As someone who has interviewed candidates for developer jobs for over a decade: this sounds like “in my day everything was better”.
Yes, there are plenty of candidates who can’t explain the piece of code they copied from Copilot. But guess what? A few years ago there were plenty of candidates who couldn’t explain the code they copied from StackOverflow. And before that, there were those who failed at the basic programming test we gave them.
We don’t hire those people. We hire the ones who use the tools at their disposal and also show they understand what they’re doing. The tools change, the requirements do not.
I think that LLMs just made it easier for people who want to know but not learn to know. Reading all those posts all over the internet required you to understand what you pasted together if you wanted it to work (not always but the barr was higher). With ChatGPT, you can just throw errors at it until you have the code you want.
While the requirements never changed, the tools sure did and they made it a lot easier to not understand.
Have you actually found that to be the case in anything complex though? I find it just forgets parts to generate something. Stuck in an infuriating loop of fucking up.
It took us around 2 hours to run our coding questions through chatgpt and see what it gives. And it gives complete shit for most of them. One or two questions we had to replace.
If a company cannot invest even a day to go through their hiring process and AI proof it, then they have a shitty hiring process. And with a shitty hiring process, you get shitty devs.
And then you get people like OP, blaming the generation while if anything its them and their company to blame… for falling behind. Got to keep up folks. Our field moves fast.
I find ChatGPT to sometimes be excellent at giving me a direction, if not outright solving the problem, when I paste errors I’m to lazy to look search. I say sometimes because othertimes it is just dead wrong.
All code I ask ChatGPT to write is usually along the lines for “I have these values that I need to verify, write code that verifies that nothing is empty and saves an error message for each that is” and then I work with the code it gives me from there. I never take it at face value.
Have you actually found that to be the case in anything complex though?
I think that using LLMs to create complex code is the wrong use of the tool. They are better at providing structure to work from rather than writing the code itself (unless it is something simple as above) in my opinion.
If a company cannot invest even a day to go through their hiring process and AI proof it, then they have a shitty hiring process. And with a shitty hiring process, you get shitty devs.
I agree with you on that.
But how do you find those people solely based on a short interview, where they can use AI tools to perform better if the interview is not held in person?
And mind you the SO was better because you needed to read a lot of answers there and try to understand what would work in your particular case. Learn how to ask smartly. Do your homework and explain the question properly so as not to get gaslit, etc. this is all now gone.
Pretty easy to come up with problems that chatGPT is useless at. You can test it pretty easily. Throw enough constraints at it and the transformer starts to loose attention and forget vital parts.
With a bit of effort you can make problems where chatGPT will actuallt give a misleading answer and candidates have to think critically.
Just like in the past it was pretty easy to come up with problems which werent easily found on SO.
Same landscape. If you put in the time and the effort to have a solid recruitment process, you get solid devs. If you have a lazy and shitty process, you get shitty devs.
The problem is not only the coding but the thinking. The AI revolution will give birth to a lot more people without critical thinking and problem solving capabilities.
apart from that, learning programming went from something one does out of calling, to something one does to get a job. The percentage of programmers that actually like coding is going down, so on average they’re going to be worse
I’ve said it before, but this is a 20-year-old problem.
After Y2K, all those shops that over-porked on devs began shedding the most pricey ones; worse in ‘at will’ states.
Who were those devs? Mentors. They shipped less code, closed fewer tickets, cost more, but their value wasn’t in tickets and code: it was investing in the next generation. And they had to go because #numbersGoUp
And they left. And the first gen of devs with no mentorship joined and started their careers. No idea about edge cases, missing middles or memory management. No lint, no warnings, build and ship and fix the bugs as they come.
And then another generation. And these were the true ‘lost boys’ of dev. C is dumb, C++ is dumb, perl is dumb, it’s all old, supply chain exploits don’t exist, I made it go so I’m done, fuck support, look at my numbers. It’s all low-attention span, baling wire and trophies because #numbersGoUp.
And let’s be fair: they’re good at this game, the new way of working where it’s a fast finish, a head-pat, and someone else’s problem. That’s what the companies want, and that’s what they built.
They say now that relying on Ai makes one never really exercise critical thought and problem-solving, and I see it when I’m forced to write fucking YAML for fucking Ansible. I let the GPTs do that for me, without worrying that I won’t learn to code YAML for Ansible. Coding YAML for Ansible is NEVER going to be on my list of things I want to remember. But we’re seeing people do that with actual work; with go and rust code, and yeah, no concept of why we want to check for completeness let alone a concept of how.
What do we do, though?
If we’re in a position to do so, FAIL some code reviews on corner cases. Fail some reviews on ISO27002 and supply chain and role sep. Fail some deployments when they’re using dev tools in prod. And use them all as teachable moments. Honestly, some of them got no mentorship in college if they went, and no mentorship in their first ten years as a pro. It’s going to be hard getting over themselves, but the sooner they realise they still have a bunch to learn, the better we can rebuild coders. The hardest part will be weaning them off GPT for the cheats. I don’t have a solution for this.
One day these new devs will proudly install a patch in the RTOS flashed into your heart monitor and that annoying beep will go away. Sleep tight.
I have seen this too much. My current gripe isn’t fresh devs, as long as they are teachable and care.
My main pain over the last several years has been the bulk of ‘give-no-shit’ perms/contractors who don’t want to think or try when they can avoid it.
They run a web of lies until it is no longer sustainable (or the project is done for contractors) and then again its someone else’s problem.
There are plenty of 10/20 year plus and devs who don’t know what they are doing and don’t care whose problem it will be as long as it isnt theirs.
I’m sick of writing coding 101 standards for 1k+ a day ‘experts’. More sick of PR feedback where it’s a battle to get things done in a maintainable manner from said ‘experts’.
No one wants mentors. The way to move up in IT is to switch jobs every 24 months. So when you’re paying mentors huge salaries to train juniors who are velocity drags into velocity boosters, you do it knowing they are going to leave and take all that investment with them for a higher paycheck.
I don’t say this is right, but that’s the reality from the paycheck side of things and I think there needs to be radical change for both sides. Like a trade union or something. Union takes responsibility for certifying skills and suitability, companies can be more confident of hires, juniors have mentors to learn from, mentors ensure juniors have aptitude and intellectual curiosity necessary to do the job well, and I guess pay is more skill/experience based so developers don’t have to hop jobs to get paid what they are worth.
Fixed typos due to my iPhone hating me.
Yeah those job hoppers are the worst. You can always tell right away what kind of person those are. I’ve had to work with a “senior” dev who had 15 years of experience and to be honest he sucked at his job. He couldn’t do simple tasks, didn’t think before he started writing code and often got stuck asking other people for help. But he got paid big bucks, because all he did his entire career was work somewhere for 2-3 years and then job hop and trade up. By the time the company figured out the dude was useless, he went on to the next company.
Such a shitty attitude, which is a shame because he was a good dude otherwise. I got along with him on a personal level. And honestly good on him for making the most he can, fuck the company. But I personally couldn’t do that, I take pride in my work.
I let the GPTs do that for me, without worrying that I won’t learn to code YAML for Ansible.
And this is the perfect use case. There’s a good chance someone has done exactly what you want, and AI can regurgitate that for you.
That’s not true of any interesting software project though.
FAIL some code reviews on corner cases. Fail some reviews on ISO27002 and supply chain and role sep. Fail some deployments when they’re using dev tools in prod. And use them all as teachable moments.
Fortunately, I work at an org that does this. It turns out that if our product breaks in prod, our customers could lose millions, which means they could go to a competitor. We build software to satisfy regulators, regulators that have the power to shut down everything if the ts aren’t crossed just so.
Maybe that’s the problem, maybe the stakes are low enough that quality isn’t important anymore. Idk, what I do know is that I go hard on reviews.
and I see it when I’m forced to write fucking YAML for fucking Ansible. I let the GPTs do that for me, without worrying that I won’t learn to code YAML for Ansible. Coding YAML for Ansible is NEVER going to be on my list of things I want to remember.
Feels like this is the attitude towards programming in general nowadays.
To be fair, YAML sucks. It’s a config language that someone thought should cover everything, but excel at nothing.
Just use TOML, JSON, or old-school INI. YAML will just give you an aneurism. Use the best tool for the job, which is often not the prettiest one.
Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away.
Antoine de Saint-Exupéry
Kids these days with their fancy stuff, you don’t need all that to write good software. YAML is the quintessential “jack of all trades, master of none” nonsense. It’s a config file, just make it easy to parse and document how to edit it. That’s it.
While there is some truth to what you said, it sounded to me too much like “old man yells at clouds” because you are over-generalizing. Not everything new is bad. Don’t get stuck in the past, that’s just as dumb as relying on AI.
What are you guys working on where chatgpt can figure it out? Honestly, I haven’t been able to get a scrap of working code beyond a trivial example out of that thing or any other LLM.
I’m forced to use Copilot at work and as far as code completion goes, it gets it right 10-15% of the times… the rest of the time it just suggests random — credible-looking — noise or hallucinates variables and shit.
I would quit, immediately.
Pay my bills. Thanks.
I’ve been dusting off the CV, for multiple other reasons.
Agreed. I wanted to test a new config in my router yesterday, which is configured using scripts. So I thought it would be a good idea for ChatGPT to figure it out for me, instead of 3 hours of me reading documentation and trying tutorials. It was a test scenario, so I thought it might do well.
It did not do well at all. The scripts were mostly correct but often in the wrong order (referencing a thing before actually defining it). Sometimes the syntax would be totally wrong and it kept mixing version 6 syntax with version 7 syntax (I’m on 7). It will also make mistakes and when I point out the mistake it says Oh you are totally right, I made a mistake. Then goes on to explain what mistake it did and output new code. However more often than not the new code contained the exact same mistake. This is probably because of a lack of training data, where it is referencing only one example and that example just had a mistake in it.
In the end I gave up on ChatGPT, searched for my testscenario and it turned out a friendly dude on a forum put together a tutorial. So I followed that and it almost worked right away. A couple of minutes of tweaking and testing and I got it working.
I’m afraid for a future where forums and such don’t exist and sources like Reddit get fucked and nuked. In an AI driven world the incentive for creating new original content is way lower. So when AI doesn’t know the answer, you are just hooped and have to re-invent the wheel yourself. In the long run this will destroy productivity and not give the gains people are hoping for at the moment.
It’s like useful information grows as fruit from trees in a digital forest we call the Internet. However, the fruit spoils over time (becomes less relevant) and requires fertile soil (educated people being online) that can be eroded away (not investing in education or infrastructure) or paved over (intellectual property law). LLMs are like processed food created in factories that lack key characteristics of more nutritious fresh ingredients you can find at a farmer’s market. Sure, you can feed more people (provide faster answers to questions) by growing a monocrop (training your LLM on a handful of generous people who publish under Creative Commons licenses like CC BY-SA on Stack Overflow), but you also risk a plague destroying your industry like how the Panama disease fungus destroyed nearly all Gros Michel banana farming (companies firing those generous software developers who “waste time” by volunteering to communities like Stack Overflow and replacing them with LLMs).
There’s some solar punk ethical fusion of LLMs and sustainable cultivation of high quality information, but we’re definitely not there yet.
To extend your metaphor: be the squirrel in the digital forest. Compulsively bury acorns for others to find in time of need. Forget about most of the burial locations so that new trees are always sprouting and spreading. Do not get attached to a single trunk ; you are made to dance across the canopy.
This is probably because of a lack of training data, where it is referencing only one example and that example just had a mistake in it.
The one example could be flawless, but the output of an LLM is influenced by all of its input. 99.999% of that input is irrelevant to your situation, so of course it’s going to degenerate the output.
What you (and everyone else) needs is a good search engine to find the needle in the haystack of human knowledge, you don’t need that haystack ground down to dust to give you a needle-shaped piece of crap with slightly more iron than average.
When I had to get up to speed on a new language, it was very helpful. It’s also great to write low to medium complexity scripts in python, powershell, bash, and making ansible tasks. That said I’ve been programming for ~30 years, and could have done those things myself if I needed, but it would take some time (a lot of it being looking up documentation and writing boilerplate code).
It’s also nice for writing C# unit tests.
However, the times I’ve been stuck on my main languages, it’s been utterly useless.
ChatGPT is extremely useful if you already know what you’re doing. It’s garbage if you’re relying on it to write code for you. There are nearly always bugs and edge cases and hallucinations and version mismatches.
It’s also probably useful for looking like you kinda know what you’re doing as a junior in a new project. I’ve seen some shit in code reviews that was clearly AI slop. Usually from exactly the developers you expect.
Yeah, I’m not even that down on using LLMs to search through and organize text that it was trained on. But in it’s current iteration? It’s fancy stack overflow, but stack overflow runs on like 6 servers. I’ll be setting up some LLM stuff self hosted to play around with it, but I’m not ditching my brain’s ability to write software any time soon.
I’ve been using (mostly) Claude to help me write an application in a language I’m not experienced with (Rust). Mostly with helping me see what I did wrong with syntax or with the borrow checker. Coming from Java, Python, and C/C++, it’s very easy to mismanage memory the exact way Rust requires it.
That being said, any new code that generates for me I end up having to fix 9 times out of 10. So in a weird way I’ve been learning more about Rust from having to correct code that’s been generated by an LLM.
I still think LLMs for the next while will be mostly useful as a hyper-spell checker for code, and not for generating new code. I often find that I would have saved time if I just tackled the problem myself and not tried to reply on an LLM. Although sometimes an LLM can give me an idea on how to solve a problem.
Same. It can generate credible-looking code, but I don’t find it very useful. Here’s what I’ve tried:
- describe a function - takes longer to read the explanation than grok the code
- generate tests - hallucinates arguments, doesn’t do proper boundary checks, etc
- looking up docs - mostly useful to find search terms for the real docs
The second was kind of useful since it provided the structure, but I still replaced 90% of it.
I’m still messing with it, but beyond solving “blank page syndrome,” it’s not that great. And for that, I mostly just copy something from elsewhere in the project anyway, which is often faster than going to the LLM.
I’m really bad at explaining what I want, because by the time I can do that, it’s faster to just build it. That said, I’m a senior dev, so I’ve been around the block a bit.
I used it a few days ago to translate a math formula into code.
Here is the formula: https://wikimedia.org/api/rest_v1/media/math/render/svg/126b6117904ad47459ad0caa791f296e69621782
It’s not the most complicated thing. I could have done it. But it would take me some time. I just input the formula directly, the desired language and the result was well done and worked flawlessly.
It saved me some time typing around. And searching online a few things.
Not in any way a new phenomenon, there’s a reason fizzbuzz was invented, there’s been a steady stream of CS graduates who can’t code their way out of a wet paper bag ever since the profession hit the mainstream.
Actually fucking interview your candidates, especially if you’re sourcing candidates from a country with for-profit education and/or rote learning cultures, both of which suck when it comes to failing people who didn’t learn anything. No BS coding tests go for “explain this code to me” kind of stuff, worst case they can understand code but suck at producing it, that’s still prime QA material right there.
We do two “code challenges”:
- Very simple, many are done in 5 min; this just weeds out the incompetent applicants, and 90% of the code is written (i.e. simulate working in an existing codebase)
- Ambiguous requirements, the point is to ask questions, and we actually have different branches depending on assumptions they made (to challenge their assumptions); i.e. simulate building a solution with product team
The first is in the first round, the second is in the technical interview. Neither are difficult, and we provide any equations they’ll need.
It’s much more important that they can reason about requirements than code something quick, because life won’t give you firm requirements, and we don’t want a ton of back and forth with product team if we can avoid it, so we need to catch most of that at the start.
In short, we’re looking for actual software engineers, not code monkeys.
Those are good approaches, I would note that the “90% is written” one is mostly about code comprehension, not writing (as in: Actually architect something), and the requirement thing is a thing that you should, IMO, learn as a junior, it’s not a prerequisite. It needs a lot of experience, and often domain knowledge new candidates have no chance of having. But, then, throwing such stuff at them and then judging them by their approach, not end result, should be fair.
The main question I ask myself, in general, is “can this person look at code from different angles”. Somewhat like rotating a cube in your mind’s eye if you get what I mean. And it might even be that they’re no good at it, but they demonstrate the ability when talking about coffee making. People who don’t get lost when you’re talking about cash registers having a common queue having better overall latency than cash registers with individual queues. Just as a carpenter would ask someone “do you like working with your hands”, the question is “do you like to rotate implication structures in your mind”.
judging them by their approach, not end result, should be fair.
Yup, that’s the approach. It’s okay if they don’t finish, I want to know how they approach the problem. We absolutely adjust our decision based on the role.
If they can extend existing code and design a new system (with minimal new code) and ask the right questions, we can work with them.
We’re a somewhat big player in a niche industry that manufactures for a large industry. Yearly profits are in the hundreds of millions of dollars, market cap is a few billion, so low end of mid cap stocks. I don’t want to doxx myself, but think of something like producing drills for oil rigs and you won’t be far off.
We have about 50 software developers across three time zones (7 or 8 scrum teams) and a pretty high requirement for correctness and very little emphasis on rapid delivery. It’s okay if it takes more time, as long as can plan around it, so we end up with estimates like 2-3 months for things that could have an MVP in under a month (in fact, we often build an MVP during estimation), with the extra time spent testing.
So yeah, it’s a nice place to work. I very rarely stay late, and it’s never because a project is late, but because of a high severity bug in prod (e.g. a customer can’t complete a task).
