How Long Will It Take for AI to Replace Programmers?

How Long Will It Take for AI to Replace Programmers?

As things stand, AI has made programmers more efficient, but it has not increased their business output. They are simply completing the same tasks faster—and, for many programmers, doing them better.

But a company with 100 employees still has only 100 employees’ worth of business work to do; a company with 1,000 employees still has only 1,000 employees’ worth.

The result is that employers spend additional money on tokens, only to lighten programmers’ workloads. There have certainly been improvements and optimizations, but their impact has not been as significant as expected.

Employers, however, are not Pang Dong Lai—a Chinese retailer famous for its unusually generous treatment of employees. Why would they willingly do that? As a result, they have split into two camps:

1. Continue embracing AI, rank employees by token usage, and push a team of 100 people to produce the business output of 1,000.
2. Begin laying people off—gradually eliminating outsourced roles, removing agreeable but unproductive employees, and dismantling middle-layer departments. The goal is to empower core employees with AI so that even after cutting a 1,000-person workforce down to 100, the remaining team can still handle the original workload.

Both approaches may look like “using AI to achieve a tenfold increase in efficiency,” but the difference is substantial. The former creates a thriving organization; the latter leaves everyone anxious and insecure.

In Silicon Valley, newly founded startups generally take the first path, while some well-known large companies are clearly taking the second.

Meta, for example, is burning through tokens at a staggering rate while aggressively cutting jobs. Everyone is watching to see how the experiment turns out.

How Long Will It Take AI to Replace Programmers?

Under the current paradigm, at least, it will not happen. Mass layoffs and outright replacement are two very different things.

1. Today’s large language models still rely on a foundation of human-built code. Without actively maintained, well-developed open-source infrastructure, AI is just as helpless as humans. At the same time, AI-assisted programming is continuously damaging that open-source foundation. It is hard to say whether AI will take over open source first, or whether AI-generated code will break the foundation before that can happen.

2. AI is highly capable when dealing with conventional business tasks that have clear objectives. But the more obscure and specialized the field, the worse it performs. AI still cannot write code that no one has ever written before.

3. Understanding the business and collaborating with surrounding teams are programmers’ greatest competitive advantages. AI still lacks enough “initiative” to take ownership of a codebase—and if it cannot take ownership of the code, it certainly cannot take ownership of the business.

4. There is not enough context. A one-million-token context window may be enormous for other kinds of work, but in programming, it may not even hold an entire business codebase—let alone its vast dependency tree, especially when those dependencies have undergone major recent changes.

5. Hallucinations remain a serious problem for large models. Without expert supervision, AI often cannot find its way out of the debugging maze. Even when it does, the tokens and time it consumes may mean that it performs no better than a programmer—and may even cost more.

Ever since programming became a profession, humanity has needed more code than the world has been able to supply qualified programmers to write. This contradiction has persisted throughout the history of the profession. Many specialized fields remain insufficiently automated precisely because there are not enough programmers or enough code. Many things are either never done, done poorly, or prohibitively expensive.

Today’s AI tools have both an elitist and a democratizing side. They dramatically raise the ceiling: the stronger a programmer already was, the more they stand to gain. But they also raise the floor, enabling more people to become competent programmers—and to work far more efficiently than before.

Of course, all of this refers to AI programming models at their current level of capability. If they undergo another qualitative leap in the near future, then all bets are off.