Software design is now cheap | Hacker News Reader

Software design is now cheap

dottedmag.net

50 pointsdottedmag77 comments

This article is lumps design and implementation together. In my experience, LLMs are really quite bad at designing anything interesting. They are sort of tolerable at implementation — they’re remarkably persistent (compared to humans, anyway), they will tirelessly use whatever framework, good or bad, you throw at them, and they will produce code that is quite painful to look at. And they’ll say that they’re architecting things and hope you’re impressed.

In my experience, LLMs are bad at designing even repetitive boring things.

(Author here) I meant "design" as in designing physical objects — and all our "programming" is "design" in this definition, because "manufacturing" is done by compilers and bundlers.

And I wouldn't write this article 3 months ago. Since then the quality of the output jumped significantly, it is now possible to put the agent into a proper harness (plan/edit/review/test) and the output is good — and if it's not, you discard it and try again, or point out a detail for the next cycle of improvements.

Yes, this requires a lot of forethought to set up, but it works.

I'm not talking only about "web things", I'm working on a project that involves engineering calculations and a lot of optimization of hot paths, both CPU and GPU.

>LLMs are really quite bad at designing anything interesting

Let’s be honest, how many devs are actually creating something interesting/unique at their work?

Most of the time, our job is just picking the right combination of well-known patterns to make the best possible trade-offs while fulfilling the requirements.

> Most of the time, our job is just picking the right combination of well-known patterns to make the best possible trade-offs while fulfilling the requirements.

Right. I don't trust LLM's to pick the right pattern. It will pick _a_ pattern and it will mostly sorta fulfill the requirements.

Today I asked an LLM (Codex whatever-the-default-is) to implement something straightforward, and it cheerfully implemented it twice, right next to each other, in the same file, and then wrote the actual code that used it and open-coded a stupendously crappy implementation of the same thing right there. The amazing thing is that the whole mess kind of worked.

Right. Kind of works is their MO at the moment. I do try to keep in mind that just because something sucks at the moment, that doesn't mean that it will always suck (especially when you pour _trillions_ of dollars into it)

The problem is that OpenAI and Claude don’t care if their tools produce good code, as long as people pay for them.

Just pick patterns yourself and let LLM fill them in with colours :)

(Author here) I found that over time I spend more time striping someone's badly designed abstractions to get to the real functionality. LLMs are surprisingly good at figuring it out, plowing through the code and documentation and finding out that a 100MB library is in reality a HTTP client for 7 REST endpoints, or something like this.

Unless you work for a consulting firm, you should be working on something new/unique.

It’s a winner-takes-all market. There are no buyers for off brand Salesforce or Uber.

That feels a bit rigid.

Many people are in position where they can’t afford risking their financial future by going all-in on startup. They just want to do honest work in exchange on paycheck and enjoy time with family after 5pm and on weekends.

There are not? So Lyft and bolt do not exists?

Same with Salesforce, there are a few hundred alternatives

I was wondering if our goal is to leverage them to think about interfaces a bit, like a slightly accelerated modeling phase and then let them loose on the implementation (and maybe later let them loose on local optimization tricks)

    > …a slightly accelerated
    > modeling phase and then
    > let them loose on
    > the implementation…

If you mean _visual_ modeling ala UML [1], then I have it on "good authority" [2] that's a sound approach…

_____

MODEL

…

The Verdict: If you provide a clear instruction like "Before you touch the code, read architecture.puml and ensure your changes do not violate the defined inheritance/dependency structure," the agent will be very effective at following it.

If you just "hope" it bears it in mind, it probably won't.

_The agent is a tool, not a mind-reader; it will take the shortest path to a passing test unless you wall that path off with your architectural models_.

…

To make it actually work, you need to turn the UML from a "suggestion" into a "blocker." You should add a section to your AGENTS.md (or CLAUDE.md ) that looks like this:

    1. Tool Trigger: By using words like "…"

…

Why this works:

…

_____

[1] https://news.ycombinator.com/item?id=46974325

[2] https://g2ww.short.gy/TheMightyBooch

UML or else, a set of types / interfaces that represent the system and its properties. thanks for the suggestion

I was going to say - with agents the only part I actually have to do is design. Well, and testing. But they don’t really do design work so much as architecture selection if you provide a design.

This is a great take. It applies the SaaS is dead theory at a lower level (libraries are dead) but has it a much more nuanced view.

Yeh even if LLMs are 10x better than today you probably still don't want to implement cryptography from scratch, but use a library.

I also like the 3d printing analogy. We will see how good LLMs get, but I will say that a lot of AI coded tools today have the same feeling as 3d printed hardware. If no engineer was involved the software is cheap and breaks under pressure because no one considered the edge cases. It looks good on the surface but if you use it for something serious it does break.

The engineer might still use an LLM/3d printer but where necessary he'll use a metal connection (write code by hand or at least tightly guide the LLM) to make the product sturdy.

> It looks good on the surface but if you use it for something serious it does break.

That's LLMs extending C and C++ Undefined Behaviour to every project regardless of language.

-------------------

EDIT: I tried articulating it in a blog post in a sleep-deprived frenzy of writing on Sunday - https://www.lelanthran.com/chap14/content.html

The time might come when the ai just writes things like http request parsers and crypto libraries on the fly. And it will be somewhat normal

But why? Even if you could have an AI do that it’s, if anything, a waste of cpu cycles. If you have a battle tested library that works and has been tested for trillions of request cycles why would you even want to write a new one that needs testing and maintenance? No matter how cheap code gen gets it doesn’t make sense. For something like a UI library, sure build something specific to your needs.

Libraries are really built for human beings, not super intelligent machines. ORMs are built because I don’t like to and can’t write complex sql with every edge case.

Same with a lot of software, software libraries are designed to work with the deficiencies of the human mind.

There’s no reason to think ai needs these libraries in the same way

Even in your scenario LLMs could write super optimized libraries (not intended by for humans) to do common tasks and share the code between them.

I’m not saying the future can’t get to an ai just producing everything. I’m saying it’s just plain inefficient to keep solving the same problem over and over.

I agree with part of this (see my comment above). That said our limitations were also how we produced mathematics. Categorizing world into fixed concepts is valuable i'd say.

This seems like today's version of "I could write Facebook in a weekend."

What are the incentives for doing that? What are the incentives for everyone else to move?

So if proven things exist for basics, what's the incentive to not use them? If everyone decides they're too heavy, they could make and publish new libraries and tools would pick those up. And since they're old, the feature-set is probably more nuanced than you expect. YAGNI is a motto for doing less to avoid creating code debt, but writing more net new code to avoid using a stable and proven library doesn't fit that.

Importing a library for everything will become a dated concept. Similar to the idea that object relational mappers might go away if the ai can just write ultra complicated hyper efficient sql for every query

If AI makes the per line cost of producing software cheaper but you still need an expensive human in the loop then the per line cost is merely cheap not free or at the cost of electricity.

Given the choice between

A) having one AI produce a library and having 1000 produce code using that library which comes with tests human in the loop vetting documentation and examples which drastically increase the chance of the 1000 AIs doing it correctly

B) Having 1001 produce the same functionality provided by the library probably on average worse and requiring more expensive hand holding

What in that benefit of B? You might have slightly higher specificity to your use case but its more likely that the only increased specificity is shit you didn't realize you needed yet and will have to prompt and guide the AI to produce.

I fail to see how AI would obviate the need to modularize and re-use code.

This take is just an intermediary take until ai takes over software engineering. In the same way, eventually self driving cars will make human drivers look dangerous

I think your thought process is not taking into account what a super logical ai can do, and how effortlessly it could generate some of this code.

> effortlessly

Real question: does your world model suppose super exponential gains in intelligence and efficiency?

no i think we are a year or two out from this, and i see glimpses of it daily

Why does the AI need SQL queries? Who needs that? It will just write its own ACID-compliant database with its own language, and while it's at it, reinvent the operating system as well. It's turtles all the way down.

It’s actually not a ridiculous concept, but I think in some ways code will go away, and the agent itself will do what the code used to do. Software of the future will be far more dynamic and on the fly. We won’t have these rigid structures

Why does the AI need hardware/chips? Why does the AI need the universe to exist? Why does the AI need math/logic to exist?

Using these preexisting will all become outdated. You will look like primitive cavemen if your agents don't build these from scratch every time you build $NEXT_BIG_THING.

Even local LLMs will be able to build these from scratch by end of 2026.

Yeah I actually think the very concept of software will see agents affect out. Agents will do what you used to code. And it will be automatic

Yeah, who needs things like PyTorch? Let agents do that!

you realize ive written some of the pytorch source code?

So that's what 10x means - 10 times the code because your app has no dependencies

ORMs have largely been fading away for a while because there are real wins of not using them.

Hyper-optimized HTTP request/response parsing? Yawn. Far less interesting.

AFAICT, the advantages of keeping context tight and focused have not gone away. So there would neeed to be pretty interesting advantages to not just doing the easy thing.

Build times too. I kinda doubt you're setting up strictly-modularized and tightly-controlled bazel builds for all your stuff to avoid extra recompilation... so why are we overcomplicating the easy stuff? Just because "it will probably function just as well"?

"leftpad"-level library inanity? Sure, even less need than before (there was never much). Substantial libraries? What's the point?

Hell, some of the most-used heavily-AI-coded software is going the opposite direction and jumping through hoops to keep using web libraries for UI even though they're terminal apps.

Wh … why would we want that?

Why would you not?

Libraries are more well tested. Unless you "regurgitate code in a library" as "not importing any libraries".

What value is there in tested code when you have the option of superlative, flawless code?

How do you determine flawlessness? How you even approximate a guarantee of it? To what specification is flawlessness judged, and can you precisely and completely relay that spec to your friendly local robot more efficiently than it can vendor / import existing libraries?

It’ll just spit the code out. I vibe coded some with cookie handling the other day that worked. Should I have done it? Nope. But the ai did it and I allowed it

The concept of using a library for everything will become outdated

Either you didn't need a library for that or you could have vibe coded it using the library even easier.

It read the library and created a custom implementation for my use case. The implementation was interoperable with a popular nextjs with library. It was a hack sure, but it also took me three minutes

What is the advantage of coding something from scratch vs importing a well tested library?

If the LLM writes better code (and it will) then its code should be favored over a library.

People aren’t mentally prepared for the LLMs spitting out flawless code

Because it’s not a forgone conclusion that it’s going to happen in our lifetime.

The value of a library is not just that it does a thing you want, but that it doesn’t do all the things you’d prefer it didn’t.

It’s easy to write a cookie parser for a simple case; clearly your robot was able to hand you one for millidollars. How confident are you that you’ve exhaustively specified the exact subset of situations your code is going to encounter, so the missing functionality doesn’t matter? How confident are you that its implementation doesn’t blow up under duress? How many tokens do you want to commit to confirming that (reasoning, test, pick your poison)?

(Author here) That's why I spent a sinificant chunk of the article talking about what should and what should not be reinvented.

My true opinion is that your questions posed here, may become irrelevant in the face of super intelligence

We should probably not be making decisions based on the assumption that super intelligence is coming.

If it does come, nothing we do now really matters. If it doesn’t come and we cash a bunch of checks assuming it will, we are screwed.

I mean ASI could just generate the pixels of a webpage at 60hz. And who needs cookies if ASI knows who you are after seeing half a frame of the dust on your wall, and then knows all the pixels and all the clicks that have transpired across all your interactions with its surfaces. And who needs webpages when ASI can conjure anything on a screen and its robots and drones can deliver any earthly desire at the cost of its inputs. That is if we’re not all made into paper clips or dead fighting for control of the ASML guild at that point.

I say all that mostly in jest, but to take your point somewhat seriously: the most accurate thing you’ve said is that no one is ready for super intelligence. What a great and terrible paroxysm it would be.

Which is the reverse of how humans design things, layers, modules. LLMs act as generalized compilers. Impressive but at the same time you end up with a static-like bunch of files instead of a system of parts (that said I'm not a great user of llms so maybe people managed to produce proto-frameworks with them, or maybe that will be the next phase.. module-oriented llm training).

I was assured manufacturing was dead because we'd all be 3D printing everything we need in our garages.

(Author here) I'm actually talking in the article why it hasn't happened.

Gotta love a world in which a tool which has ingested "all the world's libraries" is now trotted out as a solution to replace those libraries.

You know what would happen if all the people who handwrote and maintained those libraries revoked their code from the training datasets and forbid their use by the models?

:clown face emoji:

This LLM-maxxing is always a myopic one-way argument. The LLMs steal logic from the humans who invent it, then people claim those humans are no longer required. Yet, in the end, it's humans all the way down. It's never not.

> You know what would happen if all the people who handwrote and maintained those libraries revoked their code from the training datasets and forbid their use by the models?

The MCP servers combined with agentic search solved this possibility, just this year superseding RAG methods but all techniques have their place. I don't see much of a future for RAG though, given its computational intensity.

Long story short, training and fine tuning is no longer necessary for an LLM to understand the latest libraries, and therefore the "permission" to train would not even be something applicable to debate

it's a fast moving field, best not to have a strong opinion about anything

LLMs write superior code. Maybe they learned from humans, but they have seen further while standing on those shoulders.

> LLMs write superior code.

How would they know what superior code is? They're trained on all code. My expectation and experience has been that they write median code in the best-case scenario (small greenfields projects, deeply specified, etc).

The AI maximalist are really out in force in this thread.

Then you have access to better models then I do (4.6/5.3)

The code is mostly not bad, but most programmers i have worked with write far better code.

> What matters less is the mechanical knowledge of how to express the solution in code. The LLM generates code, not understanding.

I think it's the opposite -- if you have a good way to design your software (e.g., conceptual and modular), LLM will generate the understanding as well. Design does not only mean code architecture, it also means how you express the concepts in it to a user. If software isn't really understood by humans, I doubt LLMs will be able to generate working code for it anyway, so we get a design problem to solve.

This couldn't be more wrong.

LLM's are only as good as they are because we have such amazing incredible open source software everywhere. Because their job is to look at the types of really good libraries that have decades of direct and indirect wisdom poured into them, and then to be a little glue.

Yes the LLM can go make you alternatives, and it will be mostly fine-ish in many cases. But LLMs are not about pure endless frivolous frontiersing. They deeply reward and they are trained on what the settlers and town planners have done (referencing Wardley here).

And they will be far better at using those good robust well built tools (which they have latently built-in to their models some!) than they will be at re-learning and fine-tuning for your bespoke weird hodgepodge solution.

Cheap design is cheap now. Sure. But good design will be ever more important. Model's ability, their capacity, is a function of what material they can work with, and I can't for the life of me imagine shorting yourself with cheap design like proposed here. The LLM's are very good, but but honing in on good design is hard, period, and I think that judgement and character is something the next orders of magnitude of parameters is still not going to close the gap on.

I think the main point is that "reinventing the wheel" has become cheap, not software design itself.

For example, when a designer sends me the SVG icons he created, I no longer need to push back against just using a library. Instead, I can just give these icons to Claude Code and ask it to "Make like react-icons," and an hour later, my issue is solved with minimal input from me. The LLM can use all available data, since the problem is not new.

But many software problems challenge LLMs, especially with features lacking public training data, and creating solutions for these issues is certainly not cheap.

I really like the jig analogy. But that the software jig is no different from production software, I don't buy. Saying that the bytes are the same, is like saying that the atoms of a jig is the same as the atoms in the final product.

GenAI generally makes software cheaper. But there is a huge difference in how much. Prototypes and jigs may be 90% cheaper (just making up numbers here), while for production software it may be closer to 10%.

I don't buy that.

AI is taking over Senior Devs' Work is the same as IKEA is taking over carpenter's moat - no, no, and again no way.

AI lets you do some impressive stuff, I really enjoy using it. No doubt about that.

But app development, the full Software Delivery Life Cycle - boy, is AI bad. And I mean in a very extreme way.

I talked to a carpenter yesterday about IKEA. He said, people call him to integrate their IKEA stuff, especially the expensive stuff.

And AI is the same.

Configuration Handling: Works on my machine, impressive SaaS app, fast, cool, PostgreSQL etc.

And then there is the final moment: Docker, Live Server - and BOOM! deployment failed.

If you ever happen to debug and fix certain infrastructure and therefore deployment fails - you wish you were doing COBOL or x86/M68000 assembly code like it is 1987 all over again - if you happen to be a very seasoned Senior Dev with a lot of war stories to share.

If you are some vibe coder or consulting MBA - good luck.

AI fails so bad at doing certain things consistently well - and it costs company dearly.

Firing up a Landing Page in React using some Tailwind + ShadCN UI - oh well...

Software Design, Solution Architecture - the hard things are getting harder, not cheaper.

IKEA is great - for certain use cases. It made carpenter's work only more valuable. They thrived because of IKEA, they didn't suffer. In fact, there is more work for them to do. Is their business still hard, of course, but difficult in a different way (talent).

And all doomer's talking about the dev apocalypse - if AI takes over software development, who is in trouble then? Computer Science, software development? Or any and every job market out there?

Think twice. Develop and deploy ten considerably complex SaaS apps using AI and tell me how it went.

Access to information got cheaper. A fool with a tool is still a fool.

It's not cheap to develop new large AI models even when one has data. This is where the most interesting software development is. The limitation is because the required hardware is very expensive to buy and also to rent.

> Less dependency management, less glue, less accidental complexity from adapting your problem to someone else’s abstraction.

Less maintenance and flexibility. You're not really "designing software" until you have a 20+ year old product.

Vibe coders really embody the "temporarily embarrassed billionaire" mindset so perfectly.

We are maintaining and extending a bunch (15 around) large ERP/data type projects which are over 20 years old with vibe coding. We have a very strict system to keep the LLMs in bounds, keeping to standard etc and we are feeling we are close to not having to read the code; we have over 2 months of not having to touch anything after review. I designed most of those projects 20-30 years ago; all have the same design principles which are well documented (by me) so the LLM just knows what to find where and what to do with it. These are large 'living' projects (many updates over decades).

This seems like an uncharitable take. Personally I would refrain from putting a 20 year barrier around the qualification, that seems a little harsh.

TFA's take makes sense in a certain context. Getting a high-quality design which is flexible in desirable ways is now easier than ever. As the human asking an LLM for the design, maybe you shouldn't be claiming to have "designed" it, though.

I was specific to the age of the product not of the age of the developer on that project. The point is that "high quality design" is such a fleeting thing that perhaps "longevity of design" is more worth having. It's also probably the case that the latter is much harder to come by which makes it a perfect barrier of qualification.

This is ridiculous. There are numerous multi billion dollar software companies with customers less than half that age.

Is there some reason to think that poorly designed software cannot be profitable? Perhaps we shouldn't use revenue as a proxy for quality?

More to the point how much of that profit is generated from selling those customers data rather than earning those customers payments?

The problem is that if you don’t understand the code you’re taking a risk shipping that code.

There’s a reason why most vibe coded apps I’ve seen leak keys and have basic security flaws all over the place.

If you don’t know what you’re doing and you’re generating code at scale that you can’t manage you’re going to have a bad time.

The models are trained on all the slop we had to ship under time pressure and swore we’d fix later, etc. They’re not going to autocomplete the good code. They’re going to autocomplete the most common denominator code.

I don’t agree that design is cheap. Maybe for line-of-business software that doesn’t matter much.

Reply on news.ycombinator.com