Hacker News Reader

Comment by SanderNL | Hacker News Reader

The thing to remember about recent AI developments is that they speed up development itself. Copilot is already useful. It does not write its own code, but that’s not the bar. The bar to clear is, can this speed up itself and I think the answer is yes even if humans stay in the loop.

AI that can code is one the most awesome and at the same time tragically idiotic moves we can make. Even if they are only 1% effective, it’s a one percent we did not have and it makes iterating on the next version faster.

Making LLMs and other types of models is becoming a toy project already. This will only accelerate.

View on HN

We've basically already seen this with computer aided design. It's essentially the root of Moore's law.

I'm not entirely convinced AI-aided software development (via copilot or whatever) is actually going to speed up development. It makes writing code faster, but pressing the keys or implementing simple functions was never the slow or difficult part.

The problem for the last several decades has been managing complexity, and that's something AI-aided programming doesn't help with. If anything my hunch is that it makes the problem worse. The increased speed of producing code means designs become even more shortsighted, the spaghetti even more byzantine and entangled.

I upvoted when I had read up to where you wrote "doesn't help with." I felt I was reading my own thoughts when I read the last two sentences. I think it is not just software design that will become shortsighted we may even see less interest in improving programming languages. I've always hoped for people to come up with some mathematical insights (beyond functional programming) that will describe code organizing structures in a more sensible way, so that experience no longer appears as the main argument for software patterns being tools to manage complexity. Now a programming language is just another language to be modeled, and its texts generated and proofread by a fallible human.

> I think it is not just software design that will become shortsighted we may even see less interest in improving programming languages.

I believe in the contrary. AI code generators and modern languages that enforce type and memory safety are a great match, since the compiler ensures the mess doesn't grow too much.

This is an entirely class of problems than what I'm describing.

Keypressing is a hurdle for me. I know that comes with experience. Very little if any software is truly novel and it all looks and feels very similar. Everything I see is patterns and boilerplate. I do not see the insurmountable complexity anymore.

Scientists are the ones coming up with novelty. We hook up databases to UIs and manage infrastructure. Sure there is complexity but it is of a very superficial kind. Like a ball of Christmas lights.

If there is a universal pattern that lets you effortlessly design large non-trivial software systems across all domains, please do publish it. It will do more for productivity than Copilot or ChatGPT ever will.

Well, if you ask how to design “large” systems then the answer is already “don’t”. And besides that’s not a technical issue, our issues are organizational and political.

Also, yes there is an universal pattern and that is called a competent and experienced software engineer. We lack those. We have a lot of kids with fancy degrees.

There's large systems that do useful things though that don't decompose into smaller systems. Shouldn't we have those?

Compilers. VMs. Operating systems. Search engines. LLMs...?

The core tech of LLMs is famously simple. All the rest decomposes just fine. I know we suck at it and there often nontechnical forces that steer us into other directions, but I don’t see big fundamental issues AI can’t already help with.

I know that there is complexity, but the essential part of it is not our job. That is my point. We are accidental complexity managers. I know we like to think working on compilers and browsers is magic, but compared to other fields it’s nothing special.

And by the way, you and I know that exactly 6.78% of devs are employed in those domains and of those 1.12% work on core parts.

I am sorry if I come off annoying and stupid because I am both. You of course raise an interesting point it’s just that I don’t feel like going into wide(r) tangents.

Why are you working on the boring parts, just gluing together prefab components?

See my last paragraph. Also it’s all glueing prefab parts.

Software dev is a dead end job anyway, all parts of it.

Huh. I don't glue prefab parts when I develop.

You do OS development?

No, but I've built an internet search engine.

Oh right, nice one btw, but seriously how are you not glueing prefabs?

Are you writing your own db engine? Network layer?

Yes on the db engine, at least for the document indexing part. I'm not building absolutely everything from scratch, a significant part of it is bespoke. There's just not a lot of off-the-shelf stuff that's built to deal with this type of application.

You do occasionally see attempts at building internet search engines out of like elasticsearch or stuff like that, but it just doesn't scale.

You're conflating the "code" of an LLM which is just basic calculus with the vast data, fine tuning, human confirmation and filtering that make these things work.

You're also conflating the downloading and fine tuning of pretrained models with the creation of foundational models.

AI can iterate on model architecture all it wants, that's really 10% of the solution.

No, I mean basic development of all kind - not just AI directly - is sped up.

The fact that downloading and finetuning models is the only hard part of these systems is a testament to how frigging far along we came.

We forgot just how hard software development and compute in general was just a short time ago.

Yeah I agree with this, and at the same time people's expectations grow to meet the capability.

I remember when writing some C to do some basic network IO felt like witchcraft.

Writing code gets faster but programs get slower. Sometimes I wonder what a modern desktop would be capable of if every piece of code was written in assembly with as much care and purposefulness as, say, an Atari game.

Everything would be basically instantaneous.

It would likely have massive security holes you could drive a truck through.

Reply on news.ycombinator.com