The difference is that 3D printing still requires someone, somewhere to do the mechanical design work. It democratises printing but it doesn't democratise invention. I can't use words to ask a 3d printer to make something. You can't really do that with claude code yet either. But every few months it gets better at this.

The question is: How good will claude get at turning open-ended problem statements into useful software? Right now a skilled human + computer combo is the most efficient way to write a lot of software. Left on its own, claude will make mistakes and suffer from a slow accumulation of bad architectural decisions. But, will that remain the case indefinitely? I'm not convinced.

This pattern has already played out in chess and go. For a few years, a skilled Go player working in collaboration with a go AI could outcompete both computers and humans at go. But that era didn't last. Now computers can play Go at superhuman levels. Our skills are no longer required. I predict programming will follow the same trajectory.

There are already some companies using fine tuned AI models for "red team" infosec audits. Apparently they're already pretty good at finding a lot of creative bugs that humans miss. (And apparently they find an extraordinary number of security bugs in code written by AI models). It seems like a pretty obvious leap to imagine claude code implementing something similar before long. Then claude will be able to do security audits on its own output. Throw that in a reinforcement learning loop, and claude will probably become better at producing secure code than I am.

Both of those are fixed, unchanging, closed, full information games. The real world is very much not that.

Though geeks absolutely like raving about go and especially chess.

Yeah but, does that actually matter? Is that actually a reason to think LLMs won't be able to outpace humans at software development?

LLMs already deal with imperfect information in a stochastic world. They seem to keep getting better every year anyway.

I don't buy the whole "LLMs will be magic in 6 months, look at how much they've progressed in the past 6 months". Maybe they will progress as fast, maybe they won't.

If this trend continues, the models will be better than me in less than a decade. Unless progress stops, but I don’t see any reason to think that would happen.

Setting aside any implications for your analogy. This is now possible.

Workflow can be text-to-model, image-to-model, or text-to-image to model.

I’m not a fan of analogies, but here goes: Apple don’t make iPhones. But they employ an enormous number of people working on iPhone hardware, which they do not make.

If you think AI can replace everyone at Apple, then I think you’re arguing for AGI/superintelligence, and that’s the end of capitalism. So far we don’t have that.

The first part is making sure you built to your specification, the second thing is making sure you built specification was correct.

The second part is going to be the hard part for complex software and systems.

I don't know about you, but I'd much rather be shown a demo made by our end users (with claude) than get sent a 100 page spec. Especially since most specs - if you build to them - don't solve anyone's real problems.

Demo, don't memo.

Demo for the main flow is easy. The hard part is thinking through all the corner cases and their interactions, so your system robustly works in real world, interacting with the everyday chaos in a non-brittle fashion.

Lol I've been programming for 30 years.

> The complexity of these systems is crazy. Unless he meant ah HTML text area with "save" button - then sure, why not.

What do you see as the difference between an LLM making an HTML text area and a save button, and an LLM making MS word? It just sounds like a scaling problem to me. We've been scaling computers since long before I was born. My first computer was a 386 with 4mb of ram. You needed a special add-in chip to enable floating point calculations. Now look at what we have.

As far as I can tell, the only difference between opus 4.6 and some future AI model that could code up MS word is a difference in scale. Are you betting against the entire computing (software and hardware) industry being unable to scale LLMs past their current point? That seems like a really bad bet to me. Especially seeing how far they've come in the last few years. Claude code can already do some quite complex tasks. I got it to write a simple web based email client for me yesterday. It took about an hour in total. It has some bugs, but the email client works.

We scaled hard drives. We scaled down silicon chips. We scaled digital camera sensors. And display resolutions. And networking bandwidth. We went from the palm pilot to the first iphone to modern phones. Do you really think we'll be unable to scale AI models?

100% bet - no way any "AI" will be able to generate you anything close to a complex piece of software like Ms Word within reasonable time and budget. Given infinite time and money - sure, anything is possible, just like a trilling monkeys randomly printing "War and Peace" once in a trillion years in some remote galaxy. I don't even understand your confidence given how much guidance and hand holding LLMs need at the moment to produce anything useful.

There are clearly two camps - one points to existing deficiencies, another - to trends, and getting wildly different predictions.

I'm looking at the trend line. A few years ago it couldn't make a simple webpage. Now it can make a bad C compiler in thousands of dollars of tokens. What does it look like in another few years? Or another 2 decades?

I'd much rather have a conversation with them to discuss their current problems and workflow, then offer my ideas and solutions.

Not going to. Is. Actually, always has been; it isn’t that coding solutions wasn’t hard before, but verification and validation cannot be made arbitrarily cheap. This is the new moat - if your solutions require time consuming and expensive in dollar terms qa (in the widest sense), it becomes the single barrier to entry.

You can: the words are in the G-code language.

I mean: you are used to learn foreign languages in school, so you are already used to formulate your request in a different language to make yourself understood. In this case, this language is G-code.

I spent years writing a geometry and gcode generator in grasshopper. I wasn’t generating every line of gcode (my typical programs are about 500k lines), but I write the entire generator to go from curves to movements and extrusions.

I used opus to rewrite the entire thing, more cleanly, with fewer bugs and more features, in an afternoon. Admittedly it would have taken a lot longer without the domain expertise from years of staring at geometry and gcode side by side.

Produce the 3D images of xxx from various angles.xxx should be able to do yyy.

This is the tricky part. Do you know anything about mechanical engineering?

Not saying adding few novel ideas (perhaps working world models) to the current AI toolbox won't make a breakthrough, but LLMs have their limits.

https://en.wikipedia.org/wiki/Ninety%E2%80%93ninety_rule

Except that the either side of it is immensely cheaper now.

How long before those lines cross? Intuitively it feels like we have about 2-3 years before claude is better at writing code than most - or all - humans.

I don't see it in practice though.

The fundamental problem hasn't changed: these things are not reasoning. They aren't problem solving.

They're pattern matching. That gives the illusion of usefulness for coding when your problem is very similar to others, but falls apart as soon as you need any sort of depth or novelty.

I haven't seen any research or theories on how to address this fundamental limitation.

The pattern matching thing turns out to be very useful for many classes of problems, such as translating speech to a structured JSON format, or OCR, etc... but isn't particularly useful for reasoning problems like math or coding (non-trivial problems, of course).

I'm pretty excited about the applications for AI overall and it's potential to reduce human drudgery across many fields, I just think generating code in response to prompts is a poor choice of a LLM application.

Have you actually tried the latest agentic coding models?

Yesterday I asked claude to implement a working web based email client from scratch in rust which can interact with a JMAP based mail server. It did. It took about 20 minutes. The first version had a few bugs - like it was polling for mail instead of streaming emails in. But after prompting it to fix some obvious bugs, I now have a working email client.

Its missing lots of important features - like, it doesn't render HTML emails correctly. And the UI looks incredibly basic. But it wrote the whole thing in 2.5k lines of rust from scratch and it works.

This wasn't possible at all a couple of years ago. A couple of years ago I couldn't get chatgpt to port a single source file from rust to typescript without it running out of context space and introducing subtle bugs in my code. And it was rubbish at rust - it would introduce borrow checker problems and then get stuck, trying and failing to get it to compile. Now claude can write a whole web based email client in rust from scratch, no worries. I did need to manually point out some bugs in the program - claude didn't test its email client on its own. There's room for improvement for sure. But the progress is shocking.

I don't know how anyone who's actually pushed these models can claim they haven't improved much. They're lightyears ahead of where they were a few years ago. Have you actually tried them?

I've been disappointed every time.

I do use the LLMs for summarization and "a better google" and am constantly confronted with how inaccurate they are.

I haven't tried with code in the past couple months because to be completely honest, I just don't care.

I enjoy my craft, I enjoy puzzling and thinking through better ways of doing things, I like being confronted with a tedious task because it pushes me towards finding more optimal approaches.

I haven't seen any research that justifies the use of LLMs for code generation, even in the short term, and plenty that supports my concerns about mid to long term impact on quality and skills.

So the TL;DR version is: nah.

i told my boss (not fully serious) we should ban anyone with less than 5 years experience from using the ai so they learn to write and recognize good code.

The LLM is a stochastic parrot. It will never be anything else unless we develop entirely new theories.

And, pray tell, how people are going to come up with such design?

The other day I tested an AI by giving it a folder of images, each named to describe the content/use/proportions (e.g., drone-overview-hero-landscape.jpg), told it the site it was redesigning, and it did a very serviceable job that would match at least a cheap designer. On the first run, in a few seconds and with a very basic prompt. Obviously with a different AI, it could understand the image contents and skip that step easily enough.

It's kind of telling that the number of apps on Apple's app store has been decreasing in recent years. Same thing on the Android store too. Where are the successful insta-apps? I really don't believe it's happening.

https://www.appbrain.com/stats/number-of-android-apps

I've recently tried using all of the popular LLMs to generate DSP code in C++ and it's utterly terrible at it, to the point that it almost never even makes it through compilation and linking.

Can you show me the library of apps you've launched in the last few years? Surely you've made at least a few million in revenue with the ease with which you are able to launch products.

There's a really painful Dunning-Kruger process with LLMs, coupled with brutal confirmation bias that seems to have the industry and many intelligent developers totally hoodwinked.

I went through it too. I'm pretty embarrassed at the AI slop I dumped on my team, thinking the whole time how amazingly productive I was being.

I'm back to writing code by hand now. Of course I use tools to accelerate development, but it's classic stuff like macros and good code completion.

Sure, a LLM can vomit up a form faster than I can type (well, sometimes, the devil is always the details), but it completely falls apart when trying to do something the least bit interesting or novel.

They wouldn’t even know where to begin!

Even if all sandboxing is done right, programs will be depended on to store data correctly and to show correct outputs.

I'm in a similar domain, the AI is like a very energetic intern. For me to get a good result requires a clear and detailed enough prompt I could probably write expression to turn it into code. Even still, after a little back and forth it loses the plot and starts producing gibberish.

But in simpler domains or ones with lots of examples online (for instance, I had an image recognition problem that looked a lot like a typical machine learning contest) it really can rattle stuff off in seconds that would take weeks/months for a mid level engineer to do and often be higher quality.

Right in the chat, from a vague prompt.

I think exceptional work, AI tools or not, still takes exceptional people with experience and skill. But I do feel like a certain level of access to technology has been unlocked for people smart enough, but without the time or tools to dive into the real industry's tools (figma, code, data tools etc).

I think the idea that LLM's will usher in some new era where everyone and their mom are building software is a fantasy.

I am usually a bit of an AI skeptic but I can already see that this is within the realm of possibility, even if models stopped improving today. I think we underestimate how technical things like WIX or Squarespace are, to a non-technical person, but many are skilled business people who could probably work with an LLM agent to get a simple product together.

People keep saying code was never the real skill of an engineer, but rather solving business logic issues and codifying them. Well people running a business can probably do that too, and it would be interesting to see them work with an LLM to produce a product.

In the same vein, I think you underestimate how much "hidden" technical knowledge must be there to actually build a software that works most of the time (not asking for a bug-free program). To design such a program with current LLM coding agents you need to be at very least a power user, probably a very powerful one, in the domain of the program you want to build and also in the domain of general software. Maybe things will improve with LLM and agents and "make it work" will be enough for the agent to create tests, try extensively the program, finding bugs and squashing them and do all the extra work needed, who know. But we are definitely not there today.

The walls and plateaus that have been consistently pulled out from "comments of reassurance" have not materialized. If this pace holds for another year and a half, things are going to be very different. And the pipeline is absolutely overflowing with specialized compute coming online by the gigawatt for the foreseeable future.

So far the most accurate predictions in the AI space have been from the most optimistic forecasters.

No such thing as trajectory when it comes to mass behavior because it can turn on a dime if people find reason to. Thats what makes civilization so fun.

Im really tired, and exhausted of reading simple takes.

Grok is a very capable LLM that can produce decent videos. Why are most garbage? Because NOT EVERYONE HAS THE SKILL NOR THE WILL TO DO IT WELL!

I don't know if they will ever get there, but LLMs are a long ways away from having decent creative taste.

Which means they are just another tool in the artist's toolbox, not a tool that will replace the artist. Same as every other tool before it: amazing in capable hands, boring in the hands of the average person.

This matters less for text (including code) because you can always directly edit what the AI outputs. I think it's a lot harder for video.

I wonder if it would be possible to fine train an AI model on my own code. I've probably got about 100k lines of code on github. If I fed all that code into a model, it would probably get much better at programming like me. Including matching my commenting style and all of my little obsessions.

Talking about a "taste gap" sounds good. But LLMs seem like they'd be spectacularly good at learning to mimic someone's "taste" in a fine train.

True. But quantity has a quality of its own.

I'm personally delighted at the idea of outsourcing all the boring cookie cutter programming work to an AI. Things like writing CSS, plumbing between my database, backend server and web UI. Writing and maintaining tests. All the stuff that I've done 100 times before and I just hate doing by hand over and over again.

There's lots of areas where it doesn't really matter that the code it produces isn't beautifully terse and performant. Sometimes you just need to get something working. AIs can do weeks of work in an afternoon. The quality isn't as good. But for some tasks, that's an excellent trade.

It's driven by it in the sense that better tools and the democratization of them changes people's baseline expectations.

It's independent of it in that doing the baseline will not stand out. Jurassic Park's VFX stood out in 1993. They wouldn't have in 2003. They largely would've looked amateurish and derivative in 2013 (though many aspects of shot framing/tracking and such held up, the effects themselves are noticeably primitive).

Art will survive AI tools for that reason.

But commerce and "productivity" could be quite different because those are rarely about taste.

How can I proclaim what I said in the comment above? Because Ive spent the past week producing something very high quality with Grok. Has it been easy? Hell no. Could anyone just pick up and do what Ive done? Hell no. It requires things like patience, artistry, taste etc etc.

The current tech is soul-less in most people hands and it should remain used in a narrow range in this context. The last thing I want to see is low quality slop infesting the web. But hey that is not what the model producers want - they want to maximize tokens.

With Opus 4.6 I'm seeing that it copies my code style, which makes code review incredibly easy, too.

At this point, I've come around to seeing that writing code is really just for education so that you can learn the gotchas of architecture and support. And maybe just to set up the beginnings of an app, so that the LLM can mimic something that makes sense to you, for easy reading.

And all that does mean fewer jobs, to me. Two guys instead of six or more.

All that said, there's still plenty to do in infrastructure and distributed systems, optimizations, network engineering, etc. For now, anyway.

HN is a echo chamber of a very small sub group. The majority of people can’t utilize it and needs to have this further dumbed down and specialized.

That’s why marketing and conversion rate optimization works, its not all about the technical stuff, its about knowing what people need.

For funded VC companies often the game was not much different, it was just part of the expenses, sometimes a lot sometimes a smaller part. But eventually you could just buy the software you need, but that didn’t guarantee success. Their were dramatic failures and outstanding successes, and I wish it wouldn’t but most of the time the codebase was not the deciding factor. (Sometimes it was, airtable, twitch etc, bless the engineers, but I don’t believe AI would have solved these problems)

Tbh, depending on the field, even this crowd will need further dumbing down. Just look at the blog illustration slops - 99% of them are just terrible, even when the text is actually valuable. That's because people's judgement of value, outside their field of expertise, is typically really bad. A trained cook can look at some chatgpt recipe and go "this is stupid and it will taste horrible", whereas the average HN techbro/nerd (like yours truly) will think it's great -- until they actually taste it, that is.

This is the schtick though, most people wouldn't even be able to tell when they taste it. This is typically how it works, the average person simply lacks the knowledge so they don't even know what is possible.

Agreed. Honestly, and I hate to use the tired phrase, but some people are literally just built different. Those who'd be entrepreneurs would have been so in any time period with any technology.

1) I don’t disagree with the spirit of your argument

2) 3D printing has higher startup costs than code (you need to buy the damn printer)

3) YOU are making a distinction when it comes to vibe coding from non-tech people. The way these tools are being sold, the way investments are being made, is based on non-domain people developing domain specific taste.

This last part “reasonable” argument ends up serving as a bait and switch, shielding these investments. I might be wrong, but your comment doesn’t indicate that you believe the hype.

Low quality music made in bulk seems much less useful than low quality code made in bulk.

They would get amazing amounts done, but no one else could understand the internals because they were so uniquely shaped by the inner nuances of one mind.

Software exists as part of an ecosystem of related software, human communities, companies etc. Software benefits from network effects both at development time and at runtime.

With full custom software, you users / customers won't be experienced with it. AI won't automatically know all about it, or be able to diagnose errors without detailed inspection. You can't name drop it. You don't benefit from shared effort by the community / vendors. Support is more difficult.

We are also likely to see "the bar" for what constitutes good software raise over time.

All the big software companies are in a position to direct enormous token flows into their flagship products, and they have every incentive to get really good at scaling that.

Instead software development would just become a tool anybody could use in their own specific domain. For instance if a manager needs some employee scheduling software, they would simply describe their exact needs and have software customized exactly to their needs, with a UI that fits their preference, ready to go in no time, instead of finding some SaaS that probably doesn't fit exactly what they want, learning how to use it, jumping through a million hoops, dealing with updates you don't like, and then paying a perpetual rent on top of all of this.

But your hypothetical manager who needs employee scheduling software isn't paying for the coding, they're paying for someone to _figure out_ their exact needs, and with a UI that fits their preference, ready to go in no time.

I've thought a lot about this and I don't think it'll be the death of SaaS. I don't think it's the death of a software engineer either — but a major transformation of the role and the death if your career _if you do not adapt_, and fast.

Agentic coding makes software cheap, and will commoditize a large swath of SaaS that exists primarily because software used to be expensive to build and maintain. Low-value SaaS dies. High-value SaaS survives based on domain expertise, integrations, and distribution. Regulations adapt. Internal tools proliferate.

Back in the 1980s this was called "systems analysis". The role disappeared a bit before the web came along, and coders were tasked with the job or told to just guess what the exact needs are, which is why so much software is trash.

I don't know, though, Claude Opus is most of the way to being a good systems analyst, and early reports say that having an AI provide descriptions/requirements to a fleet of code-writing AIs gives better results than having a human do it.

Troubleshooting and fixing the big mess that nobody fully understands when it eventually falls over?

If that's actually the future of humans in software engineering then that sounds like a nightmare career that I want no part of. Just the same as I don't want anything to do with the gigantic mess of Cobal and Java powering legacy systems today.

And I also push back on the idea that llms can't troubleshoot and fix things, and therefore will eventually require humans again. My experience has been the opposite. I've found that llms are even better at troubleshooting and fixing an existing code base than they are at writing greenfield code from scratch.

They are much worse at code/patterns/apis that were locally created, including things created by the same LLM that's trying to fix a problem.

I think LLMs are also creating a decline in the amount of good troubleshooting information being published on the internet. So less future content to scrape.

If most software is just used by me to do a specific task, then being able to make software for me to do that task will become the norm. Following that thought, we are going to see a drastic reduction in SASS solutions, as many people who were buying a flexible-toolbox for one usecase to use occasionally, just get an llm to make them the script/software to do that task as and when they need it, without any concern for things like security, longevity, ease of use by others (for better or for worse).

I guess what im circling around is that if we define engineering as building the complex tools that have to interact with many other systems, persist, be generally useful and understandable to many people, and we consider that many people actually dont need that complexity for their use of the system, the complexity arises from it needing to serve its purpose at huge scale over time. then maybe there will be less need for enginners, but perhaps first and foremost because the problems that engineering is required to solve are much less if much more focused and bespoke solutions to peoples problems are available on demand.

As an engineer i have often felt threatened by LLMs and agents of late, but i find that if i reframe it from Agents replacing me, to Agents causing the type of problems that are even valuable to solve to shift, it feels less threatening for some reason. Ill have to mull more.

Google's weird AI browser project is kind of a step in this direction. Instead of starting with a list of programs and services and customizing your work to that workflow, you start with the task you need accomplished and the operating system creates an optimized UI flow specifically for that task.

Luckily my org has a bit of a pushback attitude towards AI systems, but it will only be a matter of time before we have to compete and adapt. It's kind of depressing, and only the strong will survive.

yes, it will enable a lot of custom one-off software but I think people are forgetting the advantages of multiple copied instances, which is what enabled software to be so successful in the first place.

Mass production of the same piece of software creates standards, every word processor uses the same format and displays it the same way.

Every date library you import will calculate two months from now the same way, therefore this is code you don't have to constantly double check in your debug sessions.

Linux costs $0. Creating a linux clone compatible with your hardware from the hardware spec sheets with an AI for complicated hardware would cost thousands to millions of dollars in tokens, and you'd end up with something that works worse than linux (or more likely something that doesn't even boot).

Even if the price falls by a thousand fold, why would you spend thousands of dollars on tokens to develop an OS when there's already one you can use?

Even if software becomes cheaper to write, it's not free, and there's a lot of software (especially libraries) out there which is free.

> Even if the price falls by a thousand fold, why would you spend thousands of dollars on tokens to develop an OS when there's already one you can use?

Why do you assume token price will only fall a thousand fold? I'm pretty sure tokens have fallen by more than that in the last few years already - at least if we're speaking about like-for-like intelligence.

I suspect AI token costs will fall exponentially over the next decade or two. Like Dennard scaling / Moore's law has for CPUs over the last 40 years. Especially given the amount of investment being poured into LLMs at the moment. Essentially the entire computing hardware industry is retooling to manufacture AI clusters.

If it costs you $1-$10 in tokens to get the AI to make a bespoke operating system for your embedded hardware, people will absolutely do it. Especially if it frees them up from supply chain attacks. Linux is free, but linux isn't well optimized for embedded systems. I think my electric piano runs linux internally. It takes 10 seconds to boot. Boo to that.

The hardware required to run these things has all ballooned in price, there are no efficiencies coming. To run Kimi2.5 4bit you're sitll spending 100k in hardware, and its not nearly as reliable as Claude. Also Agentic Tooling have made their token consumption go up to increase revenue, and models are becoming more verbose in their output (wonder why). You're smoking something.

I said like for like. You can't compare GPT5.2 tokens with GPT3.5 tokens. They're different products.

You can run local AI models today which can compete with early chatgpt releases for a fraction of what those models cost to use at the time. Thats the claim I'm making.

I think the kind of software that everybody needs (think Slack or Jira) is at the greatest risk, as everybody will want to compete in those fields, which will drive margins to 0 (and that's a good thing for customers)! However, I think small businesses pandering to specific user groups will still be viable.

The model owner can just withhold access and build all the businesses themselves.

Financial capital used to need labor capital. It doesn't anymore.

We're entering into scary territory. I would feel much better if this were all open source, but of course it isn't.

Before its too late for what? For OpenAI and Claude to privatise their models and restrict (or massively jack up the prices) for their APIs?

The genie is already out of the bottle. The transformers paper was public. The US has OpenAI, Anthropic, Grok, Google and Meta all making foundation models. China has Deepseek. And Huggingface is awash with smaller models you can run at home. Training and running your own models is really easy.

Monopolistic rent seeking over this technology is - for now - more or less impossible. It would simply be too difficult & expensive for one player to gobble up all their competitors, across multiple continents. And if they tried, I'm sure investors will happily back a new company to fight back.

The only existential threat to the model owner is everyone being a model owner, and I suspect that's the main reason why all the world's memory supply is sitting in a warehouse, unused.

It would be cool if I can brew hardware at home by getting AI to design and 3D print circuit boards with bespoke software. Alas, we are constrained by physics. At the moment.

Aggregation. Platforms that provide visibility, influence, reach.