Hacker News Reader

Comment by ori_b | Hacker News Reader

So why would you buy it off of Anthropic?

In any activity you can take shortcuts that makes it easier. It's up to you how many (if any) you want.

Take woodworking for example. When I build a kitchen cabinet, I can get lumber that's already smooth and treated, I can buy drawer tracks, I can use power tools instead of a handsaw and a screwdriver, I can use a pocket hole jig to make joints easier. I still have to do more planning and assembling than with the Ikea cabinet, which also takes more work than having a contractor do everything for me.

I'm doing it my way because it's fun for me. Other people might enjoy other parts of the process - or different things altogether.

There's a whole spectrum between doing everything from scratch and paying someone to have it done for you.

I don't understand the question. For one thing I use local models mainly, but even if I didn't, I'd be buying the tokens from cloud model providers, not the prepackaged, fully complete software itself. I buy the tokens to make what I want.

It's actually quite similar to buying the services of a programmer off Upwork to build something for me, only with LLMs it's way cheaper and faster, with a shorter feedback loop.

I think their point is that you aren't really doing the implementing, Claude (or any model really) is. If you genuinely find prompting LLMs to be fun, then by all means go for it.

What I find fun is getting the output to exactly what I want. I don't care whether I'm personally implementing something or not, and that's what many in this thread seem not to understand.

I'm just gonna hop in and say: I get it.

If I spend a weekend standing up a self-hosted media system or something, I doubt anyone would dispute that's a fun building exercise.

If I do the same thing but use an LLM to build out instead, somehow it's not.

Yeah it's not the same kind of building as what we might have done pre-LLM but it need not be any less satisfying or rewarding.

The real disconnect seems to be the classic dichotomy: people who see coding as the point and the purpose, vs people who just want an outcome.

And that's fine!

I'm just don't understand why the one camp feels the need to deride the other.

Like Lego, assembly can be fun too, and I don't have to manufacture the individual bricks from scratch just to enjoy assembling them together. But no one doesn't call that building something, similar to your experience.

And sometimes buying a statue can be fun, but nobody calls it building a statue.

Help me out with the metaphor here. Do you think designing and architecting some software, then getting an LLM to write a lot of the code, is the same as buying a computer program from a store?

The better AI gets, the closer it gets, yes.

It really depends how you use it. There's a whole spectrum from "do it yourself" to "get the AI to do everything". Its up to each person, on each project, to decide what amongst all of that work is actually worth human time.

I don't enjoy CSS, so I delegate that to an LLM. Just like I don't like writing assembly by hand, so I delegate that to a compiler. So what?

[flagged]

Interesting that you think building is just coding.

What do you think architects do? Or interior designers? Or civil engineers?

Interesting that you think coding just typing. Code is just a language where the problem is specified in fine detail; the biggest value proposition of an LLM is being able to hand-wave and let some other tool take care of guessing at detail, where you can't be bothered to specify it in full. And, part of the process of specifying in full forces you to rethink design assumptions.

Architecture certainly isn't building, and neither is interior design. Civil engineers calculate and specify the loads in excruciating detail, because if they didn't, people would die.

No, coding is the act of reifying all the things that actually matter--the requirements, the visual design, the system design, etc--into a form that a computer can execute.

The biggest value proposition of an LLM is being able to focus on the truly high-value activities while allowing the machine take care of much of that reification.

That you think architecture or interior design isn't building tells you prefer to downplay or devalue any work that isn't hands on construction. It's an interesting perspective, but it's one I'll never be able to understand or agree with.

If you give up on the details, you lose the ability to do a good job on the high level structure. The entire purpose of the high level structure is setting things up so that the details fit cleanly.

For junior developers, I absolutely agree that's a major risk, and it's something I'm deeply worried about as these tools take over the industry, as understanding those details helps inform why a given architecture is needed.

For someone who's been around the block a few times, there's little new under the sun. Coming up with good architecture for a problem space and having the LLM fill in the details is incredibly effective as a pattern.

That said, tbh I'm not sure why we keep exchanging messages. Your other comment about coding with AI assistance as being akin to buying a statue tells me you have either never coded with an agent, and so think it's a simple matter of saying "build me a to-do app" and then calling it a day, or you're experienced with these tools and deliberately misrepresenting the process.

If it's the former, then we don't share enough understanding to have a healthy conversation, and if it's the latter then there's little point in arguing with someone who insists in doing so in bad faith.

Reply on news.ycombinator.com