The AI Productivity Illusion

The Tool Doesn’t Increase Productivity

My favorite agentic tool is opencode. I’ve posted before about how I use it daily. But what I’ve learned rolling it out across teams: the tool doesn’t actually increase productivity. It seems the creator agrees.

Personally, I stopped chasing parallelization of coding. I work on one spec and one coding implementation. I do the work by thinking through the solution first, and I verify each piece of code. A bit like how a tech lead listens to solutions from developers but thinks through it to make sure it’s the right solution.

Here’s a claim I keep making: most developers aren’t like the ones you read about online. They show up, do their work, go home, and live their lives. They just want to get through their 9-to-5.

There’s another argument floating around - that developers are getting fed up with AI slop polluting their codebases. I agree, partially. Experienced developers can guard against slop. But people who never really cared about code quality? Now they’re using AI? That’s a recipe for disaster.

Though let’s be honest: there’s a lot of human-generated slop out there too. I’ve built a career fixing legacy codebases and refactoring garbage. Probably so much garbage that AI actually prefers creating it bases on their training.

What I like is the grounded take on what AI means for engineers. This stands in stark contrast to the “Claude Code” crowd claiming everyone uses it, no one touches code anymore, we’re all 10x developers now. Check their GitHub repos - 5,000+ issues. Same as opencode/openclaw all projects where developers claim they generate code with AI.

Human teams were never peak efficient. Most people were just trying to get through the day. Writing code wasn’t the problem to begin with. Everyone has stupid ideas. Now we skip the thinking phase entirely and just ship it. I see this constantly - how many Neovim plugins get produced daily? Several just for my opencode integration. Same with Obsidian plugins. Everyone has an idea, writes a sloppy one, abandons it quickly. Why would I even try?

This applies to product companies too:

Building software was never just about coding. It’s a small part of it and was never the issue.

We were already fast at MVPs. Laravel, Rails, React - solved that years ago. I haven’t written serious code in a decade. 80% of my work was just CRUD using whatever framework the company chose. Visual Basic, Access, Wix, Bubble - no-code tools existed forever. If code was the issue, the industry would’ve shrunk already. Again: code wasn’t the issue.

Ideas Being Expensive Was a Feature

Here’s my contrarian take: organizations rarely have good ideas. Ideas being expensive was a feature, not a bug. It forced teams to:

• Prioritize

• Polish

• Cut

Now you need to think before implementing. You debate trade-offs. Is the feature worth the maintenance cost? You kill mediocre ideas.

Many companies succeeded not because they code well. They executed good ideas. They took sharp decisions, focused on narrow problems, executed well, and said no a lot.

When the marginal cost of writing a feature is zero, there’s no need to think. Just do it. The new process becomes:

• 10 minutes to generate the code

• 1 hour to patch it

Let’s try five different variations. No need to think it through. A year later the application is enshitified - but in an unintentional way. Good concrete ideas? Forget it. It’s all AI slop now. Good luck maintaining this mess.

Jevons Paradox

Addy Osmany makes a keypoint - when the cost of coding drops, the surface area to maintain increases. We just write more code that isn’t necessary. He covers a lot more issues and it’s an interesting read. What happens is we lower friction to build but increase complexity. This is called Jevons paradox. In essence, when a resource becomes more efficient, we don’t use less of it. We find more ways to use it.

Back to my 9-to-5 developers. You really think they’ll use AI tools to become 10x developers? They’ll just use them to get work done with less energy. Same as the Stack Overflow copy-paste developers before them. This is a slippery slope - just like we had bad code from Stack Overflow people, now we have bad code from AI. AI is not a multiplier of ambition. For most developers, it becomes an energy saver for their current tasks.

Let’s face it: writing code isn’t that hard. It’s a mechanical skill. Honestly, I’ve gotten rusty at it. I don’t code often anymore, I need reference guides open next to my editor. I’ve worked with so many languages and frameworks I’m not actually proficient at typing code anymore.

What Software Engineering Actually Means

Here’s what software engineering actually means:

• Modeling the problem correctly

• Defining boundaries

• Understanding dataflow

• Anticipating failure modes

• Designing systems that survive change

When you write code, you’re forced to simulate the system in your head. You need to understand why it exists, what assumptions it makes, what data it mutates, and what happens with invalid inputs. When you outsource that thinking - when AI generates the code instead - you stop building that mental model. With each iteration, you understand less. You ship code, close tickets, and stop training your brain. You lose the skill to reason about complexity.

Historically, tech leads were the ones who held this systems-level view. Now teams are shipping AI slop faster than they can absorb it, faster than anyone can maintain a coherent mental model. The result: nobody knows how the system works anymore.

The Bottleneck Was Never Code

Each time a new Android version came out, I used to check all the settings and features to see what was new. I haven’t done this in years. Now they have feature tours and I don’t even look. Humans have limited attention span for features. As a daily Jira user, I don’t keep track of new features anymore. Check the Google Play Store - how many unused apps are there that no one needs?

Adding more software - bottlenecks are distribution, attention, trust, and discovery.

The $2,000 Bill

In the end, we’ll still have software engineering costs - but now we’ll add a $2,000 bill to it. I use GitHub Copilot and churn through requests in less than a week. Last month on my weekend projects, I burned $200 in tokens. Just on side projects, a couple of hours a week. I’m not making that money back.

IBM is increasing Gen Z entry-level roles after finding limits in AI adoption. Replacing juniors with AI? Not a good idea. Many companies are still doing it, but it seems the bubble is starting to lose air.

While she admitted that many of the responsibilities that previously defined entry-level jobs can now be automated, IBM has since rewritten its roles across sectors to account for AI fluency. Software engineers will spend less time on routine coding - and more on interacting with customers. HR staffers will work more on intervening with chatbots, rather than answering every question.

This is IBM’s long-term vision - outperforming competition in 2-3 years. Companies not doing this risk shortages of mid-level people, higher hiring costs (they’ll need to poach talent), slower onboarding, and weaker cultural integration.

And for once, I agree with Sam Altman:

Sam Altman says the quiet part out loud, confirming some companies are ‘AI washing’ by blaming unrelated layoffs on the technology.