Hacker News Reader

Comment by akmarinov | Hacker News Reader

Seriously guys, what am I missing that they can do that?

I’ve tried doing just basic json printing apps and the AIs (GPT-4o, Claude Opus, Gemini) often fail me after digging into what needs to be done.

They do great at scaffolding things and getting you maybe 40-50% there, then they hit a brick wall of awful code and not doing what I want.

How can you replace devs for even these simplest cases? It just doesn’t work at the end.

View on HN

I can tell you from first-hand experience that a wall of source code which only works 50% of the time is precisely the level of quality that you can expect from cheap outsourcing. I believe what AI can do is change managers' expectations so that smaller teams become more accepted. And by kicking out the worst performers, you increase average productivity, no matter what the AI does or doesn't do.

I’d believe if you did nothing with AI and just fired a big chunk of your bottom tier cheap outsourced programming “talent” your productivity would significantly increase as a result of your actual talent no longer being occupied with managing and reacting to low quality code contribution and contributor hand holding.

Agree. If someone's productivity is negative, bringing it closer to 0 is an improvement. That means keeping them busy with a useless AI is actually helpful.

The difference between cheap outsourcing by humans and LLMs is that in the former case this 50% of working code remains more or less constant, while with LLMs, I would assume, it gets exponentially lower with every iteration. I use LLMs for my daily work and personal matters a lot and I certainly admire the value they bring but the shortcomings cannot be unseen.

My experience is that the more experienced you are the less benefit you will see from coding asistants. If you are at the beginning of your career and writing a simple for loop requires serious brain power, an AI coding assistant will seriously speed you up. The same goes for task complexity. Mainstream tasks will work better than specialized ones simply due to the lack of adequate training material.

Reply on news.ycombinator.com