Mark Zuckerberg recently claimed AI will replace mid-level engineers by 2025.

As someone building AI developer tools and studying their real-world implementation, I believe this fundamentally misunderstands both the current state of AI and the role of mid-level engineers.

Here’s what Meta is missing.

Learning why model redundancy > optimization

It started with a frustrating Thursday afternoon. Our code analysis service was hitting rate limits constantly, and I was doing what any reasonable engineer would do: optimizing our token usage, implementing better queuing, and trying to squeeze maximum performance from our chosen model.

Nothing worked. Or rather, everything worked a little bit, but not enough.

Last week, my AI coding assistant generated a perfectly-structured code review suggestion.

The format was immaculate - every field properly typed, every attribute carefully specified, the suggestion clear and actionable.

There was just one problem: it fundamentally misunderstood how our authentication system worked.

“Every developer has that one post they almost didn’t write…”

For 2025, I decided to get back into technical writing.

I’ve been obsessing over getting AI to actually understand production code. Not just generate it, but really grok what’s happening in a mature codebase.

After countless nights exploring this rabbit hole, I wrote about a breakthrough I had.

Sunday morning, I posted it on HackerNews. “Why not,” I thought, “maybe someone will find it interesting.”

Then I went back to my weekend.

A messy experiment that changed how we think about AI code analysis

Last week, I watched our AI choke on a React codebase - again. As timeout errors flooded my terminal, something clicked. We’d been teaching AI to read code like a fresh bootcamp grad, not a senior developer.

Here’s what I mean.

Last month, I asked Claude to help refactor a React component. The code it wrote was beautiful - clean, well-documented, following all the best practices.

It also quietly broke our error tracking system, removed a crucial race condition check (that admittedly looked like a bug), and duplicated three utility functions with slightly different implementations.

Sound familiar?

The AI coding space is exploding right now. Cursor hit $50M ARR, Lovable.ai reached $4M in 4 weeks, and every day there’s a new “AI-powered IDE” on Product Hunt. Clearly, developers want AI assistance.

Yet despite these impressive numbers, I believe current AI coding tools are fundamentally solving the wrong problem. That’s why I’m building another one.

I’ve been running multiple WordPress blogs for my friends and family on my own VPS since ~2012. I didn’t bother checking them for updates, and surprise surprise they all got hacked.

This is my journey of how I fixed it and how the latest version of my blog was born.