RMB 20,000 for Outsourcing vs RMB 2,000 in AI Tokens: Programmers Will Not Disappear Overnight, but They Are Being Repriced First

A recent V2EX post caught my attention because of one number pair: RMB 20,000 for outsourced development versus RMB 2,000 worth of AI tokens.

The original poster said he is also a programmer and is building a business. In the past, when he had more projects than he could handle, he would outsource part of the work. Now, he feels that spending the same effort communicating with AI can be more effective. In his view, a few thousand yuan worth of tokens may do the work of several experienced outsourced developers.

I do not want to turn this into a simple claim that AI has already replaced programmers.

That would be too rough. The post itself lacks key details: what kind of project was outsourced, which model or tool was used, whether the project reached production, and who would maintain it later. The poster also admitted that AI is fast for early demos, but a project can become messy if one relies entirely on AI in the later stages.

That matches my own technical management experience. The hard part of software work is rarely just writing a few pages or connecting a few APIs. The real trouble comes from changing requirements, legacy code, permission boundaries, data consistency, rollback plans, and accountability after launch.

There is also counter-evidence. In a 2025 METR experiment, experienced open-source developers using AI tools on real issues in mature projects became 19% slower in that setting. But METR later updated its view in 2026, noting that developer productivity effects may have changed and become harder to measure as developers increasingly rely on AI and multi-agent workflows.

That is the real problem: this is not a binary situation.

AI is not useless just because it fails in some mature codebase tasks. Programmers are not finished just because AI can write code. This is a dynamic process of technical progress, market competition, and cost recalculation.

Every time models improve, employers recalculate. Every time token prices fall, outsourcing prices and hiring expectations are pressured again. Whoever uses AI effectively first forces others to follow.

Anthropic's labor market research found high exposure for computer programmers, while also showing that actual observed use still lags far behind theoretical capability. In plain language: AI has not fully replaced programmers, but the path has been opened.

Capital and competition do not wait for perfect replacement. If AI can build demos today, it will be used for demos. If it can fix bugs tomorrow, it will be used to fix bugs. If it can read legacy code later, it will be used there too. At some point, the programmer's role shifts from writing every line of code to directing and reviewing AI-generated work.

The cold part is that an individual company's optimal choice may not be society's optimal outcome. A boss wants lower cost and faster delivery. A developer wants to use AI to become more productive. A company wants to reduce inefficient roles and protect margins. But if everyone optimizes this way, the broader result may be fewer entry-level jobs, lower household income, and weaker demand.

This is why I expect state power, regulation, and industry rules to eventually step in. When a technology reshapes employment, income distribution, education paths, and social stability, it is no longer merely an internal productivity tool for companies.

At the same time, ordinary people should not be scared to death. Technology revolutions do not erase all old industries in one day. Even after washing machines became common, traditional large-scale washing services still exist in places such as Mumbai's Dhobi Ghat. AI will also leave gaps: local services, trust-based work, domain expertise, responsibility, taste, and on-site problem solving.

But there is a big difference between saying someone will still have a job and saying I personally will still have a job.

That is the part programmers need to face honestly.

The key question is not simply whether AI will replace programmers. The better questions are: is my work mostly a repeatable intermediate process that AI can imitate? Do I own judgment around requirements, system boundaries, business responsibility, and quality control? If my company asks one person with AI to do the work of two, am I the person who stays or the cost being optimized away?

The change will not be black and white. It will be gradual, contested, and uneven. But the direction is no longer something we can block with denial. Programmers will not disappear overnight. Many will be repriced first.

老花 / Easton Hua