Business Insider, 4/26/23
ChatGPT has come for software developers
When ChatGPT was released to the world in November, most of us marveled at its ability to write rap lyrics and cover letters and high-school English essays. But Adam Hughes, a software developer, was intrigued by artificial intelligence’s much-ballyhooed aptitude for writing code. So he signed up for an account and asked ChatGPT to program a modified tic-tac-toe game, giving the game some weird rules so the bot couldn’t just copy code that another human had already written. Then he quizzed it with the kind of coding questions he asks candidates in job interviews.
Whatever he threw at it, Hughes found that ChatGPT came back with something he wasn’t prepared for: very good code. It didn’t take him long to wonder what this meant for a career he loved — one that had thus far provided him with not only a good living and job security, but a sense of who he is. “I never thought I would be replaced in my job, ever, until ChatGPT,” he says. “I had an existential crisis right then and there. A lot of the knowledge that I thought was special to me, that I had put seven years into, just became obsolete.”
Coding, as an occupation, has long been considered a haven from the relentless advance of technology. Even as new gizmos replaced other jobs, the people who wrote the instructions for the machines felt untouchable. Universities rushed to expand their computer-science programs. Policymakers scrambling to futureproof the workforce stuck to one unwavering message: Learn to code! But in recent weeks, behind closed doors, I’ve heard many coders confess to a growing anxiety over the sudden advent of generative AI. Those who have been doing the automating fear they will soon be automated themselves. And if programmers aren’t safe, who is?
Much has been written about how AI is coming for white-collar jobs. Researchers at OpenAI, which created ChatGPT, recently examined the degree to which large language models could perform the 19,000 tasks that make up the 1,000 occupations across the US economy. Their conclusion: 19% of workers hold jobs in which at least half their tasks could be completed by AI. The researchers also noted two patterns among the most vulnerable jobs: They require more education and come with big salaries. “We didn’t think that would be the case,” says Ethan Mollick, a professor of management at Wharton who studies innovation. “AI was always supposed to automate dangerous, dirty tasks — not the things we want to do.”
But one white-collar skill set, the study found, is especially at risk for being automated: computer programming. The reason? Large language models like the one powering ChatGPT have been trained on huge repositories of code. Researchers at Microsoft and its subsidiary GitHub recently divided software developers into two groups — one with access to an AI coding assistant, and another without. Those assisted by AI were able to complete tasks 56% faster than the unassisted ones. “That’s a big number,” Mollick says. By comparison, the introduction of the steam engine in the mid-1800s boosted productivity at large factories by only 15%.
Tech companies have rushed to embrace generative AI, recognizing its ability to turbocharge programming. Amazon has built its own AI coding assistant, CodeWhisperer, and is encouraging its engineers to use it. Google is also asking its developers to try out new coding features in Bard, its ChatGPT competitor. Given the tech industry’s rush to deploy AI, it’s not hard to envision a near future in which we’ll need half as many engineers as we have today — or, down the line, one-tenth or one-hundredth (Emad Mostaque, the CEO of Stability AI, has gone as far as predicting “there’s no programmers in five years.”). For better or worse, the rise of AI effectively marks the end of coding as we know it.
Now, before we dive into this doomsday scenario, let’s pause for a moment and consider the case for optimism. Perhaps, as the industry’s sunnier forecasts are predicting, there’s enough of a demand for coding to employ both humans and AI. Sure, the arrival of the tractor threw a lot of farmers out of work. But coding isn’t like farming. “There’s only so much food that 7 billion people can eat,” says Zachary Tatlock, a professor of computer science at the University of Washington. “But it’s unclear if there’s any cap on the amount of software that humanity wants or needs. One way to think about it is that for the past 50 years, we have been massively underproducing. We haven’t been meeting software demand.” AI, in other words, may help humans write code faster, but we’ll still want all the humans around because we need as much software as they can build, as fast as they can build it. In the rosiest outlook, all the productivity gains from AI will turbocharge the demand for software, making the coders of the future even more sought after than they are today.
Another argument from the optimists: Even as AI takes over the bulk of coding, human coders will find new ways to make themselves useful by focusing on what AI can’t do. Consider what happened to bank tellers after the widespread adoption of ATMs. You’d think ATMs would have destroyed the profession, but surprisingly, the number of bank tellers actually grew between 1980 and 2010. Why? Because bank tellers, one analysis found, became less like checkout clerks and more like salespeople, building relationships with customers and selling them on additional services like credit cards and loans. Similarly, Tatlock envisions a future for software engineers that involves less writing of code and more verifying of all the cheap and potentially dangerous code the machines will be generating. “You probably don’t need to formally verify a widget on your website,” Tatlock says, “but you probably do want to formally verify code that goes into your driving assistant in your car or manages your insulin pump.” If today’s programmers are writers, the thinking goes, their future counterparts will be editors and fact-checkers.
So maybe, long term, human coders will survive in some new, as-yet-to-be-determined role. But even in the best-case scenario, the optimists concede, the transition will be painful. “It is going to be the case that some people’s lives are upended by this,” Tatlock says. “This happens with every technological change.” Some coders will inevitably be displaced, unable to adapt to the new way of doing things. And those who make the transition to the AI-driven future will find themselves performing tasks that are radically different from the ones they do today.
There’s only so much food that 7 billion people can eat. But it’s unclear if there’s any cap on the amount of software that humanity wants or needs.Zachary Tatlock, University of Washington
The first question is: In this evolutionary battle for survival, who is best positioned to adapt, and who’s going to get left behind? Intuitively, you would think seasoned veterans — those who already spend less time coding and more time on abstract, higher-order, strategic thinking — would be less vulnerable to AI than someone straight out of college tasked with writing piecemeal code. But in the GitHub study, it was actually the less experienced engineers who benefited more from using AI. The new technology essentially leveled the playing field between the newbies and the veterans. In a world where experience matters less, senior engineers may be the ones who lose out, because they won’t be able to justify their astronomical salaries.
Then there’s the issue of job quality. The optimists assume that AI will enable us to outsource a lot of the boring, repetitive stuff to the bots, leaving us to concentrate on more intellectually stimulating work. But what if the opposite ends up happening, and AI takes on all the fun stuff? No disrespect to my colleagues in the research department, who do vital work, but I’m a writer because I love writing; I don’t want my job to morph into one of fact-checking the hallucinogenic and error-prone tendencies of ChatGPT. What feels unnerving about generative AI is its capacity to perform the kind of highly skilled tasks that people enjoy most. “I really love programming,” says Hughes, the software developer. “I feel like I’m one of the few people who can say for sure that I’m in the career I want to be in. That’s why it’s scary to see it at risk.”
But the greatest glitch in the “it’ll be OK” scenario is something the optimists themselves admit: It’s predicated on the assumption that generative AI will keep serving as a complement to human labor, not as an outright replacement. When ATMs came along, bank tellers were able to adapt because there were still things they could do better than the machines. But go back a few decades, and you’ll find a technology that obliterated what was one of the most common jobs for young women: the mechanical switching of telephones. Placing your own calls on a rotary-dial phone was way faster and easier than going through a human switchboard operator. Many of the displaced operators dropped out of the workforce altogether — and if they kept working, they ended up in lower-paying occupations. Their fate raises the question: At what point does AI get so good at coding that there’s nothing left for a human programmer to do?
The fact we need to ask that question underscores one of the most glaring problems with AI research: Far too much of it is focused on replacing human labor rather than empowering it. Why are we deploying our best and brightest minds to get machines to do something humans can already do, instead of developing technology to help them do something entirely new? “It’s a sad use of innovation,” says Katya Klinova, the head of AI, labor, and the economy at the nonprofit Partnership on AI. There are plenty of dire problems in the world that need solving, she points out, like the urgent need for more sources of clean energy. The question we should be asking about AI isn’t how well it can perform existing human tasks, and how much money that automation will save businesses — it’s whether the technology is doing what we, as a society, would like it to do.
In the meantime, on an individual level, the best thing coders can do is to study the new technology and to focus on getting better at what AI can’t do. “I really think everybody needs to be doing their work with ChatGPT as much as they can, so they can learn what it does and what it doesn’t,” Mollick says. “The key is thinking about how you work with the system. It’s a centaur model: How do I get more work out of being half person, half horse? The best advice I have is to consider the bundle of tasks that you’re facing and ask: How do I get good at the tasks that are less likely to be replaced by a machine?”
Mollick adds that he’s watched people try ChatGPT for a minute, find themselves underwhelmed by its abilities, and then move on, comforted by their superiority over AI. But he thinks that’s dangerously shortsighted, given how quickly the technology is improving. When ChatGPT, powered by the 3.5 model of GPT, took the bar exam, for instance, it scored in the 10th percentile. But less than a year later, when GPT 4 took the test, it scored in the 90th percentile. “Assuming that this is as good as it gets strikes me as a risky assumption,” Mollick says.
“The best advice I have is to consider the bundle of tasks that you’re facing and ask: How do I get good at the tasks that are less likely to be replaced by a machine?”
Hughes has seen the same head-in-the-sand reaction from his fellow coders. After ChatGPT aced his tic-tac-toe challenge, he was scared to look his phone, for fear of seeing yet another headline about the tool’s human-like capabilities. Then, as an act of catharsis, he wrote a long post on his Medium blog — a step-by-step, worst-case scenario of how he thought AI could replace programmers over the next decade. The response was telling: Developers flooded the comments section with impassioned critiques, some of them so aggressive and toxic that Hughes felt forced to delete them. In post after post, they listed all the ways they thought they were still better coders than ChatGPT. “You are a really bad software developer if you don’t understand the number of AI limitations,” one seethed. AI, they were confident, won’t replace what they bring to the job anytime soon.
Reading the comments, I found myself thinking the critics were missing the point. AI is still in its infancy. Which means, much as with a newborn human, we need to start thinking about how it will affect our lives and our livelihoods now, before its needs outstrip our ability to keep up. For the moment, we still have time to shape the future we actually want. Sooner or later, there may come a day when we no longer do.