Comment by zamadatix | Hacker News Reader

Comment by zamadatix | Hacker News Reader

Parent Full thread

It's a quick path to having something to play with but I'd encourage folks specifically wanting to learn something new or create something for a portfolio to try to have it be more the other way around: write 90% by hand, implement the features yourself, and have AI help you with the research and debugging (but not try to automatically write the code to fix the bugs). AI is really great at all of the things listed, but I guarantee if you were in an interview and someone asked you about this project your responses about how and why everything works would be 2x-3x stronger.

If this was your 3rd time implementing an orbit simulator and you just wanted to get the bulk done quick so you could move on to work on new stuff then that's when having AI write the bulk & doing checking really works best as you're already an expert in what needs to be done rather than trying to learn about what needs to be done from someone/something else doing it for you.

I say this as an orbital dynamics simulator in C++ having been one of my first big projects I put in a portfolio myself :). And I also like to use AI often, I'm not an AI hater in general, just I like to use in certain specific ways sometimes.

Feedback on the UI side: It'd be cool if the stars used more random positions rather than sine/cosine looking patterns. There are lots of different approaches one could try (from simple random positions to very complex types of clustered randomness and brightness variation to realistic star maps). My suggestion would be to ask AI what types of approaches could be taken and try implementing whichever approach sounds the most fun!

I completely agree with what your saying, with one caveat. If I understand how the underlying physics and dynamics are working but don't know how to write the code then theres no world where I could make something like this. The way that I see it is that AI is a tool that really depends on the person.

If I were to be someone who just tells AI "implement drag" and lets it do it then sure, im not learning, but if I do my research outside and just use AI to translate what I give it into the language of computers I feel like im not only building something cool but im understanding whats happening because AI is just translating.

So basically while yes I do believe that AI can be harmful if you approach it inproperly, it helps novice programers implement cool things by just using english.

>If I understand how the underlying physics and dynamics are working but don't know how to write the code then theres no world where I could make something like this.

You could learn coding same way as you learned the physics and dynamics. Programming and physics aren't mutually exclusive. Actually every physicist is (was?) required to know (multiple) programming (languages).

>If I were to be someone who just tells AI "implement drag"

That'll mean at least you understand drag. Could do even worse, regards to learning at least rather result, prompting something like "make a cool-looking physically-realistic 2d rocket launch simulator", which we're at point that will most certainly return a functional app.

>but im understanding whats happening

Do you though? You depend on AI correctly translating your natural language input to code. Though arguably this is something LLMs excel at, since math (logic) also plays role you've to be able to at least read and review the resulting code for correctness. (Assuming you actually care about the physical accuracy that is.)

Curious what age bracket you’re in? I understand where you’re coming from, but I don’t think it will be the paradigm for the future. My own experience is that you actually learn a lot by having AI write piecemeal code and then do overall review.

He is 17

Author 17, me 31. Having AI write your physics code by telling it what to do is a relatively poor way to learn either physics or code, but I agree there are ways to use AI to learn things - just not necessarily this particular way here.

The main thing is the author did normal research and then gave it to AI to make it happen and verified it seemed to match what they asked for. In that, AI helps build the demo but it's not really adding anything to the understanding of the physics than reading about it in the first place did (akin to reading a post of someone else doing something, reviewing their code, and putting it in your project rather than trying to work with the new ideas yourself). Similarly, having AI write 90% of the code is not really helping you get familiar with the code any more than telling it which physics formulas to implement helps you get more familiar with the formulas.

OTOH, asking AI to help validate your research and understanding before going in to write the code helps you learn things like "the rocket shouldn't be flying right at takeoff just because the planet is spinning". Having AI help debug code issues you can't figure out in the code you wrote yourself (but not having it fix them for you directly) helps you really test 100% of the understanding you're trying to demonstrate instead of testing if you can say another's implementation seems reasonable.

If the only goal is to get to generate output instead of learn, then the calculus of course changes. E.g. yesterday I wanted to run an optical flow method for heart rate posted here in my browser. I wasn't really seeking to learn about it, I just didn't have python set up on my machine with the good webcam and wanted to play with the author's demo live. So I copied the article into AI and told it to make it into a single page HTML app. I didn't learn much of anything more than the article already covered in doing so, however I did achieve my goal of having some output to mess with instead of learning.

Perhaps worth adding I just completed my first college degree last year (figured it'd be fun to get one), so perhaps a bit more recent experience with the learning eras they may be about to enter than a typical 30-40 year old range.

Reply on news.ycombinator.com