As soon as Tom Smith got his hands on Codex – a new artificial intelligence technology that writes its own computer programs – he gave him an interview.
He asked if it could handle the “coding challenges” that programmers often face when they apply for high profile jobs at Silicon Valley companies like Google and Facebook. Could it write a program to replace all spaces in a sentence with hyphens? Better yet, could it write one that identifies invalid zip codes?
It did both immediately before doing several other tasks. “These are problems that would be difficult for many people to solve, including myself, and the answer would be typed in two seconds,” said Mr. Smith, an experienced programmer who runs an AI start-up called Gado Images. “It was scary to watch.”
Codex appeared to be a technology that would soon replace human workers. As Mr. Smith continued testing the system, he found that his skills went well beyond answering pre-made interview questions. It could even translate from one programming language to another.
However, after several weeks of working with this new technology, Mr. Smith believes it will not pose a threat to professional programmers. In fact, like many other experts, he sees it as a tool that will ultimately increase human productivity. It can even help a whole new generation of people learn the art of computers by showing them how to write simple pieces of code, almost like a personal tutor.
“This is a tool that can make a programmer’s life a lot easier,” said Mr. Smith.
About four years ago, researchers in laboratories like OpenAI started developing neural networks that analyzed huge amounts of prose, including thousands of digital books, Wikipedia articles, and all sorts of other texts published on the internet.
By discovering patterns in all of this text, the networks learned to predict the next word in a sequence. If someone typed a few words into this “universal language models“, You could complete the thought with full paragraphs. That way, a system – an OpenAI creation called GPT-3 – could write its own Twitter posts, speeches, poems, and news articles.
Much to the surprise of even the researchers who built the system, they could even write their own computer programs, although they were short and simple. Apparently it had learned from an innumerable number of programs published on the Internet. So OpenAI went one step further and trained a new system – Codex – with an enormous range of prose and code.
The result is a system that understands both prose and code – to a certain extent. You can ask for snow on a black background in plain English and you will be given a code that creates a virtual snow storm. If you ask for a blue bouncy ball, you will get this one too.
“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 it often makes mistakes, and while its skills are impressive, it cannot argue like a human. It can recognize or mimic what it has seen in the past, but it is not nimble enough to think on its own.
Sometimes the programs generated by Codex do not run. Or they contain security holes. Or they don’t come close to what you want them to be. OpenAI estimates that Codex produces the correct code 37 percent of the time.
When Mr. Smith deployed the system as part of a “beta” testing program this summer, the code generated was impressive. But sometimes it only worked when he made a tiny change, such as adapting a command to his particular software setup or adding a digital code required to access the Internet service he was querying.
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 basic building blocks they need or point them to new ideas. With the technology, GitHub, a popular online service for programmers, now offers Co-pilot, a tool that suggests your next line of code, much like “autocomplete” tools suggest the next word when you type in text or email.
“It’s a way of getting code written without having to write so much code,” says Jeremy Howard, who founded the Fast.ai artificial intelligence laboratory and helped develop the language technology on which OpenAI’s work is based . “It’s not always right, but it’s just close enough.”
Mr. Howard and others believe Codex could help beginners learn to code too. It is particularly good at generating simple programs from short English descriptions. And it works the other way too by explaining complex code in plain English. Some, including Joel Hellermark, an entrepreneur in Sweden, are already trying to turn the system into a teaching tool.
The rest of the AI landscape looks similar. Robots are more and more powerful. Chatbots too designed for online conversations. DeepMind, an AI laboratory in London, recently developed a system that instantly identifies the form of proteins in the human body, which is a key element in the development of new drugs and vaccines. Scientists used to take days or even years to do this. But these systems only replace a small part of what human experts can do.
In the few areas in which new machines can replace employees immediately, they are typically in jobs that the market is slow to fill. Robots, for example, are becoming increasingly useful in shipping centers that are expanding and struggling to find the workforce to keep up.
With his start-up Gado Images, Mr. Smith wanted to build a system that would automatically search the photo archives of newspapers and libraries, turn up forgotten pictures, automatically write captions and tags and share the photos with other publications and companies. But technology could only do part of the job.
It could search a huge archive of photos faster than people, identifying the types of images that might be useful, and trying out captions. But finding the best, most important photos and properly tagging them still took a skilled archivist.
“We thought these tools would completely eliminate the need for humans, but after many years we learned that it wasn’t really possible – you still needed a skilled human to review the output,” said Mr Smith. “Technology makes mistakes. And it can be biased. You still need a person to check what they did and decide what is good and what is not. “
Codex expands the capabilities of a machine, but it’s another clue that the technology works best with the human being at the control.
“AI isn’t playing out like anyone would have expected,” said Greg Brockman, OpenAI’s chief technology officer. “It felt like it was doing this job and this job, and everyone was trying to figure out which one would go first. Instead, it does not replace jobs. But it takes the drudgery out of all of them at the same time. “