News

AI Writes the Code. You Still Have to Make It Work.

A widely discussed essay on Hacker News this week puts a precise name on the anxiety many engineers have felt: AI tools are making the easy part of software development almost free, while exposing just how costly the hard part remains.

The post, written by Anuradha Weeraman and titled "The Prototype Isn't the Product," argues that AI coding assistants have collapsed the time it takes to go from idea to working demo — but the gap between a demo and a shippable product is as wide as ever. The prototype runs on a developer's laptop. It breaks under real load. It has no error handling. Turning it into something production-ready still requires the same architectural thinking and disciplined engineering it always did.

Context

Anuradha Weeraman's essay argues that AI models lack judgment — they have pattern matching paired with an eagerness to produce code. That framing aligns with broader industry data: according to research cited by software development consultancy Last Rev, Gartner found that on average only 48% of AI projects make it into production, with the journey taking roughly eight months.

What the gap looks like in practice

Weeraman walks through concrete failure modes AI tools routinely produce: queries that cause full table scans on tables with tens of millions of rows, caching strategies that create race conditions under concurrent load, and authentication logic held together by assumptions rather than solid design. Each passes the demo test. None survive a production launch.

Fixing these is not primarily a debugging task. It is an architectural reasoning task, requiring someone who already understands what the system is supposed to do and why. Last Rev's analysis concludes that failures are almost always in the engineering — not the AI itself.

What this means for job seekers

The employers asking the most interesting questions right now are not asking whether you use AI tools — they assume you do. They are asking whether you can evaluate what the tool produced. Can you spot the full table scan before it hits production? Can you identify the race condition in the caching layer? Can you rewrite the authentication flow so it is secure rather than merely functional?

That skill gap runs along a predictable fault line: experience. Engineers who have lived through a production incident caused by a scaling oversight understand why AI-generated code needs human review. Those who have only worked in demo environments often do not.

If you are building your profile right now — as a developer, a data analyst, or a product manager who commissions AI-generated work — the most valuable thing you can show an interviewer is not a portfolio of things AI helped you build. It is evidence that you understand what could go wrong and took steps to address it. A write-up of a production issue you caught, an explanation of why you refactored something AI generated, a documented trade-off you made: these signals separate people who use AI as a force multiplier from those who use it as a crutch.

Our guide to job searching in the AI era explores more ways to position these judgment-based skills in your applications.