“I write code, but it feels like a black box.” Meta engineer reveals why developers are losing control over AI

The engineer’s LinkedIn posts have garnered significant traction among IT professionals. In response to the discussion, Hamiyeh clarified that this is not a problem specifically with Meta, but a broader shift in software engineering. In his opinion, developers will have to find a new balance between speed of product creation, understanding of architecture, and validation of AI results.

The code works, but what exactly is happening inside?

“I’m an engineer at Meta, and I’ve never felt so uncomfortable doing my job,” Hamiye wrote.

He says he has been writing code lately, but the process is increasingly like working with a “black box.” AI generates code so quickly that it becomes increasingly difficult to manually check every line. If a team tries to analyze everything the model generates, development can take too long and lose the main advantage of AI: speed.

At the same time, another problem arises: an engineer may understand the general architecture of the system and its purpose, but not know all the details of the implementation.

"You can end up with systems that you built yourself, but you don’t fully understand how all the code works. You understand the big picture, but not every detail," he explained.

For a developer, this goes against the grain of the profession. Previously, understanding what you were building was an integral part of the job. Now, engineers have to determine how deeply they need to understand the implementation before agreeing to release the product.

Hamieh posed this question to other experts: Do they feel the same way?

"This is not a Meta problem — all engineering is changing"

The post sparked a lively discussion. In the comments, other engineers and technical experts debated whether the code generated by AI can be trusted and how to test it without resorting to slow manual development.

After numerous responses, Hamiye wrote a separate clarification. According to him, he was not trying to describe a specific problem with Meta: he also experiences similar feelings when working on his own projects outside the company.

"This is not a Meta problem. This is an engineering problem," he stressed.

In his opinion, the industry is experiencing large-scale changes: more and more code is being written by AI, and developers need to adapt to a new work model.

However, Hamieh does not believe that understanding the system has become unnecessary. On the contrary, he suggested several ways to maintain control over what artificial intelligence creates.

Three ways to stay in control of AI-generated code

1. Pay more attention to design.

According to Hamie, the work begins even before code generation — with a query to the AI ​​and the design of the future system. The developer must understand what exactly the model should build, what components will be in the system, and how they will interact.

2. Test more.

Hamieh emphasizes that testing becomes even more important in a world where code is generated automatically. Tests should help determine whether a system is working properly, even if an engineer hasn’t manually analyzed every line of its implementation.

3. Ask AI to explain and visualize the system.

The model can be used not only to write code, but also to explain how it works. For example, ask the agent to create diagrams, show relationships between components, and answer questions about the implementation.

In the comments, Hamiye also said that he often discusses the principles of the system with the agent and asks him to draw a diagram. He believes this helps to better understand the result.

Do developers need to understand every line?

Not everyone on the panel agreed that the inability to read every line of code was a major problem. In particular, Atlassian lead engineer Amit Gaur noted that the requirement to understand the product does not mean that you have to manually review every line.

Instead, he says, engineers can use flowcharts, dependency graphs, and other ways to visualize changes. With robust review processes in place, some details can be left out of deep analysis.

It’s more important to understand the architecture, choose the right abstractions, and know what the team is releasing to production. But syntax checking, Gaur believes, is gradually losing its former importance.

Hamieh agreed that understanding the system through diagrams and conversations with an agent helps, but stressed that even with this approach, discomfort does not go away. Another important level is to verify the results and have evidence that the system actually works.

Risk is not only about mistakes, but also about losing skills

Other commentators have drawn attention to the long-term implications of the move to AI-powered code generation.

One panelist, who works in technology consulting and data transformation and AI, called the situation an architectural crisis. He believes that if a system generates thousands of lines per second, humans will not be able to scale the verification of each one without losing speed. But superficial code approval also creates risks.

The problem can be exacerbated when the system crashes under heavy load. If the team does not understand the implementation details, finding the cause of the error can become much more difficult.

Another risk is the gradual loss of programming and debugging skills. If new generations of engineers start out working primarily with code generated by models, they may have less practical experience needed to deeply analyze and test such systems.

This is a participant’s position, not a proven consequence of AI. However, it reflects a broader question facing the industry: how to use automation without losing the competencies needed to control its outcomes.

"Tokens are worth more than engineers": EPAM Ukraine explained why AI will not replace developers

On the topic

"Tokens are worth more than engineers": EPAM Ukraine explained why AI will not replace developers

Read the country’s main IT news in our Telegram

On the topic

Read the country’s main IT news in our Telegram