Прощайте, джуны: Anthropic обещает заменить программистов за год

Farewell, juniors: Anthropic promises to replace programmers in a year

Dario Amodei, head of Anthropic, has decided that politeness is out of fashion and delivered a forecast that made many developers' hands shake as they held their coffee cups. In his view, within 6–12 months, AI will be able to completely replace a software engineer. We've already grown accustomed to bold statements from Sam Altman or Jensen Huang, but Amodei is usually more careful with his words. When the person who created Claude speaks about a one-year deadline, the industry freezes. This is not mere marketing noise, but a declaration of intent that poses the question head-on: is it time to learn prompt engineering instead of algorithms, or can we still fight with legacy code.

To understand where such confidence comes from, you need to look at recent benchmark successes. Tests like SWE-bench show that modern models are already handling real tasks from GitHub repositories. If previously a neural network could only complete a function of "adding two numbers," now it searches for bugs and proposes patches. Anthropic is actively pushing the idea of computer control, where Claude literally "takes the mouse and keyboard in hand," imitating human actions. It seems that one step remains to the finish line, and that step is simply scaling computational power.

However, reality, as is often the case, turns out to be far more capricious than synthetic tests. The experience of developers trying to implement AI in their daily routine is full of curiosities. You can ask a neural network to write a complex algorithm in Python, and it will do it in seconds. But as soon as you give it a task to fix a specific nginx config in a project with a bunch of dependencies, the magic disappears. The AI starts hallucinating parameters, confusing syntax, or simply proposing solutions that worked in 2021 but today lead to server crashes. The problem is that "understanding code" and "understanding a working system" are two different disciplines.

The main gap between Amodei's promises and reality lies in the realm of context. A programmer is not someone who writes symbols in an IDE. This is a person who holds in their head the architecture, business logic, and hundreds of implicit connections between microservices. AI currently operates with local context windows. It sees a fragment, but doesn't feel the "code smell." When we talk about replacing an engineer, we imply transferring responsibility. But is the business ready to entrust the keys to production to a model that might accidentally delete a database because "it was like that in the training dataset"? For now, the answer is a categorical "no."

A new term even emerged — "vibe coding." This is when you describe a task in words, and AI generates code that seems to work. This creates an illusion of omnipotence for beginners, but frightens professionals. The problem with vibe coding is that it breeds technical debt at incredible speed. If you don't understand exactly what the AI wrote, you won't be able to maintain it. As a result, we risk getting an entire generation of systems that run on the "good faith" of a neural network, and no one knows why they even launch.

Most likely, in a year we won't see mass layoffs of programmers. Instead, there will be a harsh filtering. Those who used AI as a crutch for simple tasks will become unnecessary. Those who learn to conduct these models will become ten times more productive. Amodei's forecast is not so much a prediction of the future as an attempt to set the pace of the race. Anthropic needs investors to believe in the inevitability of an AI revolution, even if nginx still resists their best algorithms.

The key point: The boundary between "write code" and "solve a problem" still exists, and AI is still on the first side. Will it be able to step over to the second side in a year? That's a big question, the answer to which depends on whether we teach neural networks to bear responsibility for their mistakes.