Ask HN: What does programming with AI feel like? | Hacker News Reader

Ask HN: What does programming with AI feel like?

9 pointsnotarobot1235 comments

For someone coming back to programming from before the LLM revolution, what does it feel like to be fully productive with AI coding support?

Are you more productive? How? Does it feel like programming at the speed of thought or is it more like pair programming with a less experienced colleague? Is it more/less enjoyable?

Adopting workflows like TDD changes the way you approach a problem. Have AI-centric workflows changed the way you think about programming/problem-solving?

Are there any good demos of these new and improved workflows from people who are productive and have really mastered their tools?

View on HN

I’m much more ambitious going into new languages and paradigms. Didn’t know PyTorch, but shipped projects using it. Don’t know Swift, but I’m working on an iOS app. Don’t know Kotlin but I’m going to try to write an Android app in the next month. Learning is much faster. We’re all going to become generalists again except for a very few hyper-specialist consultants.

Different tools have different experiences. Cursor feels like pair programming with a competent but junior engineer. Agent mode often starts going in loops or changing things that shouldn’t change.

ChatGPT was great early on but I use it less and less these days because it hasn’t kept up with competing products. Claude seems to be where it’s at and I love its “voice.”

If you know how to program already, it’s great. If you don’t, it may help you learn quicker. I doubt there’s much of a future for publishers or online training courses.

I'm an architect with decades of development experience. I know how the tools, technologies, and frameworks work - but I don't necessarily know all the little details like an expert would. I find the AI tools cover that well. I can architect and design the solution, create a skeleton of the solution, and then use AI to rapidly fill in the details and complete the implementation.

It's been working well for me. I don't know how well it would work for junior developers though, because I see some things AI creates that don't look right to me. Then I have to dig in more. Still, as a tool, it's been pretty awesome.

To address your last point, I think there's still a future for publishers and online training courses. They need to focus on the "big picture" - the concerns of architects and designers, and what good practices look like. The publishers also need to publish all the details, so when the AI looks "off" you have resources to double-check.

It feels like having lower mental load because it handles a certain amount of the most granular complexity for you. It’s less tiring and I can code for longer.

I already used TDD on occasion. It’s even more enjoyable now that someone else is doing the tedious work of coming up with (most of) the cases we have to test and drafting up the test code so that I just need to make sure it’s actually checking the things it needs to and that the fixtures are hooked up properly (it struggles with that sometimes).

The people who are best at it seem to come from a documentation-first perspective. Personally I enjoy that even less than TDD, and haven’t tried it because I’m just doing this for fun and what’s the point if it means doing more of the parts I dislike?

I find it is good for many things. A time-saver & productivity booster for domain that's already known aka problems that have already been solved.

For new stuff or stuff not in the public domain like that it is completely worthless. I also have to keep second guessing the results. Fortunate enough quality testing a bunch of code on a carousel to "good enough" status has never been easier.

Evolution of AI is truly revolutionizing how developers code. With AI tools like ChatGPT, Copilot, Gemini, etc more and more developers are becoming dependent on these tools to generate code. But a majority of the popular AI code generators are incapable of understanding the context of complex codes and this leads to incorrect code snippets.

We are working on Potpie (https://github.com/potpie-ai/potpie), which targets this very problem and helps the developers to build AI agents that truly understand your complex codebase and perform desired actions

Reply on news.ycombinator.com