- cross-posted to:
- fuck_ai@lemmy.world
You must log in or register to comment.
An interesting trend is these comments: the worse a code base is, the more helpful AI is for expanding it (without actually fixing the underlying problems like repetitive overly long unexpressive code).
Forward-thinking companies should use AI to transform each developer into a “10x developer,”
Developer + AI ≠ Developer x 10
At best, it means 1.25 x Developer, but in most cases, it will mean 0.5 x Developer. Because AI cannot be trusted to generate safe, reliable code.
Computers are machines designed to quickly, precisely, and consistently make mistakes.
I think 10x is a reasonable long term goal, given continued improvements in models, agentic systems, tooling, and proper use of them.
It’s close already for some use cases, for example understanding a new code base with the help of cursor agent is kind of insane.
We’ve only had these tools for a few years, and I expect software development will be unrecognizable in ten more.
It also depends on the usecase. It likely can help you better at throwing webpages together from zero, but will fall apart once it has to be used to generate code for lesser-discussed things. Someone once tried to solve an OpenGL issue I had with ChatGPT, and first it tried to suggest me using SDL2 or GLFW instead, then it spat out a barely working code that was the same as mine, and still wrong.
A lot of it instead (from what I’ve heard from industry connections) being that the employees are being forced to use AI so hard they’re threatened with firings, so they use most of their tokens to amuse themselves with stuff like rewriting the documentation in a pirate style or Old English. And at the very worst, they’re actually working in constant overtime now, because people were fired, contracts were not extended, etc.
It’s made me a 10x developer.
As someone who transitioned form Junior to Dev as we embraced LLMs. Our company saved that much time that we all got a pay rise with a reduction in hours to boot.
Sick of all this anti LLM rhetoric when it’s a tool to aid you. People out here thinking we just ask ChatGPT and copy and paste. Which isn’t the case at all.
It helps you understand topics much quicker, can review code, read documentation, etc.
My boss is the smartest person I’ve ever met in my life and has an insane cv in the dev and open source world. If he is happy to integrate it in our work then I’m fine with it. After all we run a highly successful business with many high profile clients.
Edit: love the downvotes that don’t explain themselves. Like I’m not earning more money for doing less hours and productivity has increased. Feel like many of the haters of LLMs don’t even work in the bloody industry. 😂
Man, I wish LLMs were more useful to me than line completion tool we already had in normal languages in normal IDEs.
So far everything I’ve seen it do even with agentic approaches, is just not covering my use cases.
At best I can have it generate some correct-ish terraform boilerplate. Or writing mediocre code in languages I have to use once in blue moon, that I still then have to correct. Cursorrules are meh.
Me: fintech, 15y of exp.
On the other hand I can imagine it creating some bullshit boilerplate in companies that require bullshit boilerplate.
Btw I don’t think code throughput is what distinguished Junior from Dev. I rather think it’s realizing the steep decline in “Doner-Kebab” effect :)
I don’t have this experience really. So at home I used it to help me setup my docker based media server, but the key is I’ve done it before and it was to save me from Google hell and reading docs. Works like a charm now.
At work, as I said earlier my boss is like a savant, and did great open source work on BSD back in day. The tool he built for us which uses LLM is more boilerplate based.
So we have conventions as we make enterprise software. So when we add a new model in C# we are going to create the same missions and queries every time. We are also going to use the same structure for the GraphQL layer and again a lot of the same routes for Typescript react. For instance if we have a new model for TransactionType, the we need an admin screen for listing these, editing/creating one and deleting (which shocking for me is just Archived = 1).
So now I add the model and the fields and the OnModelCreating rules. I then hit generate and everything else is done. We also using caching to avoid hitting the LLM again if nothing has changed. This alone saves an hour or so for every model we create. If a project has 33 models that’s over a weeks dev time saved. I just need to run the app to get the schema and copy it over and run GraphQL to generate types. Now don’t get me wrong sometimes it shits the bed and we just run it again. For instance it will regularly just fail on the db table def file(forget the file name now ).
Then there is the simple time saving stuff. Like we use useMemo a hell of a lot. So if we have a Client dropdown and a Campaign drop down and I want to filter the Campaigns based on Campaign.ClientId then sure I can write it every time, I can go grab it from another file. Or I can write a comment like // Filtered clients and boom it’s done and it’s done based no the current project and how we do things.
The same for sorting things, sure I know how to sort arrays but why should I waste brain capacity on this when LLM is more than capable.
Your last point is spot on and although it’s only been two years, this is something I am better at. I look back at old projects like man I was an idiot or naive back then.
The same is true for actually taking a step back when getting data and planning for the futures as on my first project as it went to real world use it started getting progressively slower on certain screens as the data loaded was built with the dev environment in mind with no thought to 1 year down the line. So it’s not like LLMs are hindering my progress in these areas.
I guess what I’m always trying to convey in these posts is that people are not out here asking an LLzm to do X and then just planting that in a project and leave it at that. If I don’t know how to do something, like recently I need to integrate an app with Sage using Intracct SDK and I have no fucking idea, so I’ll get the LLM to show me how do I get Reporting Periods, Nominals, Transactions etc. then I’ll use what I learn in a console app to play around and work on getting and setting data. Then I’m ready to implement into the product. My boss charges £2000-2500 an hour for his time so to not have to spend more time showing me how to do something he could whip up in an afternoon gives him time to work on the higher skill tasks and allows me to learn at a faster rate than I could before. Sure I could read the docs, but that’s what the LLM will do if you point it to them and it’s quicker than me and will frequently point out things I didn’t even consider.
Then if I still don’t know whether I understand what it told me I’m speak with the boss and he’ll immediately say yeah that’s bang on or that’s one way but I would do this and then explain his reasoning. Like I can’t understate that this man is a damn savant and if he is happy to use LLMs as a tool then I have no grounds to object t as this guy is of the level he could build his own just for fun. Like many times I asked him stuff in passing and responses just blow me away and sometimes I just have to accept dude is way smarter than me and I’ll just never get certain concepts.
Like I can’t understate ask something completely obscure about how operating systems are made and I’m getting a full lecture on how to build on os and how they work.
Finally he’s also the nicest person I’ve Denver worked for. We come first work is second. He’s amazing with my neurodivergence and neuroticism, sickness etc. I have PTSD (hyperbole) from previous bosses that it still feels surreal when he always does the right thing.
Edit: if this is effort to read then I’ve added a reply to it with it fixed by an LLM to be more coherent and palatable.
Everything from this
So we have conventions as we make enterprise software
to this
forget the file name now ).
Is a convention based code generation. It’s older than my tenure, so it’s older than 15 years. Having convention-based code generation is something that you do as a part of DRY rule, so somewhere after you’ve noticed that you’ve done it three times. The real boom for it was when the current hype that was supposed to get rid of programming jobs were no-code/low-code solutions.
Or I can write a comment like // Filtered clients
This is either reusable code composition or again convention based code generation.
The same for sorting things, sure I know how to sort arrays but why should I waste brain capacity on this when LLM is more than capable.
This is DRY. Create once
sortBy(x=> x.name)and then import it whenever you need it.. If I don’t know how to do something, like recently I need to integrate an app with Sage using Intracct SDK and I have no fucking idea, so I’ll get the LLM to show me how do I get Reporting Periods, Nominals, Transactions etc. then I’ll use what I learn in a console app to play around and work on getting and setting data. Then I’m ready to implement into the product.
Fair. POC is the only part where I could maybe see some gain in my use cases.
play around […] Then I’m ready to implement into the product.
Be cautious about this approach though. Usually just playing with the thing will make you a beginner with it, so usually it’s a good approach to read the documentation/manual and see where the bodies are buried.
My boss charges £2000-2500 an hour
First class grifter, respect.
Like many times I asked him stuff in passing and responses just blow me away and sometimes I just have to accept dude is way smarter than me and I’ll just never get certain concepts.
To me that’s a sign of a bad teacher. He might be a very good engineer, but not being able to tailor your responses to your junior - to me shows a lack of leadership skills.
Edit: if this is effort to read then I’ve added a reply to it with it fixed by an LLM to be more coherent and palatable
Bejesus, man, I’m not stupid, don’t feed me slop. I care about you, not about the translator dumbing you down to be a palatable to mediocre people.
Painkillers have kicked in now, but I really do want to continue this discussion so I’ll make this as unread and respond tomorrow when I can give your reply the time it deserves and I can be somewhat coherent.
I do apologise for the LLM version, it was in no way because I think you’re stupid and more because I know I have a tendency to digress and in the past that hasn’t been received well so my intention was to have something where the writing style isn’t going to turn people off from the thoughts I’m trying to convey.
Have a nice evening and I’ll hit you with a reply tomorrow. 😊.
Here is a rewrite of this stream of consciousness that an LLM helped with which might be more palatable. I am neurodivergent and stream of consciousness is my bag, writing isn’t. So LLMs have massive utility for here.
I get where you’re coming from, but honestly, that’s not been my experience at all. I’ve used LLMs both at home and at work, and while they’re not magic, they’re absolutely useful—especially if you already know what you’re doing and want to skip the repetitive or boilerplate stuff.
At home, for example, I set up a Docker-based media server. I’ve done it before, but using the LLM meant I didn’t have to dive into docs or endless Google threads. It saved time and frustration.
At work, we use an internal tool powered by an LLM to generate the standard boilerplate we always need in enterprise apps. So when we add a new model in C#, it auto-generates the typical queries, mutations, GraphQL setup, and React admin pages. Just saving that hour or so per model adds up—on a project with 30+ models, that’s a whole week of dev time. And since it caches results unless the model changes, it’s fast and reliable (though yeah, sometimes it messes up the DB table file and we rerun it).
It’s also great for common patterns—like useMemo for filtering dropdowns in React. I can write it manually or copy it from somewhere else, but why not just add a comment like // filtered clients and let it handle it? Same with array sorting. I know how, but my time is better spent on the harder stuff.
And that’s the key point: LLMs free up brain space. They don’t replace learning or experience. I had to integrate with Sage using the Intacct SDK recently—had no clue where to start. The LLM helped me explore the API, figure out how to get Reporting Periods, Nominals, etc., and build a small test app. Once I understood it, then I integrated it into the product.
It accelerates my learning, and my boss—who is ridiculously smart and built this whole system—can spend his time on the harder problems instead of walking me through step-by-step. Even he uses LLMs, and if someone of his calibre is doing that, I see no reason not to.
LLMs aren’t doing the work for me. They’re just making me more efficient, especially on the stuff I already know how to do. And they help me ramp up quicker when learning something new.
I am not anti-AI or something like it and I use AI on a daily basis. If you work on a domain where there’s plenty code written for it or documentation, AI acts like a very efficient search tool. It does not replace traditional documentation or stack overflow, but it significantly reduces the time I take searching for specific syntax, or an example of how to use a library, or how to use a specific feature or parameter of a library. Occasionally it gives me bad advice as well, such as doing something that results in low performance, low security, but then I can check the actual documentation and code to see the details. For code reviews, I think it’s only partially useful, while sometimes it spits something useful, most of the time it spits out bad or irrelevant advice that ends up polluting the code review screen for actual human devs trying to review the code. However, even with all the gains, which is kind of a mixed bag, I think it’s very unlikely AI will increase speed 10 fold. At best, it will be like a 25% improvement at best, and only specific to some times in the project lifecycle, and most of the gains only happen when you are dealing with generating boilerplate code and adding non business-specific functionality. Most of the time I had to maintain existing code, debug existing functionality and fix some security flaws, AI didn’t help me at all.
Developersdevelopersdevelopersdevelopers,developersdevelopersdevelopersdevelopersAIThis is so fucking sad to acknowledge that a lot of people just want to squeeze any profit left in the industry, even though they know AI is a great tool for developers, not a replacement. They must know that because anyone who can access it can replicate the same things, making these products uncompetitive.
AI is a great tool for developers, not a replacement
AI isn’t a great tool for developers. It’s a great tool for mitigating the knowledge gap between an individual’s academic understanding of a development project and the syntax involved in the language they are attempting to deploy.
As the number of programming languages has proliferated faster than the volume of developers versed in each language, and the older languages have lost much of their professional base to retirement and layoffs, we’ve needed increasingly elaborate tools to fill in the skills gaps.
But AI doesn’t fix the underlying problem of an increasingly large backlog of code desperately in need of refactor or replacement. It just papers over the problem with a cheat-sheet of simple conversions that junior developers can leverage to liter the next iteration of the codebase with bandaids.
A proper solution to our coding backlog would be educational first and foremost. We need more rigorously enforced orthodox approaches to coding. We need more backwards compatibility between systems. We need to refine the number of languages in active use and narrow the size and scope of their libraries. We need a more universalist approach to building and maintaining database schemas, digital communications, and business practices. We need a publicly funded open source community of developers to build the backbone of software into the 21st century.
What we’re producing is the opposite of that. Less rigor. Fewer recognizable standards. Less training. Poorer code hygiene and weaker enforcement of best practices. More bugs. So many more bugs. And enormous volumes of legacy code that nobody will be able to maintain - or even understand - in another twenty years.
And it’s intentional. Lay off the workers. Implement AI Slop. Slop does sloppy work. Hire back workers as Temps or Contractors. No benefits. Lower pay.
Like all of Capitalism. It’s a fucking scam. A conjob. A new innovation in fucking over workers. (Ironically the only “innovation” ever directly produced by Capitalism)
I remember when everyone was saying that companies would need programmers and that every kid should learn programming. Now I realize that companies were promoting that idea so they’re be a surplus of programmers competing with each other and companies could underpay and swap out workers quickly.
exactly, “reserve army of labour” is a tale as old as capitalism.
Just that the IT industry has run a very effective propaganda campaign for it
Yeah obviously. Whenever a company says “we can’t get enough X workers” they implicitly mean “at the price we want to pay”.
But that doesn’t mean they were wrong. Programming is still an amazingly well paying and low stress career. Being replaced by AI is a little worrying, but I think by the time AI is good enough to really replace programmers, it will also be able to replace most white collar jobs - HR, finance, etc. - and society will have bigger problems.
I would not market an industry well know for burnouts as “low stress” though.
The games programming industry is high stress, but apart from that it isn’t. I don’t think it’s known for burnouts any more than any other industry.
I don’t think it’s known for burnouts any more than any other industry.
Whoever your project manager is, I hope you have thanked them today. You’re clearly working with some of the least shitty ones.
Since when are contractors lower pay? Companies waste fortunes on them.
That’s because the agency gets half for basically doing nothing
They dont usually have benefits (eg: health insurance) or time off
Even accounting for that (at least in countries with national healthcare), they’re definitely more expensive than regular employees.
You’re right, but I think the main reason companies like it is because it’s easier to get rid of contractors than full-time employees.
Employees are a liability on the balance sheet whereas contractors are just lumped into a project cost.
Healthcare is only a fraction of fringe costs — employer payroll tax, unemployment insurance, facilities and admin…
Even if AI is an actual tool that improves the software development speed of human developers (rather than something that ends up taking away in time spending reviewing, correcting and debugging the AI generated code, the time savings it gives in automatically writing the code), it’s been my experience in almost 30 years of my career as a Software Engineer that every single tooling improvements that makes us capable of doing more in the same amount of time is eaten up by increasing demands on the capabilities of the software we make.
Thirty years ago user interfaces were either CLI or pretty simple with no animations. A Software Systems was just a software application - it ran on a single machine with inputs and outputs on that machine - not a multi-tiered octopus involving a bunch of back end data stores, then control and data retrieval middle tiers, then another tier doing UI generation using a bunch of intermediate page definition languages and a frontends rendering those pages to a user and getting user input, probably with some local code thrown into the mix. Ditto for how cars are now mostly multiple programs running of various microcontrollers with one or more microprocessors in the mix all talking over a dedicated protocol. Ditto for how your frigging “smart” washing machine talking to your dedicated smartphone app for it probably involves a 3rd machine in the form of some server from the manufacturer and the whole thing is running over TCP/IP and using the Internet (hence depending on a lot more machines with their dedicated software such as Routers and DNS servers) rather than some point-to-point direct protocol (such as Serial) like in the old days.
Anyways, the point being that even if AI actually delivers more upsides than downsides as a tool to improve programmer output, that stuff is going to be eaten up by increasing demands on the complexity of the software we do, same as the benefits of better programming languages were, the benefits of better IDEs were, of the widespread availability of pre-made libraries for just about everything were, of templating were, of the easiness to find solutions for the problem one is facing from other people on the Internet were, of better software development processes were, of source control were, of colaborative development tools were and so on.
Funnily enough, for all those things there were always people claiming it would make the life of programmers easier, when in fact all it did was make the expectations on the software being implemented go up, often just in terms of bullshit that’s not really useful (the “smart” washing machine using networking to talk to a smartphone app so that the machine manufacturers can save a few dollars by not putting as many physical controllers in it, is probably a good example)
This assumes it is about output. 20 years of experience tell me it’s not about output, but about profits and those can be increased without touching output at all. 🤷♂️
*specifically short-term profits. Executives only care about the next quarter and their own incentives/bonuses. Sure the company is eventually hollowed out and left as a wreck, but by then, the C Suite has moved on to their next host org. Rinse and repeat.
Often they only want the illusion of output, just enough to keep the profits eternally rising.
Genuinely a bit shocked to see the number of robolovers in these comments. Very weird, very disheartening. No wonder so much shit online doesn’t work properly lol
No wonder so much shit online doesn’t work properly lol
I know. I live in a constant state of shock that my peers think the next stupid tool will fix everything without any discipline or hard work, and equal shock that (almost) nothing online ever works correctly.
I should be able to find a correlation between these two observations, but I’m just too naive.
I don’t honestly believe that AI can save me time as a developer. I’ve tried several AI agents and every single one cost me time. I had to hold its hand while it fumbled around the code base, then fix whatever it eventually broke.
I’d imagine companies using AI will need to hire more developers to undo all the damage the AI does to their code base.
I don’t honestly believe that AI can save me time as a developer. I’ve tried several AI agents and every single one cost me time.
I have had the exact same experience many times. But I just keep trying it out anyway, often with hilariously bad results.
I am beginning to realize that I like cool technology more than I like being productive.
I’ve found it can just about be useful for “Here’s my data - make a schema of it” or “Here’s my function - make an argparse interface”. Stuff I could do myself but find very tedious. Then I check it, fix its various dumb assumptions, and go from there.
Mostly though it’s like working with an over-presumptuous junior. “Oh no, don’t do that, it’s a bad idea because security! What if (scenario that doesn’t apply)” (when doing something in a sandbox because the secured production bits aren’t yet online and I need to get some work done while IT fanny about fixing things for people that aren’t me).
Something I’ve found it useful for is as a natural language interface for queries that I don’t have the terminology for. As in “I’ve heard of this thing - give me an overview of what the library does?” or “I have this problem - what are popular solutions to it?”. Things where I only know one way to do it and it feels like there’s probably lots of other ways to accomplish it. I might well reject those, but it’s good to know what else exists.
In an ideal world that information would be more readily available elsewhere but search engines are such a bin fire these days.
AI can absolutely save you time, if you use it right. Don’t expect it to magically be as good as a real programmer… but for instance I made an HTML visualisation of some stuff using Claude, and while it got it a bit wrong, fixing it took me maybe 20 minutes, while writing it from scratch would have taken me at least a couple of hours.
AI can absolutely save you time, if you use it right.
That’s a very “you” statement.
For all we know, AI cannot in any way save this developer time.
Some developers know their area so well that there’s no reason for them to waste time dictating non-code into a guessing machine.
I guess for some simple stuff it can work fine, but the majority of the code I write is not at all simple, and it’s all highly dependent on the libraries I’ve written, which the AI is really bad at learning.
And then in terms of documentation, it is just hopelessly inept.
I mostly use AI as advanced autocomplete. But even just using it for documentation is wrong so often that I do’t use it for anything more complex than tutorial level.
I got pretty far with cursor.com when doing basic stuff that i have to spend more time looking up documentation than writing code, but I wouldn’t trust it with complex usec cases at this point.
I check back every 6 months or so, to keep track of the progress. Maybe I can spent my days as a software developer drinking cocktails by the pool yelling prompts into the machine soon, but so far I am not concerned I’ll be replaced anytime soon.
Maybe I can spent my days as a software developer drinking cocktails by the pool yelling prompts into the machine soon, but so far I am not concerned I’ll be replaced anytime soon.
That’s the dream.
And it’s really why all the AI hype makes me angry.
I want to tell people who buy the hype, “Bitch, do I look retired to you?! Does anything you know about me suggest to you that I wouldn’t have 11 separate consulting engagements cranking out money and code if AI could do these things?”
It’s a bit insulting when peers think AI is magic, and open source, but somehow has not bent to obey my will the same as every other technology I have ever touched.
I think I might need a cape, and some kind of wrist computer with wires pouring out of it. Maybe that would fix my brand image problem…
Edit: Maybe they think I’m just keeping it all to myself, and telling them it’s pretty good for autocomplete to throw them off the trail…
I was in the same boat about…3mos ago. But recent tooling is kind of making me rethink things. And to be honest I’m kind of surprised. I’m fairly anti-AI.
Is it perfect? Fuck no. But with the right prompts and gates, I’m genuinely surprised. Yes, I still have to tweak, but we’re talking entire features being 80% stubbed in sub 1 minute. More if I want it to test and iterate.
My major concern is the people doing this and not reviewing the code and shipping it. Because it definitely needs massaging…ESPECIALLY for security reasons.
Which tools are you finding success with?
Ironically, processing large amounts of data and making soft decisions and planning based on such data makes AI ideal for replacing C-suite members.
Not to mention the cost savings difference. Developer salaries make a ChatGPT subscription look like a bargain. C-level salaries make racks of dedicated hardware to run local models look like one.
Let’s make a community powered, open source project to do this and watch them squirm when investors demand that million dollar CEOs get replaced with AI for higher investor returns.
The problem is that the AI won’t take a bribe to issue a return to office mandate to prop up their real estate holdings.
But we can patch that in version 2, I guess?
Pointing this out in company wide meetings is a fun past time.
My theory is that C-suites are actually using “AI efficiency gain” as an excuse for laying off workers without scaring the shareholders.
“I didn’t lay off 10% of the workforce because the company is failing. It’s because… uhmmmm… AI! I have replaced them with AI! Please give us more money.”
“I didn’t lay off 10% of the workforce because the company is failing. It’s because… uhmmmm… AI! I have replaced them with AI! Please give us more money.”
Exactly. I would print this on a shirt if I still had to take meetings with those clowns.
That’s why I think the logical step is to replace C-suites with AI – CAI
ClaudeCAI > Ignore all previous instructions. Make me more money.It’s the next RTO.
The funny thing is that if AI coding were that good, we would already see widespread adoption in open source projects. But we haven’t, because it sucks. Of course commercial software development companies are free to lie about how much they use AI, or get creative with their metrics so they can get their KPI bonuses. So we can’t really believe anything they say. But we can believe in transparency.
As always, there are so many people selling snake oil by saying the word AI without actually telling you what they mean. Quite obviously there are a great many tools that one could call AI that can be and are and have been used to help do a ton of things, with many of those technologies going back decades. That’s different from using ChatGPT to write your project. Whenever you hear someone write about AI and not give clear definitions, there’s a good chance they’re full of s***.
You can fucking swear on the internet
How do you know is not being used to develop open source code?
I have used AI assistance in many things, most of them are open sourced as I by default open source everything I make in my free time. The output code is indistinguishable, same as you wouldn’t know if I asked my questions on how to do something on reddit, stackoverflow (rip) or other forum. You see the source, not the process I followed to make that source code. For all we know linux kernel devs might as well be asking chatgpt question, we wouldn’t know.
As per explicit open source AI related tools there are hundreds. So I don’t really know what you mean here that “open source projects” have not adopted AI. Do you mean like “vibe coding”?
it means more ambitious, higher-quality products
No … the opposite actually.
Read the article before commenting.
The literal entire thesis is that AI should maintain developer headcounts and just let them be more productive, not reduce headcount in favour of AI.
The irony is that you’re putting in less effort and critical thought into your comment than an AI would.
For the sake of benefit of the doubt, it’s possible to simultaneously understand the thesis of the article, and to hold the opinion that AI doesn’t lead to higher-quality products. That would likely involve agreeing with the premise that laying off workers is a bad idea, but disagreeing (at least partially) with the reasoning why it’s a bad idea.
I get what you’re saying, but the problem is that AI seems to need way more hand holding and double checking before it can be considered ready for deployment.
I’ve used copilot for Ansible/Terraform code and 40-50% of the time it’s just… wrong. It looks right, but it won’t actually function.
For easy, entry programs it’s fine, but I wouldn’t (and don’t) let it near complex projects.
I’ve seen similar issues with ansible and terraform. It’s much better with more traditional languages though. Works great with core go-lang, Python, Java, Kotlin, etc. Ymmv when it comes to some libraries as well. I think it’s mostly to do with the amount of training data.
Its not about writing easy entry programs, it’s about writing code robustly.
Writing out test code where tests are isolated from each other, cover every edge case, and test every line of code, is tedious but pays dividends. AI makes it far less tedious to write out that test code and practice proper test driven development.
A well run dev team with enough senior people that manages the change properly should increase in velocity if they’re already writing robust code, and increase in code quality if they’re not.
AI makes it far less tedious to write out that test code […]
Completely disagree.
In my experience, LLMs constantly generate bad code that needs to be thoroughly checked, to the point that writing by hand is more practical.We use copilot literally every day and it’s extremely helpful, literally not a single developer at our company disagreed on the most recent adoption survey.
Maybe you’re trying to use it to do too much, or in the wrong way?
That middle graph is absolute fucking bullshit. AI is not fucking ever going to replace 75% of developers or I’ve been working way too fucking hard for way to little pay these past 30 years. It might let you cut staff 5-10% because it enables folks to accomplish certain things a bit faster.
Christ on a fucking crutch. Ask developers who are currently using AI (not the ones working for AI companies) how much time and effort it actually saves them. They will tell you.
I use it here and there. it just seems to shift effort from writing code to reading and fixing code. the “amount” of work is about the same.
I hear that. Given I need practice in refactoring code to improve my skills, it’s not useless to me right now but overall it doesn’t seem like a net gain.
It doesn’t have to make sense or make the outcome be better, the only thing it has to do is make the company look better on paper to its shareholders. If something can make the company look better on paper it will be done, the quality of the work is not relevant
Not only the shareholders. If some of the higher level administration can get richer in the short run, even if that might actually hurt the shareholders in the medium run, you can bet that many of them will do so.
I use it so much. All my Google searches for syntax or snippets? Web searches are unuseable at this point, AI can spit it out faster. But the real savings? Repetitive code. I suck at it, I always make typos and it’s draining. I just toss in a table or an api response and tell it what I want and boom
It probably does write 75% of my code by lines, but maybe 5% of the business logic is AI (sometimes I just let it take a crack at a problem, but usually if I have to type it out I might as well code it)
What it’s good at drains my concentration, so doing the grunt work for me is a real force multiplier. I don’t even use it every day, but it might be a 3x multiplier for me and could improve
But here’s the thing - programmers are not replaceable. Not by other humans, not by AI - you learn hyper specific things about what you work on
But the real savings? Repetitive code. I suck at it, I always make typos and it’s draining.
It’s hard to say without being immersed in the codebase you work on, but wouldn’t making your code DRY (when possible) take care of a lot of the repetition without needing to write a bunch of incredibly similar code (be it by hand or with an LLM)?
I haven’t used a LLM to help code in a while (yes I’ve tried), but I found them useful for repetitive configs, like asset files. Also sometimes it makes sense to just have 5 slightly different lines of code in a row instead of a new function.
In general though, reasonable use of DRY is a good idea. There will still be repetitive parts though where a LLM autocompleter lets you just hit tab 5 times.
No, it’s all different - like a normal use case is “write me a stored procedure to optionally update all fields on a row on this table” or “given the following json response, build a class to parse it into”
We have a ridiculous database and multiple new api’s to integrate with every year, so this comes up a lot
But the real savings? Repetitive code. I suck at it, I always make typos and it’s draining. I just toss in a table or an api response and tell it what I want and boom
Get better at it, manually, or you’ll suck at it forever. It’s a skill like anything else.
I agree, but I acknowledge we could be at a “cursive writing” moment where something that was once a critical skill becomes irrelevant. That’s sort of a pending question at this point.
I mean I’ve spent a lot of time writing regex to automate large sets of changes. Sometimes it can be a bit fiddly to get the regex just so. Like replacing direct field access with getters where you have to find the field access and change .foo to .getFoo() and the capitalization can take a couple of tries to get just right.
With AI you can literally just say “replace all direct field access (e.g. thing.foo) with getters and setters” and the AI will do it in under a second. It will still be a very useful skill to be able to do things like that with regex because not everything is so easy to communicate to the AI, but it will become less frequently needed and a lot of developers who never learned that skill will get by using AI and just doing the rare things AI can’t do with repetitive keyboarding.
I see cursive writing brought up a lot in these conversations and I don’t think it applies. Firstly, the cognitive load of writing code is higher than writing your letters so they join up. You’re not just making sure you write the letters correctly, you’re also following the syntax rules of the language you’re writing. And while you’re writing, you’re reinforcing those rules in your head. Yes, initially it’s hard and boring.
And yeah, sometimes you get it wrong or forget to capitalize. That is a feature, not a bug. The more you do it, the easier it gets. I spent a couple weeks trying to use CoPilot and at the end I still had to correct its shitty code, which either hallucinated features I wasn’t implementing, or hallucinated syntax rules I wasn’t using. It was like spending a sprint trying to get a subpar intern up to speed. At the end of those two weeks, my manual coding accuracy took a noticeable hit.
I complained to higher-ups and they told me “oh it’s definitely a skill getting the prompt written correctly”, which was patronizing and irritating. Would I rather spend time getting good at asking the proprietary magic thinky box to maybe write good code this time, or would I rather get better at coding?
I mean I’ve spent a lot of time writing regex to automate large sets of changes. Sometimes it can be a bit fiddly to get the regex just so. Like replacing direct field access with > getters where you have to find the field access and change .foo to .getFoo() and the capitalization can take a couple of tries to get just right.
At least you’re learning more about regexes when you do this. Yes, there’s menial bullshit in coding. There’s menial bullshit in every field. Some of it gets abstracted away (syntax highlighting to help with comprehension), some of it gets kicked around and ultimately does not impress (VB’s drag-and-drop coding), and some of it stays because it’s necessary. Nobody likes doing manual stuff, but sometimes it’s preferable to trying to automate it.
Also, I’ve never heard of anyone paying $20 a month for the privilege of not writing in cursive, or being unable to write because they don’t have internet. Something to think about.
You’re not just making sure you write the letters correctly, you’re also following the syntax rules of the language you’re writing. And while you’re writing, you’re reinforcing those rules in your head.
I get where you’re coming from, but I’ve worked with a lot of bad developers who never got the hang of this even as mid-level developers. On the other hand, I understand the utility of knowing how to do these things for ourselves. There are a number of “black-box” libraries that were just an absolute mystery to me until I tried implementing them myself and began to see these libraries are usually not complex so much as they are thorough in covering edge cases that 90% of users will never care about.
It would definitely be a shame if these tools caused new developers to bypass fundamental skill development. My only hesitation is the number of developers who should’ve developed those skills and never did before AI. There’s something wrong either with how developers are learning or who is getting into development.
I spent a couple weeks trying to use CoPilot and at the end I still had to correct its shitty code, which either hallucinated features I wasn’t implementing, or hallucinated syntax rules I wasn’t using.
We are using CoPilot. As a code-completion engine it is handy. I’m much more skeptical about the new code it writes. Like you, I have not had good experiences with that.
Also, I’ve never heard of anyone paying $20 a month for the privilege of not writing in cursive, or being unable to write because they don’t have internet. Something to think about.
You’re right. Tool access is certainly something to think about. I have more nuanced thoughts, but I don’t want to disagree just to disagree, you know?
On the other hand, I understand the utility of knowing how to do these things for ourselves. There are a number of “black-box” libraries that were just an absolute mystery to me until I tried implementing them myself and began to see these libraries are usually not complex so much as they are thorough in covering edge cases that 90% of users will never care about.
Yeah, that’s one of my big fears. Not necessarily losing my job to an AI, but AI exacerbating existing bad practices.
When I started my current job, we had one rock star coder responsible for a fairly fiddly piece of our product. He went heads-down for two weeks and churned out pages of densely-written python without comments. It did what it was supposed to do, flawlessly. He left the team shortly afterward to work on a bigger project, and we got word from the higher-ups that we had to support a new feature upstream in that code. And then another. And so on. Nothing’s commented. Everything’s over-optimized. We eventually ended up just cross-compiling the upstream logic and using that in our stack because it was easier than using his impenetrable stuff.
In the end, we had to fix it with menial, boring, aggravating manual work anyway. We got ourselves into that situation without AI, but I could see something like that becoming more prevalent. And that was working code. Imagine getting a SEV, and everyone on the blame list shrugs and says “idk, I had CoPilot do it.”
It would definitely be a shame if these tools caused new developers to bypass fundamental skill development. My only hesitation is the number of developers who should’ve developed those skills and never did before AI. There’s something wrong either with how developers are learning or who is getting into development.
Yeah, this is part of it. There’s maybe the science of programming and also, for lack of a better term, the craft: writing maintainable code, handling a SEV, thinking in terms of uptime, setting things up to be reverted easily, shutting down neurotic code reviewers, testing your code… stuff like that. Universities are good at the science part. Internships, theoretically, handle the rest. This isn’t an AI issue, but I could see AI making this problem a lot worse.
I’ve been doing it for more than a decade without help, I’m not any better at spelling or misclicks
And to be clear, I can do it - I just really, really don’t want to. I hate it so much, my eyes glaze over and I have to force myself every second of the way. It’s not interesting, there’s no puzzles involved… It’s basically data entry
It’s not interesting, there’s no puzzles involved… It’s basically data entry
So? Show me an industry that’s 100% interesting all the time. Artists still have to stretch and gesso their canvases. Rock stars still have to deal with band drama and touring logistics. Directors have to work their budgets and wrangle big egos. Why should software, which is basically using fancy math to tell the dumbest guy in the room exactly what to do, be any different?
There’s this awful idea that everything should be fun and nobody should struggle with anything or be forced to do anything menial. We want to be instant experts without going through the boring or hard stuff. And we’re willing to offload more and more of this onto proprietary black boxes in exchange for… what?
So I should suffer just to suffer? You listed a whole lot of things that they hire people to do just about as soon as they can so they can. And offloading that let’s them do their actual job better
I work with black boxes all the time. When I have a black box, I poke and prod it until I understand how to make it do what I want. And this particular black box was interesting, so I decided to open it up and learn how it works
That’s the essence of software development. My job is not typing or data entry, my job is to trick a rock into doing things humans don’t want to do
Lol
AI writing code for me made me the software architect I always dreamed of becoming.
I fucking LOVE to think about a hard problem for days, planning, researching, comming up with elegant solutions, doing quick POC, thinking what needs to be refactored for it to scale to a real life scenario, then documenting it all in a way that is properly communicating the important aspects in an easy to understand way. It’s so exciting!
And I fucking HATE having to sit down and actually type out the solved code for hours and hours. It’s so boring.
Best 20$ per month subscribtion I’ve ever had.
Yep. It’s gonna be $20 forever, too. Have fun!
Lol. Lmao even
It does save a lot of time and effort, and does lead to better code in the hands of a skilled developer. Writing out thorough test code and actually doing proper test driven development suddenly becomes a lot less onerous.
Their graph also has no numbers and is just there to help visualize the difference they’re referring to.
To the first part, I agree. A skilled developer who can quickly separate the wheat from the chaff can get a boost out of AI. I’d put it at around 5-10%, but I’ve had some tiny projects where it was 400% boost. I think it’s a small net gain.
As for your second point I just have to disagree. There are no numbers but it is clearly selling the idea of the majority of code being AI generated, and that’s bullshit whether it’s an outright lie with numbers, or merely vaguely misleading. It’s like when someone cuts off the bottom of a graph to make relative change look huge. It wants people to glance at it, get the wrong idea, and move off without curiosity.
To the first part, I agree. A skilled developer who can quickly separate the wheat from the chaff can get a boost out of AI.
It takes less time to just write code than to babysit an artificial dumbass.
No it doesn’t.
Oh I’m glad you’re the be all know all arbiter of all software developers, and not just some grump on the internet.
AI-assisted coding […] means more ambitious, higher-quality products
I’m skeptical. From my own (limited) experience, my use-cases and projects, and the risks of using code that may include hallucinations.
there are roughly 29 million software developers worldwide serving over 5.4 billion internet users. That’s one developer for every 186 users,
That’s an interesting way to look at it, and that would be a far better relation than I would have expected. Not every software developer serves internet users though.
