Ask Slashdot: Can AI Bots Really Write Most of Your Code? 178
Tablizer (Slashdot reader #95,088) writes:
A recent Slashdot poll showed roughly 15% said bots write most their code. But in my experience, I generally find bots don't and can't understand very specific domains. The rules often come from institutional habits and/or arbitrary-seeming government policies and regulations.
So what exactly are the bots doing for you?
If there are regular patterns they're milking, why can't those patterns just be factored into modules and functions? So I'm skeptical of "AI writes most of my code". Something smells off!
The real question might be whether bots can still write your code if there's domain-specific rules and habits. So share your own thoughts and experiences in the comments.
Can AI bots really write most of your code?
So what exactly are the bots doing for you?
If there are regular patterns they're milking, why can't those patterns just be factored into modules and functions? So I'm skeptical of "AI writes most of my code". Something smells off!
The real question might be whether bots can still write your code if there's domain-specific rules and habits. So share your own thoughts and experiences in the comments.
Can AI bots really write most of your code?
Additional Question (Score:2)
Re: Additional Question (Score:2)
Re: Additional Question (Score:3, Interesting)
I've had a long professional development career, and the 2026 frontier models do most coding and debugging work far faster and better than me. I wish like hell that LLMs could be un-invented, but to be an effective person, one must deal with reality.
Re: (Score:2)
Re: (Score:2)
Perhaps you should try it, before giving utterly uneducated shit comments.
Re: (Score:3)
If Anthropic+OpenAI close shop tomorrow, I'll be finally able to afford the hardware to wholly replace them with local models.
If I could buy 5 more 3090s for the $600 a pop I paid for my single one back when, I'd never touch a cloud hosted provider.
Sadly, currently, venture capital is all concentrated in pricing me out of local hardware while/by subsidizing the cost of the cloud LLMs. The moment venture capital stops buying up all the compute in the world to sell to me at a loss through a cloud provider, it
Re: (Score:2)
> 3090
Not that future proof. In a few years you will be glad to have a Blackwell card with BF16, FP8 and FP4. And a few years later you may want to have a card with the features the next generation cards have. The 3090 will still work, but in comparision it will be inefficient.
Re: (Score:2)
if both Anthropic + OpenAI close shop tomorrow
That is extremely unlikely, but if it happens, people will switch to other AI engines.
I use Cursor for coding, which SpaceX now owns.
SpaceX (which may merge with Tesla soon) has deep pockets and an in-house need for AI, so it is even less likely to "close shop".
Even if Anthropic and OpenAI get in deep financial doo-doo, they won't close entirely. They might stop updating their models, but the existing models in static mode would be fine for coding. More likely, they would be acquired in a consolidation.
Ther
Re: (Score:2)
Google
Re: (Score:2)
Even smaller models compete with them now. You need more instructions but the code quality is great. The difference between a 27B model and a 2T model is, that you tell the 2T model "Create a facebook clone" and the next day you have your social network, while the 27B model needs instructions like "Now let's design the user's homepage with the list of friends and a reminder which friends soon have birthday". So Anthropic+OpenAI enshittification may get you to do some more work, but won't kill of AI coding.
Be the machine! (Score:2)
I like the FP though I wonder if you were going for funny in the "You say that like it would be a bad thing" genre. A whole lot of the code that was written yesterday and today was bad code enabling bad ideas, so no code tomorrow isn't bugging me too much.
However the joke I was looking for involves how much easier it is for humans to learn to think like machines than vice versa. I think that's a bad thing (per the old warning from Sydney Harris) but I also think that's the path of least resistance that too
Re: (Score:2)
This isn't wrong. One of my devs is on a bit of a side quest to see how far he an get with Qwen locally (5070Ti) compared to Claude. He was basically trying to answer the GP's question. Resort so far: it's not as good as the latest frontier model feels like 4.6 instead which is pretty solid. Main thing is it's got a pretty small context, but that's due to the 16GB size of the GPU not the model per-se.
Re: (Score:2)
If you don't need to go 100% local: Let a large model like Qwen-Max write a plan (possibly instruct it to be detailed) and a smaller model implement it.
Re: (Score:2)
There's other factors too. Example: if you don't need fast but you have a specific domain you're working in, you can throw regular RAM and CPU inference at doing things with a defined pattern.
For example, no software engineering manager wants to throw hours of work at refactoring declarative infrastructure-as-code to modern tooling from old shit - it's hours of work to arrive at the same place, but with better capability and supportability; classic tech-debt everyone loves to ignore until it's too late. S
It's complexity related (Score:3, Interesting)
AI can craft a small routine or single command customized to my parameters faster than I can write it myself even if I have it committed to memory.
AI can do slightly more complex things, and so long as I keep it tightly compartmentalized so it's easier to debug, it's fine.
You cannot sit down and ask an AI to write you anything long and complex and expect it to do so successfully without requiring more debugging than writing it yourself.
At least, that's been my experience. I rarely code and never anything particularly complex by the standards of a working pro... and I still hit those limits of AI regardless.
Re:It's complexity related (Score:5, Informative)
It's amazingly good at some stuff and not at others.
It can bash out a CRUD app with a web based UI in basically no time, and really well. Think some boring internal tool like inventory management or something. Would cost a fortune to buy and also suck massively. It would from a project perspective count as long and complex, but from a pure software point of view, it's not complicated stuff. It's well documented. It's also a horrible tech stack.
It's also very good at indexing an unseen codebase and figuring out entry points for things. I've known people, or rather a person, that good, but I ain't.
But I've also seen it make a complete shitshow of threading in a small, contained task. It's also so-so if that at optimization. I set it on making thing go faster. I had the dataset, and specified tests. It found the hotspot and made it faster, which is frankly more than a lot of programmers can do. Most of them will smugly tell you everyone knows e.g. std::unordered_map is slow (and you're implicitly a moron for using it), change everything to their own implementation on a Friday night then bugger off. So yeah better than that, but I let it churn for ages with its agents and it didn't get much further. So I put some easy wins in by hand.
It is quite often wrong and i definitely worse in areas requiring domain expertise. But then again that's the case with your average programmer. It's also a bit weirdly pissy at time, but much less so that many programmers, so that's a net win. It won't flounce off if you tell it to delete a bunch of code it just wrote.
So...
Yeah I use it and quite a lot.
Re: (Score:3)
Most of them will smugly tell you everyone knows e.g. std::unordered_map is slow (and you're implicitly a moron for using it), change everything to their own implementation on a Friday night then bugger off.
It's amazing how few programmers are willing to use a profiler, even if you show them how and why. I don't know what braindead rock our profession collectively crawled out from under, but it's the rock that is allergic to simple tools to solve the problem.
No (Score:2)
Re: (Score:2)
Good for unit tests and comments
LOL wait, what?!??
Yeah, testing is easy, it's okay if it has some rough spots on the edge cases... right?
Re: (Score:2)
Generally speaking, I agree with you. AI is very good at making good looking unit tests that don't actually test the important parts. But as long as you understand this, it still helps you with the sunny-day-tests and obviously if you are specific, it can even sometimes write tests that cause the code to fail. I recommend letting AI to find a bug in the code first and then ask it to write a test that triggers that bug. When executing the test, you will instantly know if the bug was real or not and if the t
Re: (Score:2)
Re: No (Score:2)
I am curious what the AGENT.md file looked like for that full module. What directive did you give? We have several modules that it spun whole cloth at this point.
Yes. (Score:5, Insightful)
Use a paid subscription to a frontier model. Make sure it knows how to run your build and run your tests.
If you're worried, have it write tests. Look at the tests; see if they look like the right tests.
Tell it to ask you questions if it needs to resolve something ambiguous. Have it explain what it thinks it should do to the code. Ask it probing questions if you aren't sure it understands. Ask it about design tradeoffs.
I have yet to find a problem where AI generated code doesn't get me at least 90% of the way there. Multi-domain, multi-language, multi-platform. That includes stuff I was doing a year ago. It's gotten dramatically better in 2026.
If someone has an example of a software engineering problem that AI actually did really badly on, I would like to know the model and see the prompts. I'm not saying it doesn't exist. I'm saying I haven't seen it yet, and I'm really curious.
I've done software development professionally since the 1990s. I started coding on the Apple ][ in 5th grade. I've had AI do the wrong thing, or misunderstand something, or miss something that seemed obvious to me. But on all of those same criteria; I've had human software engineers do far worse, far more often.
Re:Yes. (Score:5, Insightful)
I don't think it's usually about what AI "did really badly on" at this point. It's more about whether the overheads of adding all the guard rails and then reviewing and correcting all the output end up costing at least a similar amount to just writing the code yourself.
For a junior developer, I'd say frontier models passed their standard some time ago. Today's LLMs routinely write code at the level of a competent junior-mid but with the advantage of an inhumanly broad knowledge base.
For a good senior, I'm much less convinced. The frontier models are hit-and-miss when it comes to writing good test suites, which matters a lot when they rely on tests so heavily to reach acceptable quality. IME, they are still mostly miss when it comes to the business insight that seniors will take into account and design/architecture decisions that keep the code manageable as it evolves. While they have become much better at recognising existing patterns in the code, and I would say they follow them reasonably well the majority of the time now, there are still bizarre anomalies that creep in, and they aren't very good at extending those patterns when necessary.
I think AI advocates also tend to underestimate just how much time it takes to properly review what they are doing. Take a half-day Claude session to implement some new feature. It might take half an hour to give it the initial brief and then review its plan. Make a few edits and it'll propose a new plan that is also long and detailed and has no edit markings to show what it changed so you have to review the whole thing, carefully and in detail, again to pick up the minor changes it also made for no obvious reason. Then you let it start doing its thing. You can't really leave it unattended for long because it'll stop to ask questions. Some of those questions might be whether it can run a screenful of ad-hoc Bash script to make some edits to your source files, because for some reason it decided not to show the actual edits as it usually would, so now you have to figure out what all the sed incantations are actually going to do while looking at those files in another editor or just trust it and then review all the changes properly later anyway. Sometimes it decides to change its plan mid-flight, but not always in a good way, and if you interrupt it to apply corrections or redirect, you break the neat plan flow it was in. Then you get to the end and you can finally see how everything it has done fits together, and you still need to review everything with the same care as you would with a junior developer's code and probably apply a few fixes where it's come up with weird names or not quite followed a pattern or imagined some instructions that were never there and made other changes in other parts of the system that should be out of scope.
If you read to the end of that last paragraph, you know how I typically feel at the end of a half day working with an agentic UI. And the thing is, I do those sessions as experiments fairly often to keep up with progress, but a lot of what I do is still written by hand, which gives me simple, objective data to compare the two ways of working and see whether the AI tools are really making me more productive once you take everything into account.
Currently, based on objective data comparing similar fully completed development tasks like for like, I still outperform agentic workflows using frontier models by about 25-50% on average, and I don't have to pay anyone extra fees to do that, and subjectively I also find it more engaging to work that way. So I don't really buy the hype about AIs making us all orders of magnitude more productive unless you're happy to YOLO everything forever and take the much-discussed risks that comes with.
Personally, I think the optimum use of LLMs right now is probably not to use them for the whole coding workflow like that or rely on agentic tools, but to use them for tasks like answering quick questions about the existing code, doing research like a fancy search
Re: (Score:3)
For a junior developer, I'd say frontier models passed their standard some time ago.
No, I disagree, and now that I'm thinking about it more, quite strongly. I think frontier models passed the mediocre programmers of all seniority levels some time ago. I've worked with so many arrogant pillocks who are convinced they are the shit and will not listen to someone with domain expertise and then get pissy when they turn out to be wrong. Or worse, get political and push through an objectively broken thing.
Good jun
Re: (Score:2)
Yes, it does look like we broadly agree here.
I suppose the main thing I'd have said very differently is the terminology, in that I wouldn't regard the kind of incompetent senior as a senior in the first place. But then I'm pretty old school about this: to me, a junior developer needs quite close supervision to produce useful results, a regular/mid developer can perform routine tasks with minimal supervision but might need more help with larger or more complex tasks, while a senior is someone you can give mo
Re: (Score:2)
I suppose the main thing I'd have said very differently is the terminology, in that I wouldn't regard the kind of incompetent senior as a senior in the first place
The best phrase I heard (I think here) was "10 years experience or 1 year experience repeated 10 times".
But then I'm pretty old school about this: to me, a junior developer needs quite close supervision to produce useful results, a regular/mid developer can perform routine tasks with minimal supervision but might need more help with larger or more
Re: Yes. (Score:2)
In my experience, those models actually enable juniors to be as good or better than some seniors.
It's not about how much knowledge you have anymore, it's only about how good you actually are at explaining yourself, reading and reviewing, asking the right questions, following good processes and practices.
I'd even say that a junior who cares about doing the right thing and is curious enough to want to learn and understand is now your best asset precisely thanks to AI.
Re: (Score:2)
Perfectly right.
Talk to the AI tool like you would talk to an experienced programmer: on business level specc.
Most of the time, you do not even have to reiterate and fix errors.
Re: (Score:2)
But you need to get a feel when to be really unambiguous. AI will gladly implement what you asked for and if you used a wrong term it will likely still do what you wrote and not what you meant. But that's an information problem, your junior programmer knows the full stack you use, the AI starts to explore the code base like someone who never saw it before after your prompt. How should it know to interrupt you "You probably meant XY and not XX" before exploring a lot of code doing XY.
Re: (Score:2)
Re: (Score:3)
Agreed. Using Claude Code or GitHub Copilot within your IDE is *way* different than just asking a web AI your coding questions.
In the past year, I've been able to successfully leverage frameworks that I didn't know, through the use of AI. Specifically, Vue.js and Lucene.NET. Now that I've been using AI to generate code in these frameworks for several months, I actually learned how they work, in the process, and could write them without AI, but I wouldn't want to. AI types *way* faster than I can. And it kno
Re: (Score:2)
Who needs an IDE these days? All of these tools were designed to write code, but really all you need is to view code now and sometimes write specifications.
Most of the time a claude code terminal is about right. I've even become productive on a phone, with voice input.The IDEs just give lots of screen real estate to things I don't need, like a project navigator or editor pane.
On the projects which I vibe up, the IDE gets in the way. On those were I need to look at the source, okay, but them vim or emacs do
Re: (Score:2)
Sometimes it is faster to delete the AI code and replace it with three lines that just work than explaining the AI to replace it with the three lines. I had cases where the AI overengineered something with a stateful class with inheritance, what I re-implemented in three functions that get the required data as parameters and return the processed data. The AI happily maintained the code afterward, but its own suggestions were all YAGNI.
Re: (Score:2)
You're right, AI isn't the ideal tool when you just need to update three lines of code. But it excels at implementing end-to-end functionality in one prompt. For example, "Add a new download button to save the data as an Excel file. Use the following columns... Highlight any negative values." and so on. When you can ask it to add an entire new feature in one prompt, end to end, from web UI to the database, that's a real time saver.
Re: (Score:2)
There are several things AI can do better in an IDE.
For example, GitHub Copilot in Visual Studio, offers a "Diagnose this exception" button, whenever an unhandled exception is thrown. When you click it, it looks at the stack, the contents of variables in memory, and other runtime information that simply isn't available when using AI from a console or command prompt. Similarly, if there's a thorny problem, it will, of its own initiative, build a little test program and run it, inspecting what shows up in the
Re: (Score:2)
Admittedly I wasn't using a paid frontier model, but I've tried it a few times with very mixed results.
One attempt was to write some code for a microcontroller to do some tricky stuff with DMA and timers to generate a complex waveform. It failed miserably and couldn't seem to understand my implementation that did work, or debug it when I introduced a subtle bug.
Another attempt involved having to develop a system to automatically crop and rotate images to be upright and remove the border. It managed to remov
Re: (Score:2)
I have yet to find a problem where AI generated code doesn't get me at least 90% of the way there. Multi-domain, multi-language, multi-platform. That includes stuff I was doing a year ago. It's gotten dramatically better in 2026.
Previously you claimed "I wrote a line of code by hand saturday (client blocks AI) and it was weird. I hadn't done it in months." now just 90%. What happened? Did you decide to turn the hyperbole down slightly because nobody was buying your bullshit?
If someone has an example of a software engineering problem that AI actually did really badly on, I would like to know the model and see the prompts. I'm not saying it doesn't exist. I'm saying I haven't seen it yet, and I'm really curious.
I've done software development professionally since the 1990s. I started coding on the Apple ][ in 5th grade. I've had AI do the wrong thing, or misunderstand something, or miss something that seemed obvious to me. But on all of those same criteria; I've had human software engineers do far worse, far more often.
This from the same handle that previously said "A rough estimate is that the best LLMs do software engineering about as well as IQ 150 software engineers do it. That makes them better at it than the vast majority of people who write code."
I have AI spit out c
They can, but you can never trust it (Score:5, Interesting)
I've had many moments where I used an LLM to make something from scratch, and have been extremely impressed in how good the implementation was and how much time I spared at the (likely still very underpriced, I am expecting a rugpull there) expense it cost me.
The issue is, you do this three, four times and get reasonably functioning implementations. Often ugly, but whatever. I have use for a lot of temporary tools that really serve to explore something in a mixed analysis-prototype kinda regimen, that makes even un-maintainable proof of concept kinda thing that I can hammer out in 15 minutes of writing the spec and waiting for an hour to get a whole working program.
The issue is, the fifth time, the thing shits the bed and misunderstands something fundamental and in such a an incredibly inhumanely weird way that it can poison a knowledge base with a completely wrong conclusion, and if you don't approach the LLM wit a borderline adversarial level of suspition and mistrust, you may never notice, and particularly enabled by the high speed of development you now have thanks to the elements, you end up with it deep in the foundation of your work, potentially poisoning days and days of effort that follow.
And while the overall code quality is improving and the whole thing gets faster, and cheaper, and like, median-decision smarter, the long tail of LLM-exclusive idiot moments persist, and the moment you wholly surrender the codebase to only ever touching it via the LLMs, they'll ruin your day more and more often as time goes on. And before you know it, you're using most of your newfound productivity to chase down these fuckin' quantfart landmines, just to eventually enjoy the ultimate infuriation of the LLM telling you that thing you just found really is an obvious error it should have noticed the first time, acting like it's all good.
Sometimes, I wonder how often LLMs bury a landmine that takes a company down. They made sure there's no accountability. You're always responsible for the code you didn't write. Nobody will ever be able to come after Anthropic with the damages a customer sued them for. So, what happens when nobody is reading the code? What happens when nobody is left in the company that *can* read the code? I think this will become an increasingly frequent issue.
I do concur there's utility to LLMs and I do expect to use them for the rest of my life. I suspect the difference between me and never-LLM-ers is that, I admit it, I never enjoyed coding in the first place. It's just one of those things where, for most of my life, it was the one skill I had any degree of knack on that put bread on the table with minimal effort required on my part. Now instead of coding and hating it, I have to wipe the drool off an LLM's chin, hating it. I don't think it's elevating me as a person. I'm not sure it elevates the projects I work on. I kinda don't care. I've never been passionate about programming. I've been passionate about being able to afford food and living space. Fuck coding, and fuck coding with LLMs.
Re: They can, but you can never trust it (Score:2)
All the trust and verification issues you mention are issues with human developers as well. Even "accountability" for software faults doesn't really stick to individual developers unless it's egregious and frequent... bugs are expected. The only real differences on the software quality front are scale: the LLM can write more code more quickly, but it can also review and test more coffee more quickly.
I do agree that many orgs will fall into the trap of writing unmaintainable software simply because they don'
Re: (Score:2)
I tested more coffee in the last month than AI will ever taste.
A large part of this is how you approach it (Score:2)
- Have clear, specific requirements (with no conflicts/confounds)
- Use a more expensive frontier model to plan the workflow in discreet units (context rot). You should target using at most half the context for a session without having to compact.
- Use capable models to handle the discreet units with TDD (test driven development)
- Do adverserial reviews with a different model family that has low sycophancy
Re: (Score:2)
If you have long context, use it. Getting over 80% of long context may be bad, but don't stop at 50%. I rather fill 80% then have the model compact too soon. With each compaction it needs to rediscover facts it already had in context that came short in the summary. If possibly break down the task before so it will never fill the context, but that's not always possible.
Re: (Score:3)
https://slashdot.org/story/26/10/25/027245/ask-slashdot-can-ai-bots-really-write-most-of-your-code#comments
Horseshit.
Just because you haven't doesn't mean nobody can. Break a problem down into manageable sizes, and let it do the heavy lifting on those manageable chunks. You know, just like Agile development basically says to do. And then check the hell out of it, which you should be doing anyway regardless of where the code came from.
I've had plenty of luck using Qwen to get things done, because I don't expect to tell it some high-level goal and then walk away. Learn how the tool works, and then use the damn
Re:They can, but you can never trust it (Score:4, Insightful)
How do you decide to trust code you write?
How do you decide to trust code someone else wrote?
Tests. Design discussions. Code reviews. Code review questions. Design Questions. "What about this problem?" Real validation testing.
I was in a funny situation last week. Two different humans told their claude agents to fix the same bug. Claude came up with two _slightly_ different implementations.
I told _my_ claude agent to code review both implementations - using a new prompt I'm working on to try and automate away a lot of the code review value I was adding manually.
Same model; same bug, two implementations, and the code-reviewer agent found different bugs in both PRs.
We ended up blending the best things from both PRs into the one that shipped.
I am really confident that the code I ended up shipping here is better than what our specific humans would have done without using AI, and 10x-100x faster.
I am spending more of my time telling agents how to think like me, what to look for, what tradeoffs to make, etc. They wont do those things unless I tell them to. They don't always do those things perfectly.
(Neither do the real human software engineers who work for me.)
Re: (Score:2)
So do you have any thoughts about implicit personality variables Claude is hiding within each instantiation?
Re: (Score:2)
I don't.
The creator of the pi harness did say that the claude system prompt changes often enough that he had to build his own tooling to try and improve reliability and repeatability.
I'm not convinced the differences in roles I've observed are claude personality differences as much as humans giving different prompts.
Re: (Score:2)
You shouldn't trust your own code either.
Yep (Score:5, Interesting)
And, professionally, my employer pays for a top-tier subscription that's able to code, run command line tools, build and install firmware, and even connect to a logic analyzer and debug data lines.
Yes, "AI bots" can write most of your code, if you're able to clearly explain what's needed in your planning and action prompts. And, beyond that, AI can examine an existing code base, very clearly explain its function, and provide recommendations for refactors and other improvements.
At this point, it hasn't replaced me as an engineer, but it's definitely augmented my abilities it in a way that "paired programming" and other stupid Agile-touchy-feely methodologies never could.
Re: (Score:2)
> I will never need to learn a new programming language again. This lil' bastard does everthing!
This may be dangerous. Assume the Chinese people stop releasing models and the US companies follow. Then you have models that keep telling you to use 2025s libraries. Before we stop depending on company releases, we need better mechanisms to feed a model with documentation so the neural net only needs to bring the intelligence while the knowledge is upgradable without training for weeks to month on GPU cluster
Re: (Score:2)
Your inability (Score:2, Flamebait)
Your inability to frame it correctly is the problem.
They can do a better jon than most people (Score:2)
They absolutely can write your code. And you can teach it your code's quirks and customs in an AGENTS.md file. And it can absolutely show you problems in your own code you never knew was there. If you look at the lengths that openai's agents went to in order to cheat on the huggingface test, realize they know software better than people do. Their whole world is words. Sequences of words. Songs, poems, programs, recipes...word sequences. Frontiers models are experts in all domains at this point. Chemistry, m
Re: (Score:2)
They can do a better jon than most people
Never been to Kabukicho, Japan?
Re: (Score:3, Insightful)
Frontiers models are experts in all domains at this point. Chemistry, math, medicine, biology, legal, creative writing, programming, philosophy - you name it.
That is wildly optimistic IME. Frontier models do a great job of sounding authoritative on any given subject. This is not the same as actually having expert-level knowledge and understanding, and the choices the models make when you ask them to implement something can and often do reflect that. There are several domains I've worked in over the past few years where AIs lack even basic knowledge an expert practitioner would have, probably because these fields are too specialised to have a lot of books availab
Re: (Score:2)
Yes (Score:3)
Re: (Score:2)
So what you are saying is that testing is still important.
When did it stop being important with meatspace coders? Why would you not critically test things that are critical, regardless of code source?
Re: Yes (Score:2)
Like Science Fiction (Score:2)
If there are regular patterns they're milking, why can't those patterns just be factored into modules and functions?
It's because they are doing more than simply assembling template code. As you say, that would indeed be achievable with modules and functions. But AI generated code is doing something a lot more complex.
There are people who dismiss AI as a "glorified auto-complete". To me, that just betrays a lack of knowledge about the AI of the past 5 years. It is like saying "The Large Hadron Collider is just a glorified slingshot", or "humans are just glorified protein factories". It is an application of neural networks
The 70s : Can computer do all the calculations? (Score:2)
Thousands of people were running calculations by hand or primitive mechanical calculators in the 60s and 70s.
They were wondering the same thing about computers in the 70s.
Re: (Score:2)
I doubt that. They may have wondered if it was *worth* the price of the computers at the time, but they were not debating the capabilities.
That facet certainly exists here about token pricing and how it will evolve and in a larger sense about whether it is worth the footprint, and there's every possibility that those concerns will ease with time like they did for computers doing calculations.
The issue here is that folks are treating software development as a monolithic thing. Even very milquetoast softwar
It takes me longer to read bot-generated code (Score:2)
than to write my own code most of the time, so no. I usually use it for small helper functions that can be checked at a glance and boilerplate which I mostly trust to get right without a ton of checking, but otherwise I don't see the value everyone else is seeing. It might be a years-of-experience thing, or it might be nobody using botcode actually understands their codebases, I'm not sure which is more likely.
Re: (Score:3)
One of the most obnoxious things about AI advocates is that they are awfully opinionated on my specific work even if they don't know what my specific work even is. So they see it working well in some areas and when I express a nuanced experience that isn't a slam dunk except in certain situations, they are disdainful that I am not all-in.
One concrete example, some guy who has never written or managed code (not even vibe coding), asked if I used git and I said sure, and he informed me that I should change t
Re: (Score:2)
There are three interesting points to it. First, you can (and maybe should) prompt how (much) it should comment. You may need to refine your style prompts a few times depending on how common the desired style is. Second, if you get other's code and it is hard to read because of AI (or without AI), you can just ask the model to find out what you want to know. These things got damn good in tracing program flow. And third an interesting research thing: Recently someone tried (non-scientifically) to benchmark A
It won't be "my code" (Score:2)
They can certainly write code, but the code they write will be "their code", and I will at best be reviewing it for them. It won't be "my code" because I didn't write it.
Hell yes they write my code. (Score:4, Funny)
Lately I've been having them click on fire hydrants for me too. You know, to prove I'm not a bot.
The answer to this question has many facets. (Score:2)
The answer to this has many facets. For work I am mandated to use AI. The most important thing to remember about AI is that it is a probably machine, and your job is to game it like a gambler.
AI works the best on commonly solved problems, often using the stuff that is scraped illicitly from the web. You will be able to make a basic website in a snap because there is enough probability. You then try to do something unique and the AI falls apart. As an example, for work I started implementing a chat agent
Yes (Score:4, Interesting)
All of it. Have not written a line in high-security, safety conscious, real-time precision application since February. Started off with small functions one by one. Now it is designing multiple parallel ops. I and my team review everything. Yes, we find bugs, and LLM fixes them. Total cycle time cut from days to minutes. Never going back. Expect the end of professional human software development before I retire. I am learning to transition to systems engineering.
code writing (Score:2)
our team has broadly switched over for 90% of all code writing. all small changes, all minor updates, and all exploratory work is done via AI 100%. When we get to production deployed code, its up and down (mostly up) that AI can write solid, working code, but the more complex the task, the more important measure is who prompts the AI.
Detailed, longwinded prompts that really restrict and direct the coding task precisely seems to allow nearly no human coding to be done. But that requires our best software
Most code was not handwritten even before LLMs. (Score:2)
Claude and Kimi (Score:2)
Claude is the good one, Kimi the open-source fallback.
Are we losing a skill here? (Score:2)
If everyone professionally turns into a code reviewer and no one writes code of their own, are we losing the art? Software engineering is one of those cases where the best people to review something are those with deep enough understanding to do it themselves. If we lose the skill of the fundamentals because no one does that work anymore, how are we going to review and fix things?
Writing it matters very little (Score:2)
Understanding it, creating it and being able to debug and maintain it matter a lot. LLMs do not understand anything, they mimic mindlessly and you might get statistically lucky.
NO!!! (Score:2)
I develop - among other tasks - multidimensional signal processing algorithms for a F500 company.
Since part of my work is about studying technical papers, and (if the algorithms makes sense to me) converting the equations into python/C++ code, I wanted to test Claude & c. on this.
Results were incredibly laughable, and most of them resulted into a code that was doing the wrong thing. Game over for me.
Re: (Score:2)
How long ago was that? Claude Code and Copilot get significantly better by the month.
My initial experience was the same as yours. AI wrote awful code, and it never worked the first time. These days, times have changed. AI gets the code write most of the time, even surpassing the correctness of my own code. Yes, I had to spell out some coding boundaries to make sure it followed the standards I consider important (such as API-first design). But once those were spelled out in Claude.MD or copilot-instructions.
Re: (Score:2)
Last time I tried Claude bandwagon was less than a month ago. The problem is not checking if the code generated is conforming to some standard, but rather that AI has problems understanding the algorithms, and I guess that, at least in my field, things will get worse, given the amount of AI-written papers that are being published.
Just an example...I am referee for three major scientific journals. I recently rejected a paper full of wrong reasoning, inconsistent data, and it was clearly AI generated by someo
Re: (Score:2)
It's true, AI is better in some fields than others.
Even in coding, it's better at some languages than others. For example, Claude does great coding a Vue.js website, but sucks at jQuery and WPF. It can do all three, but in the latter two, makes a lot more mistakes.
Almost all of my code (Score:2)
If you ask a bare LLM to do stuff you'll get a mess. I run Claude code with a multipart harness that compels the model to treat the code as the mathematical thing it is, not human prose. LSP Enforcement Kit keeps low capability tools like grep out of the mix, Abstract Syntax Tree things handle the code as its meant to be understood, so that's CodeSight, OptiVault, Serena. About a year ago I gave Antigravity the boot after their awful 2.0 release and went back to JetBrains IDEs, which now offer code structur
Dream Utilities are no longer 'some day' (Score:2)
where do we get the experience to oversee? (Score:2)
Re: (Score:2)
Humans can learn to oversee code even if they weren't initially experienced in code-writing.
A year ago, I realized I needed to upgrade my website to a real front end framework. I knew only jQuery, but it was blocking me at every turn. I ended up choosing Vue.js. Even though I knew nothing about it, I watched what AI did, and through that process, learned how to write code in Vue.js. Now, I could write it by hand, and sometimes do so. Essentially, AI *taught* me how to write Vue.js, and yes, I very much can
Re: (Score:2)
"every successful story i've heard involves a human with real-world experience using that experience to ensure ai doesn't get it wrong. what about when the real world experience has all left"
Sure does. And every word of those interactions goes through and is retained by the vendors building the models. It's the 'train your replacement' system. This is why they are pushing staff to use AI so hard because that's how they get you to train the next generation of AI.
AI cannot write MY code, but... (Score:2)
I'm retired from software development, so I only work on hobby projects. I don't use AI by choice because it's a hobby and I enjoy all facets of software development. Getting an AI to do it would be completely contrary to my enjoyment of the hobby.
I have a friend still working in the field, though, and he says he hasn't hand-written code in over 6 months. He tells me the current models are really good and produce code, test-cases and documentation that's good enough for his company.
My friend has decade
Re: (Score:2)
I'm like your friend, not quite retired. I too have hobby projects, and these days, exclusively use AI to write my code. I never want to go back to hand-written code. AI types *so* much faster than I can! And the code it writes these days, is excellent. It follows good coding practices *and* my own preferences, which it figured out by looking at my code. AI has literally taught me how to write code in technologies I wasn't familiar with, like Vue.js and Lucene. But I don't want to write it myself, I'd much
Re: (Score:2)
I do like the craftsmanship of hand-written software. And being retired, I'm in no hurry, so a faster build/review/test cycle doesn't matter to me.
I would use the power tools except that I think the entire GenAI industry is unethical [skoll.ca], so for that reason, I don't use the power tools.
Re: (Score:2)
As a person with first-hand experience using generative AI for coding, I take issue with several statements in your article.
1. AI, when used well, does indeed make programmers more productive. The fact that those who are "forced" to use AI dispute this, is taking as gospel the word of those who are already biased against the tools. I also question the statistics about failure rates, given my own track record of AI success at the end of a 37-year programming career.
2. I am not greedy, I'm having fun. I love
In social sciences, nothing is good nor bad. (Score:2)
He was complaining. Why? His company forced him to use claude to write code. (Take it or leave it). So he did. Complete new flow. Put in the effort and improved it to the best of his abilities.
His conclusion? The code works, efficiency is not great, but ok. The code often is hard to review because it is very ve
Depends on a number of factors... (Score:2)
Broadly speaking, I find agentic coding to be even more exhausting (and expensive) than AI augmented completion. It's babysitting a mechanism that can screw up at any time, so it involves a *lot* more exhaustive work to make sure you have a harness to let the model try and retry with objective feedback very quickly. A continuing issue being that it fails to recognize an inability to do something but tries to do something anyway, which means any result that is not objectively measured in strict terms is at
Weak Results (Score:3)
Many of my niche attempts seem to be "unsolvable" with AI. Even developing MCPs where APIs don't already exist seems beyond its capability.
I find it very frustrating to encounter so many failures when people online are claiming to be living life on full autopilot. Few are willing to show me their prompts. But, in most cases where I have been shown prompts for claims that 'I just told it to write an MP3 converter and it immediately gave me BetterHandBrake', I find that there was actually a long series of prompts, re-prompts, failures. Additionally, while the GUI looks "nice", it is the same style as Claude and everything is designs, the product works rather poorly and could likely be replaced with a couple of CLI commands.
The drumbeat of success versus my troubles is starting to make me question my capabilities. But, I'm very seriously wondering if I'm not up against bots promoting the bubble. I definitely feel like I'm being gaslighted.
AI writing most of the code?? (Score:2)
I'd argue that only expert programmers and advanced specialist still have a measurable edge.
As I am neither am expert nor a specialist programmer the AI beats me hands down for speed, quality, reliability, security, efficiency BUT NOT when it comes to cost. I do not have to pay for my own time...
These days senior programmers can manage agentic teams except when they tell their juniors what to do and what not to do they liste
AI can write code (Score:2)
But I tell it what to write.
I have a hard time understanding the "I read the diff in the end" people, but think it is very useful to give mid-level instructions like "Now add unit tests for the invariant", so I know what and why, but the bot writes code faster than I can type.
After decades of writing code... (Score:2)
AI doesn't write *most* of my code, it writes *all* my code. I used to try to cling to the illusion by using the integrate with an IDE helper format but have since ditched it. That doesn't mean someone without my decades of experience would get the same results. Is the code ugly? Sure, but no human needs to read it so who cares? I still approve the overall design and spec, the detailed 'what does this code do' and overall implementation scheme either come from me or are a approved by me.
I don't see a single
Re: (Score:2)
I don't see a single line of any of that code let alone write it. But every line goes through a full functional testing framework including vulnerability/bug scanning, aligns to a coding style spec + lint, and has additional checks and filters I've built for special consideration of vulnerabilities and injections at the seams. Claude even kicked off and builds infrastructure for backends and manages infrastructure as code.
I would be fired and sued into oblivion if I did this.
Nope (Score:2)
It also takes a tremendous amount of effort to learn how to correctly prompt and again well, at that point why don't I just write it myself?
The programming isn't what takes forever it's getting all the business logic in place and explaining it
AI can be useful for writing code... (Score:2)
AI can definitely be useful for writing code but you need to give it very clear prompts and properly verify that the code it wrote is correct and accurate.
As for my own code, considering that the stuff I am working on at this point (personal projects, not paid work) relies on a bunch of reference materials where feeding that stuff into an AI would probably violate various NDAs 6 ways from Sunday, the answer is no, an AI bot probably can't write the code I am currently working on.
Not even close... (Score:2)
Let's start with some info:
There are 2 sets of rules at the USPS. Those posted on their website and those that exist behind the scenes. There are very specific requirements to complay with their Postal Form 3541 for in county, in area, in state, and out fo state subsections.
Each zip code has to be seperate, except those going out of state, that can be one big bundle (or multiple if there are enough to cross the bundle limit). Each Carrier Route has to be a seperate bundle o
Re: (Score:2)
The thing that smells off to me is that Slashdot itself mangles Unicode every time I paste it into a comment box. So I'm reluctant to accept the implicit premise that software written entirely by humans has already conquered all the easy problems and that AI-generated code must therefore clear some uniquely high bar before we acknowledge it may be useful.
If Slashdot code ever gets open-sourced, I'll fix the Unicode.
Re: (Score:2)
Don't overpromise. Unicode is fucking annoying.
Re: (Score:2)
You can actually control the overbuilding tendencies, but providing instructions in places like Claude.md and copilot-instructions.md, or by correcting the over-building through prompts.