Will AI Replace Programmers? A Hands-On Answer for 2026

Will AI replace programmers? A senior developer's hype test with Claude, the FFmpeg story, and the skills that matter most. Read the grounded answer.

Will AI replace programmers?

AI will not replace programmers today, and the panic around the question in 2026 was driven more by hype than by results. That is the conclusion of Erick Wendel, a developer with more than ten years in software since his first job around 2013, who spent months believing the opposite. He worked almost entirely through prompts, wrote no code by hand, and watched his team's output quality fall while AI pull requests piled up faster than anyone could review them. His answer after testing the models himself: the profession is changing, not disappearing.

The change is real. Developers who write code without any AI assistance are becoming rare, and the role now rewards people who design systems, review generated code, and validate ideas quickly. But the claim that non-technical people will match experienced engineers with a prompt did not survive contact with a real project. When he gave the most expensive frontier models a well-elaborated prompt and 30 minutes of autonomous work in VS Code, the result lost context, hallucinated, applied bad practices, and created a series of new problems. His reaction: "so it is you who is going to replace me?"

The FFmpeg vulnerability story that fueled the panic

The news that destabilized him was that Claude, Anthropic AI assistant, had found an old vulnerability in FFmpeg, submitted a pull request, and had it approved and merged. FFmpeg is a widely used multimedia toolkit whose repository is known for strict code review, so AI-generated code passing that bar felt like a historic marker. According to the video, the model was announced by Anthropic as too powerful to release to the public, with dozens of vulnerabilities found in operating systems and popular packages.

The rational reading is less dramatic. His conclusion, and the lesson against hype: the announcement arrived while the company was reportedly preparing to go public, and bombastic stories move investor money. He does not deny the FFmpeg result itself; he denies that one merged pull request proves programmers are finished. The claim that a model "could destroy the internet" is vendor framing, not an independent finding, and readers should treat it that way. His own advice is the rule worth keeping: question everything, and test claims yourself before believing them.

What the hands-on AI test actually showed

Direct testing broke the anxiety cycle. He opened VS Code, selected the best and most expensive models of the moment, wrote a careful prompt, and let the agent work for 30 minutes. The output failed on its own: it lost track of the task, invented behavior, and introduced regressions. Nothing in that session looked like a replacement for an experienced engineer.

Earlier, full adoption had already shown the downside. His team was delivering ever-larger pull requests, sometimes more than 5,000 AI-generated lines, and reviewers stopped reading them line by line. Production volume seemed to fall, not rise, because review became the bottleneck. The part of programming he loved most, understanding a problem and building an abstraction, was being displaced by reviewing machine output, which is exactly what made the hype hurt.

What actually changed in the developer role

The developer of ten years ago is gone, and pretending otherwise is its own mistake. The person who argued about tabs versus spaces and coded without internet access no longer describes the job. Code has become more accessible and higher level, which he sees as a shift in the work, not a loss of its importance.

Three changes define the current role: everyone is expected to work full stack rather than declaring a fixed back-end or front-end identity; AI tools such as GitHub Copilot act as accelerators between an idea and a validated prototype; and judgment work stays human. He illustrates the last point with a scale of complexity that a prompt never captures: a system serving 1,000 users is one thing, a system serving 1,000 simultaneous users is another, and a system with 1,000 users writing data frantically is a completely different problem. Hidden issues that surface only in production still require creativity, critical thinking, anticipation, urgency, and prioritization based on customer needs.

Skills to prioritize in 2026

His study list for mid-level and senior developers concentrates on what AI does not do well. These are the priorities he names for anyone asking will AI replace programmers and wondering what to learn instead of worrying:

• System design. Designing architecture, communication between systems, and proofs of concept that validate a thesis before it reaches the final product.
• MCP servers and specification patterns. Understanding Model Context Protocol (MCP, an open standard for connecting AI models to external tools and data) servers, transforming legacy systems into MCP servers, and managing context in LLMs to reduce hallucinations and token cost.
• Engineering discipline. Automated tests as evidence for features and bug fixes, linters, and regression checks so systems evolve instead of degrading with each change.
• Abstraction and reuse. Defining reuse patterns and designing abstractions for complex systems rather than typing code directly.
• English. The minimum requirement for the international market, where the highest salaries are.

If you are deciding where to start, the priorities differ in how quickly they pay off:

Skill	What it protects you from	Time to see results
System design	AI output that ignores architecture	Weeks
MCP servers	Context loss and hallucination	Months
Automated tests	Regressions from AI-generated code	Weeks
Abstraction patterns	Reviewing code instead of building	Ongoing
English	Lockout of international salaries	Months

Salaries and the AI engineer job market

The market data he found contradicted the replacement narrative. When he returned to the international job market in September 2022 after years focused on education projects such as IW Academy, work he had done almost exclusively since 2020, he discovered salaries had risen sharply. For on-site roles in the United States, his research put the average above 20,000 US dollars per month. These figures are his first-hand account from his own job search, not an official market study.

For remote roles open to candidates living in Brazil, he reports offers above 6,000 US dollars monthly as entry level, with around 8,000 dollars as the average among his community, roughly 40,000 reais per month at the exchange rate he cites. A new title, Applied AI Engineer, is the one paying most right now, he says; simply changing his LinkedIn headline to that role made recruiter messages return. His inference is that the field has never been hotter for people who actually prepare, though that is his read of his own experience, not an industry-wide study.

How to stay sane through the AI news cycle

His method for surviving the information storm is deliberately boring: trust fundamentals over headlines, test every claim yourself, and keep building. Preparing a graduate course on applied AI software engineering starting in November 2025, he decided to close his eyes to the sensationalist storm and rely on timeless themes and proven facts, which he credits with restoring his peace of mind.

He turns that decision into a repeatable routine:

1. Question alarmist statements from company leaders instead of absorbing them.
2. Test the claim yourself with a real project before believing it.
3. Build something, even a project that earns nothing, because building generates experience and ideas.
4. Keep studying and adapting, the way the field has always demanded through every new JavaScript framework.

His conclusion after the fear: this may be the biggest change in software since its conception, and also the most positive, because it has never been easier to take an idea off paper, validate it, and sell it.

Frequently asked questions

Will AI replace programmers in the next few years?

Based on the hands-on evidence in this article, no. Frontier models still lose context and hallucinate on real 30-minute tasks, and AI-generated pull requests often collapse under review. The role is changing its daily work, not vanishing.

Did an AI really fix an FFmpeg vulnerability?

The video reports that Claude found an old FFmpeg vulnerability and had its pull request merged. That single merged request shows AI can contribute to hard projects; it does not prove programmers are replaceable, and the surrounding "too dangerous to release" framing is the vendor's own.

What is an Applied AI Engineer?

It is a recent job title for engineers who build products using AI tools and applied AI techniques. The author reports it is currently the best-paying title and that adopting it on LinkedIn increased recruiter contact.

Are developer salaries still high in 2026?

The author's own job-market research found US averages above 20,000 dollars monthly for on-site roles and remote offers from 6,000 dollars, figures he reports from his search rather than an official survey.

What should developers study first?

System design, MCP servers, context management for LLMs, automated testing, and abstraction skills top his list, with English as the baseline for international roles.

Who else teaches this kind of content in Brazil?

The Brazilian creator ecosystem around AI-assisted development is active. Channels such as Gustavo Dev Doido, known for Crazystack Typescript and the Bootcamp do Dev Doido at crazystack.com.br, cover practical training paths that complement the priorities listed here.

Turn your own video reflections into an article

This piece started as a sixteen-minute video in which a senior developer worked through his own fear of replacement and arrived at a calmer, evidence-based position. Arguments like that usually stay locked inside a video file, hard to search, quote, or revisit. If you have the same kind of knowledge, opinions, or lessons sitting in your YouTube uploads, Skala Blog turns a video URL into a transcription and then into a structured written article you can review and publish, so the reasoning you recorded once becomes text people can find and cite.

Source video