THE ECONOMIST: The developing divide as superstar coders are raking it in, but not all of them

Lucas Beyer is not a celebrity. But in Silicon Valley’s rarefied world of machine-learning talent, he is seen as one. A former researcher at OpenAI, Mr Beyer announced last month that he was leaving the artificial-intelligence (AI) lab behind ChatGPT to join Meta, a social-media giant with big AI ambitions of its own.

With rumours swirling that Mark Zuckerberg, Meta’s boss, was offering packages worth $US100m ($151m) to poach AI whizzes, Mr Beyer clarified that he had not secured a nine-figure deal. That he needed to say so at all reflects the extent of the frenzy.

Yet the race for a handful of superstar software engineers masks a slump for everyone else.

As ChatGPT-like generative AI changes how code is written, companies are rethinking how many programmers they need. In America job postings for software developers have dropped by more than two-thirds since the beginning of 2022, according to data from Indeed, a recruitment site.

In January Marc Benioff, the boss of Salesforce, a maker of business software, said his firm will not add any more software engineers this year, owing to productivity gains from AI tools.

In May Microsoft, a tech colossus, cut around 6000 jobs, many of them in engineering.

The upshot has been a divergence in pay. Gergely Orosz of The Pragmatic Engineer, a newsletter, and levels. FYI, a website that tracks employee compensation, analysed salaries for software developers in America with at least five years of experience who began their roles between the start of 2022 and the start of 2025.

At the top of the pay scale were elite AI labs such as OpenAI, as well as hedge funds such as Jane Street that are also betting heavily on machine learning.

In this tier, median pay exceeded $US400,000 a year. Below that were tech giants including Alphabet, Microsoft and, until recently, Meta, where median pay was closer to $US300,000. Experienced developers at most other companies earn much less.

Their median was around $US180,000.

The top tier of coders is tiny. One recruiter estimates that only a few hundred people globally have the right mix of mathematical fluency, research chops and engineering experience to develop cutting-edge AI models.

They tend to have cut their teeth at a handful of AI outfits, including OpenAI, Anthropic and Alphabet’s DeepMind lab. Some, like Mr Beyer, have worked at more than one of them. Most live in a few tech hubs, especially San Francisco and London. Hiring is clubby.

Mr Orosz says that the top firms mostly recruit from each other and rarely consider outsiders. Top AI engineers also do more than code. They must be able to explain their work to product teams and help shape business strategy.

Life is very different for the mere mortals. During the pandemic, tech companies swelled their engineering ranks to meet booming demand for digital services. As interest rates rose, they retrenched just as fast. Lay-offs peaked in early 2023, and hiring has yet to recover.

Partly that reflects caution. But the change is also structural. Generative AI is automating the drudgery of programming. Microsoft claims Copilot, its AI assistant, boosts developer productivity by 10 to 30 per cent, particularly on repetitive tasks.

Many developers now use AI tools to autocomplete functions or suggest bug fixes. Others let AI agents write entire chunks of code, then polish the output.

As a result, businesses may need fewer engineers to build and maintain products. Many are pausing hiring altogether while they take stock.

“Most companies want to see what they can do with their current workforce,” says Zuhayeer Musa, co-founder of levels.fyi. For firms at the AI frontier, he adds, engineers are viewed as an investment, worth tens of millions of dollars if they can help make more efficient use of computing infrastructure that costs billions. For other companies, programmers are often seen simply as another cost to be cut.

Dario Amodei, the boss of Anthropic, has predicted that AI could soon be writing all code. That is probably optimistic.

But the direction of travel is clear: the future of programming will belong less to those who write it, and more to those who build the machines that do.