Hacker News Reader

Comment by bigstrat2003 | Hacker News Reader

bigstrat2003·

As an engineer with 20 years experience, I can say this is bogus. LLMs still - right now, today, with the latest model - make errors that humans would never make. They cannot be trusted. So you have to carefully review everything they produce, which negates the productivity gains. You may as well just write your own code, it will be similarly fast and better quality.

View on HN

I think both of you could be correct based on a whole host of qualifiers like current SDLC landscape, specific business domain, current delivery pipelines, etc.

You will at least need to use LLM's to check for your manually introduced inconsistencies and security problems, like we see right now in kernels and libraries we considered peak of human achievement.

If an LLM can spot those better than a human, I'd wager it can write code without introducing them in the first place.

Eh, I dunno -- I have an LLM write a feature, then create a PR. Subsequent bot reviews list tons of "problems" (some relevant, others not so much). Why couldn't the first LLM get it right the first time?

Why don’t humans get it right the first time?

Ask your LLM to review the PR before submitting it. Then make it fix the problems. Then repeat the process until there are no more problems. At each step, document where it went wrong in a guide; next time, it can use that guide and will make less errors.

In other words, treat it like a junior that has infinite enthusiasm and talent but zero existing habits or wisdom. If you are not doing something like this, then you are going to have a bad time.

It doesn't really if you're already reviewing you team's code anyway. In that case it still saves the time writing the software

Reply on news.ycombinator.com