Hacker News Reader

Comment by edg5000 | Hacker News Reader

> state of the art LLMs are able to complete large subtasks or medium size projects alone, almost unassisted, given a good set of hints about what the end result should be

No. I agree with the author, but it's hyperbolic of him to phrase it like this. If you have solid domain knowledge, you'll steer the model with detailed specs. It will carry those out competently and multiply your productivity. However, the quality of the output still reflects your state of knowledge. It just provides leverage. Given the best tractors, a good farmer will have much better yields than a shit one. Without good direction, even Opus 4.5 tends to create massive code repetion. Easy to avoid if you know what you are doing, albeit in a refactor pass.

View on HN

I feel like a lot of the disagreement over this "large project" capability is that "large project" can mean anything. It can mean something that has a trillion github repos to work with, or it can mean something that is basically uncharted territory.

If this only works for people with like 10+ years of domain experience, doesnt that make this an Anti-AI article? Whole vibe coding sells on the promise that it works and it works for every tom and their mom.

This conflates two things.

One is LLMs writing code. Not everything and not for everyone. But they are useful for most of the code being written. It is useful.

What it does not do (yet, if ever) is bridging the gap from "idea" to a working solution. This is precisely where all the low-code ideas of the past decades fell apart. Translating an idea in to formal rules is very, very hard.

Think of all of the "just add a button there"-type comments we've all suffered.

"How hard can it be to add just a button?"

Yes that’s how I see it too. It’s a productivity multiplier, but depends on what you put in.

Sure Opus can work fully on its own by just telling it “add a button that does X”, but do that 20 times and the good turns into mush. Steer the model with detailed tech specs on the other hand, and the output becomes magical

Didn't somebody built a rather decent and fully compliant html parser by copy pasting 8000+ tests from another project?

That qualifies as a good set of hints about what the end result should be.

Reply on news.ycombinator.com