Ask HN: Do you know of any company that went back to hand-written code?
62
aantipurist about 3 hours ago 87 comments
RU version is available. Content is displayed in original English for accuracy.
This is intended as a question about the current phase of the LLM hype cycle, and, at the same time, as a reality check about whether reluctance to rely on LLMs is economically viable for an average developer.
You've definitely seen that many companies progressed from "use LLMs for anything and everything" policies to limiting budgets and scopes of approved LLM usage. It wouldn't be unexpected to see that some companies might decide that generating more code faster doesn't meaningfully contribute to their bottom line, and forego LLMs in favour of human authoring, but does it ever happen?
I'm not trying to point out any (de)merits of the new way of software development, I simply want to know if there are many/any companies that went back to the old ways.
Discussion (87 Comments)Read Original on HackerNews
Having said that, for all our startups “interesting” code I use no AI, not even tab complete. For a few reasons:
- having derived the code and built a mental model is something like 90% of the work, the code artefact being 10%. (You can probably verify this yourself, if you’ve implemented something once and then someone deleted it all, you could rewrite it MUCH faster the second time).
- once I have this model / vision in my mind, I’m streaming it from my mind into reality via code and having code suggestions pop up breaks that flow state for me.
- using Claude code even on throwaway scripts can mask problems. Why is doing this thing so difficult that it needs an AI to write the code? And usually when I’m writing those scripts I have a thought or insight about the system. in other words it’s nice to have that low stakes time.
- models are still frustratingly bad at spatial reasoning and I don’t see this improving anytime soon, which is a problem for a lot of graphics code.
- the code models produce still isn’t what I’d consider great code; subjective of course but it’s our product and we think having great code in our core technology is well worth it in the long term!
I do still use Claude code on various commoditised pieces (account UI, etc) and I use Claude as a great search engine, though its explanations of complex topics still kind of sucks and a human blog is almost always better.
We are a small company of 15 engs.
You have evidence video here, that the AWS AI Division (04:23) went literally full vibe coding:
"I was laid off from my AI software engineering job at Amazon" - https://youtu.be/ZAom6f4-vh0?t=263
The whole division has been dismantled...
After figuring out the exact product we need with some very quick development cycles with AI, we now have some problems that I can't keep under control because of AI. So, currently under consideration to rewrite the core functionality without AI so we can keep it simple, understandable and slower to change.
The interesting part is that AI could do this, but managing developers with AI has become very difficult to get them to slow down and build stable simple things.
EDIT: I found it interesting how negative the comments were and I thought it would be easier to add context to the root.
My developers wanted to use AI. I was a skeptic for several reasons. But I agreed that it was worth committing to and seeing what came of it. And all in all, the rapid prototyping is excellent.
However, as parts of the software are past the point of rapid prototyping, I talked with my team and everyone agreed that the core piece deserved a re-factor/re-write because of the mess it was. Classic tech debt of deprecated code hanging around and prototype features with patchwork additions that now are a poor way of executing the intended function.
So, we agreed to do a significant re-factor and did. However, quickly it got incredibly messy again. Partially due to AI using git to look at prior commits and add things back.
And so, it was essentially an unsuccessful refactor and we find ourselves with the same issue. They also agree that using AI is the problem here because it's very easy to get new things. And the temptation to solve a problem quickly is high (I am stressing the opposite, to go slow), but it can be hard to say no to a co-worker who isn't dealing with the technical issues.
So, both myself and my developers agree that AI is the issue in creating a messy that has resulted in unstable code. The reason I say that I am considering this choice is because while they acknowledge the issues it's causing, they aren't excited about returning to coding without them. Which is also why I'm considering the right move here, since the goal is both a happy team and well functioning software.
What many businesses are going through is the industrialization of software development.
There definitely going to be a lot of fortune 500 companies absolutely wrecked by the amount of changes to their code base that no one will be "responsible" for because none of the workers will really know if it was their agent that changed something.
There could be some outliers who found the correct tools at the right time, but I don't think they'll suddenly be thrust into the money, because good code is definitely not the only thing that lets people succeed.
That is mutually incoherent. If AI can reliably do it, developers who reliably did it before AI would reliably did it again.
Sounds like AI cant reliably do it and your developers struggle to control its outputs.
What you've said also makes clear why nobody would have an incentive to play like a team player. You have this comprehension work you think is important and you've already devalued it by saying that humans would be a waste if AI could be used.
If I were a human working for you I would conclude that regardless of your exact words, your actions create the conditions whereby it is far safer to try to get the AI to bullshit its way through things than to risk trying to do them well as a human. You clearly articulate here that the bullshit is all you care for. If that is what will satisfy you most readily, why would anyone working for you aim higher anymore?
Try “I am considering”.
Hacker News has long ago decided that making comments on the quality of the writing in the comments is not welcome. Correcting spelling, correcting word usage, changing phrasing, and making other suggestions to improve the grammar of a post is not something that Hacker News readers want to encounter in the comments section.
It bothers me every day because there are gobs of cases where the writing is horrible, but I understand why the rule exists. It would make the comments section insufferable if everybody were just correcting each other's grammar the whole time.
Solve the leadership problem and the AI thing becomes an advantage instead of a problem.
So did Commonwealth Bank of Australia and IBM: https://www.cnbc.com/2026/07/01/employers-who-laid-off-worke...
And I'm sure many others who didn't publicize it because they have to keep this circular economy going.
Please read your sources completely before posting them only reading the title
> COO Kumar Galhotra said Ford had been over-relying on automated quality systems without getting results, per Bloomberg. The returning engineers rebuilt the data pipelines feeding Ford's AI training, mentored junior staff, and reprogrammed the automated systems they had originally been brought in to replace.
Similarly with Commonwealth Bank and IBM, which are cited in your second link. None of these companies are saying they're not going to use AI and they're going to go back to manual labor. What they are saying is that they laid off workers prematurely.
The big lesson is highlighted well by IBM's head of HR, which is you can use AI, but you have to continue to invest in humans:
> “If we don’t continue to invest in entry-level hires, what happens in three-five years?,” IBM’s chief human resources officer, Nickle LaMoreaux, said at a Charter AI Summit in New York. “There’s no pipeline; the well simply dries up,” LaMoreaux added.
I get what you mean, but in general circular economy means something different, positive in my opinion. Here is what wikipedia has to say:
> a model of resource production and consumption that involves sharing, leasing, reusing, repairing, refurbishing, and recycling materials and products, to extend product life cycle for as long as possible
In the EU it is a hot topic for a lot of non-profit and social innovation work
I don't understand why people would even entertain that. They are completely fine using random libraries written by random people but not having AI helping with a small refactoring? Nobody is making you use AI to do the entire work.
I see developers using languages with garbage collectors and projects based on heavyweight frameworks while saying "we can't use AI because we still need to understand the code" or "I love coding and AI doesn't allow me to do what I love"
If you are one these people... Dude, you aren't doing the work yourself already. Back in the day memory management was one of the hardest pieces of the programming work. A lot of time went also to building stable and bug free string libraries, math libraries, UI toolkits etc. Without shifting complexity (and tedious work) elsewhere, we wouldn't have been able to build the software ecosystem we have today
I agree that we still haven’t figured out what’s the best way to use AI tools, but over the time it would mature and people would come up with patterns, conventions, and design principles so the quality of what is produced will be maintained or increased while keeping the productivity gains.
I'm glad there is a tool that let's these companies have even shittier code shipped even faster. The faster they burn down the better.
Every single person I talk too that is enthusiastic about AI is so because they can't code for shit. Sure, then the AI seems better. But it never really is. Learn to code, that inertia you are talking about is the technical debt. It will drag you down into the abyss eventually
I use AI coding tools both at work and in my hobby projects. It very clear to me they're nuclear-powered footguns, and we have a long, long way yet in developing practices, structures and workflows that will enable true benefits while minimizing the absolutely toxic baggage/fallout/side effects.
I look at the big corp push for AI adoption as another instance of boards of directors choosing the large investors / their own personal gain over the success of the corporations they govern (first instance being the "back to the office" push in attempt to shore up commercial real estate). I think the circular economy of the AI bubble has spread far and wide, and we're watching a lot of invested players trying to keep it going.
LLMs are great at this. I'd be very surprised if it wasn't possible to improve the actionable notes.
Most human-made doctor notes I've seen have not been comprehensive, anyway, and the doctor has had to spend a good chunk of a rushed appointment scribbling notes rather than interacting with the patient.
Like, did somebody seriously think through the meaning and implication of that disclaimer and still write it?
Assuming of course everything else stays the same (quality, etc.)
Does anyone know any company that went back to hand-written code because it decided drawbacks of ai generated code or some other concerns outweight the output benefits?
I don't see how one can hold "speed of code generation" as a good metric while, simultaneously, holding "number of lines generated in a unit of time" as a bad metric. And more worrisome, I'm not sure how one would reward one without rewarding the other creating a perverse incentive. But I'm open to be proven wrong.
LLMs can be very useful for all sorts of things, but large scale code generation is IMHO the least interesting use case. And even when LLMs are successfully used for code generation, I feel like progress has been reset to the early 60s and people are now rediscovering all the failed software development approaches (eg "spec-driven" is pretty much the equivalent of "waterfall", I'm now waiting for UML diagrams to make a comeback as the next big thing with an "agentic engineering" label slapped on ;)
Classic interactive "vibecoding" with quick turnaround times (but much quicker than now please, don't make me wait and let me slip out of the flow) might actually turn out to be the most useful way to build software with LLMs (or let's better say "prototypes"). Because everything else currently looks like we're building up too much bureaucracy around the software development process again, and just in time when we finally got rid of that shit (now we have that absurd amount of .md files in pseudo-human-language as "skills", "rules", "context", "memory", ...) like it was common in the 70s when the work was split between "software architects" who only do the high level design, and "implementers" who only bake that design into code. This was obviously a stupid idea and I don't know why the "AI bros" seem to be so keen on repeating that mistake.
Every approach that builds a "human language specification" that's separate from source code written in a much more precise programming language is doomed to fail (eg the code is the spec!), and that's nothing new, we've known this for decades, but that brain virus of "upfront software architecture" always keeps creeping back into the minds of people at the first opportunity.
The main things I don't like about LLM is that frontier models are paid (so no unlimited "tokens"), subscribtion based payments, local models are "not ready yet".
Where did anyone advocate for "large scale code generation"? LLMs are a fantastic way to go from, "We thought this this feature" to "It's shipped and in people's hands". That could've been 100 lines or 1,000 lines but that's not the point.
The last time I've checked ai still had an edge adding new features, but the team collectively lost the project knowledge and any problem discovered that llm could not fix took significantly longer to correct - they were effectively working on a new to them codebase.
We're working in a really deep area where having full understanding of our code is more important than speed.
Besides, in this area, clients getting one whiff of AI code would be an immediate deal-killer
If you implement AI code review suggestions then you're using AI code, just the hard way.
Also I don't see why your clients would be ok with having AI review and influence your code. Seems hypocritical, or like you're trying to cheat them.
Re "the hard way", I disagree. While it may seem like the hard way, in the medium and long term this is turning out to be the easy way. We have basically no technical debt, we can fix issues rapidly and (intangible I know) we can stand behind our codebase with absolute confidence.
Same. But that doesn’t have to be in contrast to using AI tools. There’s a lot of grunt in day to day coding as well.
The problem is not the tools
Pre AI it took 100 engineers 5 years to get into a legacy code situation. Once you’re in a legacy code situation it’s very hard to add new features and your code is full of bugs. Fortunately most old companies with legacy code are making loads of money so they can pay the increased development costs to add features to their legacy code.
Now with AI five engineers can build a legacy codebase in six months.
That doesn't sound like a good thing in the long term.
> no one can deny the productivity gains AI brings to software development.
What are those "productivity gains"? Rapidly building hundreds of the wrong things that people do not want?
It means you cannot stop and the competition only moves far more quicker and its a forever race to the bottom.
Do you know of any company that banned compilers and went back to hand written assembly?
Do you know of any company that banned stack overflow and went back to figuring everything out?
The point GP is making, which I understand seems dismissive, is that AI is a tool that people are not retreating from, because it offers the same sorts of gains as previous tool improvements we've seen in the field: from compilers to IDEs to sites like Stack Overflow. I think adding AI to that group of improvements makes a ton of sense, as it has the capacity to provide similar gains in productivity.