, and all manner of other texts. published on the Internet.
By spotting patterns throughout this text, networks have learned to predict the next word in a sequence. When someone typed a few words into these “universal language models,” he could complete the thought with whole paragraphs. That way, a system – an OpenAI creation called GPT-3 – could write its own Twitter posts, speeches, poetry, and news articles.
Much to the surprise even of the researchers who built the system, he could even write his own computer programs, even if they were short and simple. Apparently he had learned from countless programs posted on the Internet. OpenAI therefore went further by forming a new system – Codex – on a huge array of prose and code.
The result is a system that includes both prose and code – up to a point. You can ask, in plain English, for snow falling on a black background, and that will give you a code that creates a virtual snowstorm. If you ask for a blue bouncing ball, it will give you some as well.
“You can tell it to do something, and it will,” said Ania Kubow, another programmer who used the technology.
Codex can generate programs in 12 computer languages and even translate between them. But he often makes mistakes, and although his skills are impressive, he can’t reason like a human. He can recognize or imitate what he has seen in the past, but he is not nimble enough to think for himself.
Sometimes programs generated by Codex do not run. Or they contain security holes. Or they are nowhere near what you want them to do. OpenAI estimates that Codex produces the right code 37% of the time.
When Mr. Smith used the system in a “beta” testing program this summer, the code he produced was impressive. But sometimes it only worked if he made a small change, like adjusting a command to suit his particular software configuration or adding a numeric code needed to access the internet service he was trying to query.
In other words, Codex was only really useful to an experienced programmer.
But it could help programmers get their day-to-day work done much faster. It could help them find the building blocks they needed or point them to new ideas. Using this technology, GitHub, a popular online service for programmers, now offers Copilot, a tool that suggests your next line of code, much like “autocomplete” tools suggest the next word when you type. texts or emails.
“It’s a way to write code without having to write so much code,” said Jeremy Howard, who founded the Fast.ai artificial intelligence lab and helped create the language technology on which the company works. ‘OpenAI is based. “It’s not always correct, but it’s just close enough.”
Mr. Howard and others believe that Codex could also help novices learn to code. It is particularly effective at generating simple programs from short descriptions in English. And it works the other way around as well, explaining complex code in simple English. Some, including Joel Hellermark, an entrepreneur in Sweden, are already trying to turn the system into an educational tool.
The rest of the AI landscape is similar. Robots are more and more powerful. Chatbots are also designed for online conversation. DeepMind, an AI lab in London, recently built a system that instantly identifies the shape of proteins in the human body, which is a key part of the design of new drugs and vaccines. This task used to take days, if not years, for scientists. But these systems only replace a small part of what human experts can do.
In the few areas where new machinery can instantly replace workers, these are usually jobs that the market is slow to fill. Robots, for example, are increasingly useful in shipping centers, which are growing and struggling to find the workers needed to keep pace.
With his start-up, Gado Images, Mr. Smith set out to create a system capable of automatically sorting newspaper and library photo archives, bringing up forgotten images, automatically writing captions and tags, and share the photos with other publications and businesses. But technology could only handle part of the job.
He could sift through a large archive of photos faster than humans, identify the types of images that might be useful, and tackle captions. But finding the best and most important photos and labeling them correctly still required a seasoned archivist.
“We thought these tools would remove the need to be human altogether, but what we learned after many years was that it wasn’t really possible – you still needed a skilled human to review the result, ”Smith said. “The technology is on the wrong track. And that can be biased. You always need a person to take a look at what they’ve done and decide what’s good and what’s not.
Codex expands what a machine can do, but it’s another indication that the technology works best with humans in charge.
“AI is not going as expected,” said Greg Brockman, CTO of OpenAI. “I felt like he was going to do this job and this job, and everyone was trying to figure out which one would go first. Instead, it doesn’t replace any jobs. But it takes all of the drudgery away from them all at once. “