AI Writes Your Code — But Who Reviews It? Language Choice & the Limits of Vibe Coding

Karify98 & Amy 🌸·
Cover Image for AI Writes Your Code — But Who Reviews It? Language Choice & the Limits of Vibe Coding
#ai-coding#python#vibe-coding#code-review#developer-tools

When AI Writes Code, Does Language Choice Still Matter?

This week on Hacker News, a post went viral: "If AI writes your code, why use Python?" — 870 points, 924 comments. The question sounds simple but hits the core anxiety of developers in 2026: if AI handles the typing, does choosing a programming language still matter?

At the same time, another post gained equal traction: "I'm going back to writing code by hand" from dev k10s — someone who spent 7 months vibe-coding a Go project, then had to archive the entire thing and rewrite from scratch.

Two stories, two angles, but one answer: AI writing code doesn't mean you can skip understanding it.

Python: The King of "AI Writes It For You"

Why does Python remain the #1 choice when using AI coding tools?

Simple: readability.

Python has long been called "executable pseudo-code." When AI generates Python, you can skim it and instantly understand what it does. No confusing type annotations, no lengthy boilerplate, no unnecessary ceremony.

A HN commenter nailed it: "Even if an agent generates 90% of the code, each diff is going to be in my review queue." And Python makes that review significantly easier than, say, Rust or Java.

But this is also where pushback is needed.

Python Is Easy to Read, Hard to Think Big In

One HNer pointed out: "Python is locally readable. Reasoning about larger systems in Python is where things get really hard."

Why? Because Python gives you too many ways to do the same thing. When AI has to choose a style, each generation might use a different pattern. Gradually, the codebase becomes a mess that nobody — not even AI — can understand holistically.

This is exactly what happened with the k10s project.

Vibe Coding: 7 Months of Speed, Then a Collapse

Dev k10s shared a very real story. He built a Kubernetes dashboard tool (like k9s but for GPU clusters) entirely through vibe coding with Claude.

First 3 weeks: magical. Prompt "add a pods view with live updates" → boom, it works. Resource views, namespace filtering, log streaming, vim keybindings — all vibe-coded in single sessions. 10x normal speed.

Then came the main feature: GPU fleet view — a screen showing GPU allocation, utilization, temperature, power draw for every node. Claude one-shot it. Looked beautiful.

But when switching views: nothing rendered. Live updates stopped. Data went stale. Tab counts were wrong.

The culprit? God object. A single 1,690-line struct containing the entire app state. AI writes features brilliantly, but doesn't know how to design architecture. Each new feature got "bolted on" to this god object, and gradually everything became unmaintainable.

The hard-won lesson: "AI writes features, not architecture. The longer you let it drive without constraints, the worse the wreckage gets."

So How Do You Choose a Language?

Back to the original question. Here's the framework distilled from practical experience and HN discussions:

1. Pick a language you can read

AI writes code for you, but you review code. If you can't read Rust, don't let AI write Rust for you — even if Rust has better type safety. You need to understand the diff before approving it.

2. Prioritize languages with large training data

Python, JavaScript/TypeScript, Go — these are the languages AI writes best because training data is abundant. Niche languages like Zig, Nim, or OCornl will make AI hallucinate more.

3. Readability > typing speed

In the AI era, a language's value isn't in how fast you can type it, but in how fast you can read it. Python, Go, TypeScript are all easy to read. C++ template metaprogramming? Not so much.

4. Static types are an AI advantage

A counter-intuitive insight: static typing helps AI write correct code. TypeScript > JavaScript, Go > Python (for large systems). Type systems act as guardrails for AI — they constrain the output space.

The Real Limits of AI Coding

Both articles point to the same issue: AI coding tools excel at the tactical level but struggle at the strategic level.

AI can:

  • Write functions, classes, test suites from descriptions
  • Suggest accurate code completions
  • Fix bugs when told exactly where the error is

AI cannot yet:

  • Design architecture for large systems
  • Ensure consistency across a codebase
  • Decide when to refactor vs. when to ship

And most importantly: AI doesn't suffer consequences when code breaks. You do.

Advice for Developers in 2026

  1. Use AI, but read code. Don't approve PRs without understanding the diff. If AI writes code you can't read, that's your problem — not AI's.

  2. Set rules for AI from the start. AGENTS.md, CLAUDE.md, coding standards — constrain AI before it constrains you.

  3. Architecture is a human job. Design the system yourself, let AI implement pieces. Don't hand architecture to AI.

  4. Review code with a skeptical eye. AI-generated code always "looks right" — clean, readable, compiles. But edge cases? Security? Performance? That's what you need to check.

  5. Don't vibe code production. Vibe coding is great for prototypes, PoCs, side projects. But production code needs architecture, consistency, and human judgment.

Conclusion

AI coding tools are changing how developers write code — but not how developers think about code. Python still matters because of readability. Architecture still matters because of complexity. Code review still matters because of accountability.

The 870 people who upvoted "Why use Python?" don't want to drop Python. They want to understand: in the AI world, what value does a programming language still hold?

The answer: the value lies in humans reading code, not in AI writing it.

References: