Business

OpenAI Is Making Coding As Easy As Talking to a Smart Speaker

Plus: The early days of programming, an existential investigation, and bipartisanship before our very eyes.
alphabet letters
OpenAI's latest creation, Codex, writes computer code—in some instances very well.Photograph: Carmen Martínez Torrón/Getty Images

If you buy something using links in our stories, we may earn a commission. Learn more.

Hi, everyone. Hope you had a great infrastructure week! Maybe by the time we get those new train routes—in 2035—it’ll be safe to ride them again.

The Plain View

A few weeks ago, I went to an office in the Mission District of San Francisco to get an advance peek at a computer program that promises to disrupt computer programming. I was visiting OpenAI, a company devoted to “artificial general intelligence” (as opposed to humans, who have, at least on good days, non-artificial general intelligence). Their latest creation is called Codex, and it writes computer code, in some instances very well.

I was impressed as OpenAI CTO Greg Brockman and his fellow cofounder Wojciech Zaremba, who is a key creator of Codex, took it through its paces in a live demo. They asked Codex to express some text, grab some images, make a web page, and put the page on the internet. Then, employing the casual language one might use in conversation, they built a simple game by grabbing web-based images of helicopters and getting them to fly across the screen and blast enemies. As the game took shape feature by feature, I realized I was viewing a transformation of the famous “flow state” that good coders feel when they get rolling. Until now that flow involved an intense inner dialog. Now it’s more like a conversation with a robot companion.

The simple commands that Codex translates from language to code break no ground on what a computer can do, but ideally it will wash away the painstaking minutes that even an expert coder would expend to get such tasks done. At one point in the demo, OpenAI CEO Sam Altman, who was kibitzing the session, remarked that in his programming days it would have taken him half an hour to do a task Codex accomplished in seconds. Codex—which knows how to use a dozen computer languages, ranging from Python to JavaScript to HTML—figures out on its own which is suited for a given task and sets about churning out clean code.

OpenAI’s previous blockbuster was GPT-3, a “natural language” system that is able to generate plausible prose expressions from a variety of cues. This isn’t exactly the same thing, but you can call it a sibling—both cases violate the natural order of what people do and what machines do. But, for now at least, the two also have similar limitations. GPT-3 can’t match up with Nabokov right now, but can probably whip up a reasonable jacket copy. In the same vein, government IT workers won’t be replacing the social security infrastructure with a few Codex commands anytime soon. But while GPT-3’s successes come by producing documents that just clear the credibility bar, the bite-size bits that Codex creates match or improve on human code, in a fraction of the time. “This model is much better than I am,” says Brockman. “Because we've trained it on billions of lines of publicly available source code, it’s seen every way that people use all the functions that are out there, and it's very good at mapping that to the context of the domain that you're operating in.”

Still, at this point, Codex can’t pull off even a majority of requests that people give it. The version announced this week executes about 37 percent of tasks. But that’s an improvement from the previous version, which launched as part of a Github product called Copilot; that was successful with only 27 percent of requests. OpenAI’s expectation is that Codex will get much, much better. It’s only a baby!

As I watched the demo, my first thought was, There goes a lot of jobs. Brockman begs to differ. He argues that real programming isn’t the gritty kind of work that Codex does, translating commands to code, but rather the more conceptual work that humans are better at, for now at least. “Programming is really about having a dream,'' he says, “It’s about having this picture of what you want to build, understanding your user, asking yourself, ‘How ambitious should we make this thing, or should we get it done by the deadline?’”

But wait. Zaremba says OpenAI’s tests show that when using Codex, programmers produce software more than two times faster. I’m not a math whiz, but it seems to me that if my company employs 10 programmers, doubling their productivity means I can let go of five, right? Not according to Brockman, who says that by making it possible to write more computer programs, people will create code to make a variety of tasks easier and more productive, and demand for coders will rise.

Codex, says Brockman, will revolutionize not just writing code but learning how to code. Hadi Partovi, the founder of Code.org, thinks that computer education will improve when students can instantly see how Codex tackles a problem and learn from it. Also, the empowerment that newbies get from instantly executing commands will encourage them to learn, as opposed to the olden days when people wrote buggy code that made them feel like losers and turned them off from the pursuit.

Still, the cynical side of me questions the value of teaching people how to create code just at the moment that computers are learning to do it on their own. But Partovi is on the same page as Brockman and insists that the many programmers of the future will be the dreamer types, not grunts cranking out code. “As the rote work of coding becomes easier,” says Partovi, “computer science education can focus on higher-level computational thinking concepts such as designing interfaces, algorithms, and data structures.” He agrees with Brockman that Codex will not put people out of work but increase the demand for computer programmers, because we’re about to unlock the power of coding to transform new domains. (By the way, OpenAI plans to make money by charging companies who use Codex, at a pricing scheme yet to be determined.)

In other words, they are saying that my fictional company with 10 programmers—who can now do the work of 20 people by using Codex—might hire 10 more programmers. That’s 40 times more code being written for my company!

The point of Codex and its successors, then, is to immerse us in a sea of source code, directed toward all sorts of previously non-automated tasks. Brockman envisions something similar to the Star Trek computer that can basically fulfill any request you give it—like Alexa, only it does everything. “And one day,” he adds, “maybe you can build a machine that's fully as capable as a human. We’re building, building, building toward that. And all along the way, we're building useful things.”

Codex, can you fill out my unemployment form?

Time Travel

In my 1984 book Hackers, I write about the very early times of computer programming, beginning with the days when programs were not typed on a keyboard but keyed into punch cards and fed to the computer by a priesthood of authorized tenders. The programmers would wait, sometimes for hours or even days, to learn whether the code worked. In this passage I talk about the IBM 704 at MIT and the students of John McCarthy, a professor who first coined the phrase artificial intelligence:

There were secrets to those IBM machines, painstakingly learned by some of the older people at MIT with access to the 704 and friends among the Priesthood. Amazingly, a few of those programmers, grad students working with McCarthy, had even written a program that utilized one of the rows of tiny lights: the lights would be lit in such an order that it looked like a little ball was being pass from right to left: if an operator hit a switch at just the right time, the motion of the lights could be reversed—Computer Ping-Pong! This obviously was the kind of thing that you’d show off to impress your peers, who would then take a look at the actual program you had written to see how it was done.

To top the program, someone else might try to do the same thing with fewer instructions—a worthy endeavor, since there was so little room in the small “memory” of computers those days that not many instructions could fit into them. John McCarthy had once noticed how his graduate students who loitered around the 704 would work over their computer programs to get the most out of the fewest instructions, and get the program compressed so the fewest cards would need to be fed to the machine. Shaving off an instruction or two was almost an obsession with them. McCarthy compared those students to ski bums. They got the same primal thrill from “maximizing code” as fanatic skiers got from swooshing frantically down a hill. So the practice of taking a computer program and trying to cut off instructions without affecting the outcome came to be known as “program bumming,” and you would often hear people mumbling things like, “Maybe I can bum a few instructions out and get the octal correction card loader down to three cards instead of four.”

Ask Me One Thing

Carl asks, “This may sound odd, but can you swear that Josh Harris is a real person?”

Your question is indeed odd, Carl. You are referring to the cofounder of the long defunct internet company Pseudo who famously lost a fortune. Harris was a pioneer, if you want to call it that, in oversharing on social media well before everyone else did. He also had a habit of dressing up as a clown called Luuvy. He was the subject of a documentary called We Live in Public, though Carl suspects it might be a fictional mockumentary. Carl, I’m not swearing on anything, but I see no evidence that Harris doesn’t exist, despite his company being called Pseudo. He’s got a pretty dense internet trail, and if this were all faked, it would be one of the great hoaxes of all time.

WIRED once ran an excerpt of a book about Harris, and I trust our fact-checkers would sniff out deception of that sort. The author, Andrew Smith, also confirmed to me that Harris is real beyond a doubt and that Jonah Hill is going to play him in a TV series (!). Also, in the course of looking into this, I discovered that I am Facebook friends with Harris. And Wikipedia says that he is alive and currently living in Las Vegas. Of course, those last two factoids mean nothing. They are just dots on the internet. They all could be deleted or rearranged. And, yes, “Josh Harris,” who many journalists interviewed, and whose image we’ve seen on many cameras, could be any human claiming to be Josh Harris. The same goes for the person who responded to the message I just sent to Josh Harris on Facebook! What if Josh Harris’ parents just put that name on his birth certificate while secretly agreeing his name was something else? Would he be Josh Harris or not? Who is anyone, really?

You can submit questions to mail@wired.com. Write ASK LEVY in the subject line.

End Times Chronicle

A bipartisan bill actually gets through the Senate? Maybe the answer to the climate crisis is hell freezing over.

Last but Not Least

Our spiritual advice columnist tells you what to do when AI flirts with you. That’s what happens when we let programmers dream.

It took the internet to expose the dark side of hedgehogs.

The secret database that denies patients pain medicine to stave off their agony.

Phantom ships are making waves in conflict zones.

If you buy something using links in our stories, we may earn a commission. This helps support our journalism. Learn more.

More Great WIRED Stories