Categories
Reactions

The Dark side of AI in Software Engineering 2026.09.30

I was watching a YouTube video by Kevin Powell titled “We’re all fed up.”

Related Reddit Discussion is here.

At 4:25, he asks if we know why that could be, noting that while code output is no longer a blocker, software engineers are still slow to deliver value.

Here is my answer

There is no need for a study. I have 14 years of professional full-stack experience, and I have not written code for over a year now. I agree with this tweet 100%. The productivity loss comes from how Claude Code writes and talks. Everything it puts out is like a legal document that consumes souls to understand. And it produces a lot of those texts that need to be read and understood. Every time you ask a change in that output it regenerates the whole document which is wild, because you cannot trust it, so you read it again and again. And it burns you out. The other side is not reading it, but then when engineers get tickets and want to know what to build, there is miscommunication, missing information, a lot of text that takes hours to read but gives no usable information, and overuse of jargon. Also, when this approach is used, and you want to follow up on what this software should do, what the business wants to achieve, and what the decided architecture is, no one can answer anything, and everyone points to someone else. I saw both sides of these examples. That is why, I’d say, it is slower than before. And maintenance will be a bottleneck very soon and will cause a lot of companies to go down.

Right now, I am in a situation where finding information takes days, but building it with Claude Code takes a few hours. Testing it again will probably take days again. It will be an AWS CDK infrastructure-as-code project change, but no one knows the features that need to be connected at the infrastructure level, nor how things connect. It is a multi-team project, so the information is spread across multiple projects, and no single source of truth exists.

At 6:39 he says it takes the soul of developers and there is no feeling of victory.

For example, when assembly lines came into existence, they killed handcrafting many things.

Answer

Now, yes. But as I already talked about (in the Career growth in the age of AI post), the challenge is no longer writing code. But to actually know what to write, architect it well, and communicate the entire process properly. Also, feature testing and product validation in the market. Of course, maintenance as well, both in production (logs, bugs) and at the feature (usability) level.

Shipping code fast will create more work related to those other areas. We need to build new skills to use AI better, and learn the necessary soft skills to get back the driving seat from what insane KamiKaze is happening right now.

By Botond Bertalan

I love programming and architecting code that solves real business problems and gives value for the end-user.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.