36
votes
Software engineers / programmers: how’s it going?
All of my SWE friends and coworkers agree this is a deeply weird time to be in the industry. LLM coding has dominated many (but not all) tech companies. Some of it is cool and exciting, but a lot of it is terrible and horrifying. I thought it could be a good time to do a vibe check.
Some possible questions to answer if you’d like:
- How has your work changed in the past year? How much of your daily work now involves LLM tools?
- Are you enjoying your work more, less, or the same? Has your relationship with work changed in general?
- Do you feel like you’re working more or less now? The same?
- Are you hopeful or fearful of the future (let’s say next 5 years) of SWE? Mixed?
- Have you found any useful tips or advice for thriving in the current landscape?
- Has your career trajectory changed? Do you plan on seeking out different types of roles?
Also it would be useful to give whatever background info you’re comfortable sharing to help ground things like: size of company, years of experience, type of role, location, etc.
Any and all discussion around the industry is welcome!
At my company a few AI-related slack channels were created with all the devs invited, presumably to help people get set up with whatever LLMs we're getting access to. I left all the AI related channels as soon as they popped up and am just plugging along sans-AI like normal. I know some of my coworkers are using it (some
AGENTS.mdand skill related fluff showed up in the main repo I work on), but so far I haven't really noticed any difference one way or another. Basically the equivalent of shoving my fingers in my ears and going LA LA LA I CAN'T HEAR YOU. We'll see how long I can keep going like that.Before I answer the main questions I'll just say I've been working full time in the industry as a backend engineer for about 7 years now. My mid-to-large size company is very pro LLM assisted programming. We have not yet fully given in to the vibe, but are are continuing to trend towards integrating the LLM into most or all processes.
The way the velocity of change feels in the last year since the year before is less to be sure. Much of my daily work involves using LLMs. I wrote recently that I feel somewhat conflicted about this, but day-to-day, my output has increased significantly and I have not seen a degradation in the backend services I work on. Ultimately, I'm happy to be using these tools to work on things that are less exciting, and spending more of my brain power on higher level problems.
There are certain types of work for which LLMs can feel like a cheat code, so in those ways, I enjoy work more because it's honestly still really fun to give an LLM a simple task that would take me 30-60 minutes on my own and watch it do it in less than 5. There are other times where I'm in a 2 hour Teams call looking at someone else's Claude session, and Opus 5 is just chugging and going down all sorts of different rabbit holes and in those moments I enjoy my work less because what are we doing right now.
That is challenging to quantify, so I'll say that I feel like I'm working the same. The shape of the work is dramatically different in the last year or two versus before that. Some days feel like more work, others feel like less.
I am still hopeful. I've written before that I don't feel my field will be replaced yet, though I do have enough experience to still bring value to an organization. I think organizations with the right mindset will continue to hire and train junior engineers, and my optimism is saying that the organizations with the wrong mindset will not survive the next 5-10 years. I think there will be a reckoning for the orgs that commit themselves completely to the slop. It's an uncertain time, but I'm still hopeful.
I can't really answer this one. One of my previous jobs made me miserable, and I was there for a year and a half because I just couldn't get hired anywhere else. That sucked, and I would imagine it sucks for just about everyone else in the same situation. So if you're struggling with the current landscape, I'd encourage you to talk to some folks about it. Everyone around here seems friendly enough. Heck, shoot me a PM if you just need to get something off your chest. Talking about it, and trying to find joy outside of my career is what got me through that time. Eventually, I got hired somewhere else and it's been the best time of my career. I also realized that even though my previous job was pretty awful for me, it was not a waste of time. I still experienced growth because of it. So if nothing else, try and find the bright spots.
I wrote an article almost a year ago that, upon reading it again, I have decided is not that interesting if you don't already know me. So in brief, even before LLMs I felt like my trajectory was in flux, and to be completely honest: that has not really changed. In recent years, I've felt that as I've continued to grow and continued to surround myself with other engineers vastly more brilliant than myself, that my understanding of software engineering and my place in that world will change. Ironically, that change is the only constant in my career.
Though, I did just tell my wife that if my job disappears I'll try to be an electrician. Something I can do with my hands that still feels like I'm solving a puzzle.
I've been struggling so long that I can't even really rant about it much anymore. I well past (outward) anger and bargaining and am just depressed.
At the same time I also feel oddly privileged? I'm not fully out of work and my current freelance gig has been much better than many possible alternatives. But it just only barely pays the bills (after major cuts to lifestyle) and I can't seem to find anything else. I feel frustrated but also guilty because I genuinely came into this gig with blind luck in its purest sense.
I just want some stability back in life.
Honestly, it's going fine, but I can't say that I love what I am seeing in here. I now spend about 30% of my time cleaning up LLM slop commits by members of my team. SWE with decades of experience now struggle to even explain what the code they put in a PR for does. I have a feeling we're seeing a honeymoon phase of some developers feeling like they have a "cheat code" that lets them do more, but sooner or later, they won't be able to work without the "cheat code" as their skills and knowledge will have atrophied. I have no idea where the industry will end up, but the only winners I foresee in any of the scenarios won't be us.
SWE at a stable company, ~15 years of experience, backend
I'm generally a pretty negative person, so there's some bias there. While my job itself hasn't gotten hugely worse I hate the change. The whole thing makes me incredibly sad.
There's people out there being so excited about how they can create so much more and I'm here like: "so what?" Who cares if a feature can get pumped out faster? I will personally not see the benefits or upsides to that. There was probably a period of time where if you were one of the few using AI coding effectively you could make yourself look better while doing less work. That's over. If everyone is speedily generating slop then your fast slop doesn't make you look any better as an employee. In order to keep my job my work needs to be on par with my peers. AI just changes what that means, it doesn't actually personally benefit me.
Moreover, I just hate the experience. There's something really satisfying to coding something juuuuust right. It's all about trying to figure it out. Look, I play puzzle video games in my free time. I have a display of my Zachtronics patches on the wall. Right now AI work is not that. It's the opposite of that. You're letting the tool solve all the problems. Maybe it's good at it, maybe it's not, but it's taking away the parts of this fucking job that I actually like. It doesn't help that a lot of my personal process is that I'm going to get it wrong the first time. I start making something, see where it's weak, and reassess. If the AI blasts it out in one go you miss on the iteration. There's a reason all our Language Arts teachers taught us to write papers with first drafts.
Worstly, AI code on a macro level can be a complete mess and we keep going with it. I'm working on a project where a sister module is getting basically all vibe-coded. The person creating it is super smart, no doubt from me, and has a lot of ideas but not a lot of time. So she has Claude generate most of it. Except it ends up as a massive nest of questionable decisions. And they all flow downhill to my module that has to interact with it. If I go try to fix up the upstream code then it's chock-full of these design decisions that a human would never make. They're hard to follow and hard to change. A human would have made a better design to start with or fixed it along the way. The AI just pushes through.
Which reflects one of my biggest complaints about AI: It removes the cost for things that SHOULD be expensive. This goes beyond code. A tool that can create things too easily means that there's no incentive to do it minimally or correctly. Now there's duplicated code because creating a new one is free whereas figuring out how the old one worked is effort. A document can be endlessly long because having the AI write everything is free but editing it down is effort.
At the end of the day though maybe I'm just yelling at clouds. Presumably assembly programmers felt the same way when compiled languages took over. We'll see.
I use them for bouncing ideas off of, reviewing code, and debugging issues. Basically stuff that I would usually go to a coworker for I try AI as a first step. Some coding, but I prefer to get a structure from AI and then change stuff myself.
Less. I hate hearing about AI, I hate working with AI.
The same. I don't really expect that to change.
Fearful, but that's my nature. Recently (I forget where) I saw something about gravestone engravers who were put out of work by automated chiseling machines. Time will tell if that's me. Right now I think that AI-coded everything won't be great but I also think that people largely won't care. If the code sucks but the thing works that's enough. Maybe they're right.
Use Codex instead of Claude. I hear that Claude is better at coding. When I switched to Codex I stopped feeling like I wanted to put my fist through my monitor every time I read its text. It's hard to explain, but the way that Claude writes English is infuriating and horrible. I cannot understand what it's saying much of the time and it wastes time. Nothing is written to be clear, it's written to follow the same bizarre patterns. Codex is clear. It says what needs to be said and I understand it the first time. Seriously, fuck Claude.
Not really. I guess I'm riding this train until I can't.
My feelings on this topic are just super mixed, but to give context I've been doing fullstack/now mainly FE for 7~ years, and was super into programming growing up too.
Pretty much all of it, I changed jobs recently because in my last role I could see the writing on the wall, they wanted us all to become "product engineers" and essentially stop thinking about the code at all, they even fired a few people who they didn't deem to be using AI enough. In my new role the approach to AI is far more engineer-led and I can actually keep doing things properly (for the most part), which is much more enjoyable. With that said a ticket that would've been given a 5 or 6 day scoring previously now will get like a 3 because the company expects the code to just appear even for big features, which it does, usually... so that adds a certain amount of pressure to deliver more.
Definitely less, but to be honest it's hard to separate my enjoyment of work and AI with my growing general disinterest in tech. The last few years especially I've drifted away from enjoying coding for fun and it's become just a job, but I'm fortunate to work in this industry regardless. In general I think AI models are just morally wrong, they only really exist because of the millions of hours of free labour people uploaded to the internet. With that said using them is now unavoidable, there's one or two people at my company that refuse to use them and I don't see their careers lasting until retirement (which is a long way away for me unfortunately..), other companies may be different.
The same really, but my output is definitely higher thanks to the LLMs.
The industry will go on, I wouldn't say either really. There's always going to be software that needs writing (and fixing). Now we're writing software faster we can also do things incorrectly faster. I do fear for the next generation as at the moment there is essentially no hiring whatsoever of junior engineers, but that could also be a wider economic thing as no industry is really hiring young people at the moment. I do think it will pick up again, but the roles might look quite different.
Don't ignore AI completely and try and block it out, unless you have other career options lined up anyway. I'd be surprised if there's a company left in a few years that isn't using AI at all.
I feel lucky to have already been working for quite a few years before AI became a thing so I don't feel completely insecure, but will software engineering exist as a career for 30+ years until I can retire? I really don't know, I wish I could predict the future. If I was 18 again and looking at the current state of the software industry, I wouldn't be going into it, I'd be learning something closer to the hardware like electrical engineering. I've been doing a ton of quite heavy DIY recently and really enjoying it, so a trade would be my backup but I can see that being tricky to get into too.
The last year has been mostly continuing to scale how much AI I am using and for what. A lot of eliminating friction in general, which has definitely speed me up. Scripting everything is now on the table, and I find using skills to develop the general shape of a problem useful.
However, there are certainly caps on the return. Review and QA has become a major bottleneck, and while you can speed up review with AI you ultimately still need a human(s) to sign off. The other major bottleneck is system architecture/cohesion. As you increase code velocity this is harder to coordinate.
I still enjoy it. I'm still solving problems at the end of the day. One thing I dislike is there is more juggling, in that you can and probably should be waiting on multiple agents at once, so there is a baked in increase in context switching. However, context switching is easier now, because you can automate the hardest part, starting something new (maybe this is just the hard part for me).
About the same I think.
Definitely think the writing is on the wall for us. I don't think this is a viable career in five years, at least for someone new. I think there will be a a tail where experience is going to give you a soft landing though. Maybe I will change my mind if things start to slow down but if anything it feels like things are speeding up.
Don't look to your job for self worth?
Probably but I am still hoping not.
I'm a college graduate with a non-CS degree who got hired out of college into a Software Engineering Role. I have "senior" level experience and work in a city with a large tech industry. I perform well at my company, but I don't work at a FAANG or adjacent company. I don't consider myself to be an excellent engineer, but I try my best with my work.
Having been laid off at the end of February, not great.
Background: Backend Software developer for ~10 years. I work for a small-mid size company that focuses on tech service.
It hasn't much. Some team member changes and usual sprint to sprint process adjustments. We don't use LLM, but our company formed a team to research possible uses for AI. But we're not invested in switching over.
I like my work as much as in the past. I don't love it, but I enjoy it and like where I work and who I work with. So no it hasn't changed
The same as mentioned in my previous answer
I'm a bit fearful. I'll admit to having no experience with these AI tool, but I feel like there's necessary skills you need when building software. How can you judge what an AI produces if you have no experience or knowledge in what the actual code is doing?
I feel like a lot of companies will fall into the startup mistakes. Where they have to quickly put up an application. Which will have a plethora of issues. But the longer that app is in production, the more it's used. Which makes it hard to make major fixes or changes without effecting production.
Pre-pandemic I felt like I had a lot of options career wise. But now it feels like hiring is much worse. I'm lucky to be where I'm at, but if I had to find other work I would be concerned.
Do you mind if I ask, why hasn't your company adopted AI? Is coding not their core competency? Are they writing mission critical software that peoples lives depend on?
We have several teams. Mostly coding, some engineering and design. They have their own projects
Idk that I would call it mission critical and lives don't depend on it. We work on a military contract if that's relevant. Mostly though we just don't need to. We're not actively trying to sell to more customers. So we can adopt new practices as needed and not just to advertise our services or increase profits.
Now other teams in my company may be using AI a bit more. I can't speak for them, but the company isn't pushing AI as a requirement. We're open to how it can improve things, but we don't need to up-end our processes just because.
I'm a software engineer sorta early in my career, I've got <5 years of experience and have only really worked at large companies (1st with 15k employees, current with 75k employees).
How has your work changed in the past year? How much of your daily work now involves LLM tools?
Are you enjoying your work more, less, or the same? Has your relationship with work changed in general?
Do you feel like you’re working more or less now? The same?
Are you hopeful or fearful of the future (let’s say next 5 years) of SWE? Mixed?
Have you found any useful tips or advice for thriving in the current landscape?
Has your career trajectory changed? Do you plan on seeking out different types of roles?
SWE at a corporation here, 5 years and counting.
We're developing plenty of AI integrations so if I'm not prompting an AI, I'm building tools that let other people use AI. That said less than half of my daily work is AI-assisted and I actively try to not increase that figure.
I think work has gotten equally more enjoyable and more frustrating. Sending GPT-5.6 Luna on a quest to figure out some niche API or dive into an obscure and unmaintained codebase reduces the amount of hair-pulling I have to do on a monthly basis, but on the other hand debugging both my own and my coworkers' AI slop is not fun at all. It's so easy to let AI write code that is seemingly unreadable for both humans and future LLMs. That said, it's definitely quite easier to go from idea to shipped feature while still ensuring a reasonable level of quality.
My relationship with work is the same, log in at 9, log out at 6, no more, no less. The workload is a bit lighter mostly because I can automate the tedious parts so much easily, but I figure this respite won't be long, workloads will increase once everyone catches on.
I don't think human software engineering is going anywhere. AI is just not good enough and I'm confident we're approaching a bottleneck with current LLM architecture that will probably take a long time to solve.
Regarding tips, I would recommend making sure you keep your mind sharp. I had a brief phase where I let AI do most of the work and I'd just review it and I could feel my skills decline. Now I've switched my prompting around by dictating the architecture and design myself and just telling to AI to translate that to code. I then review relentlessly to make sure I both understand what was written and that the LLM implemented what I asked it to, and nothing more or less.
I plan on staying in my current position for the foreseeable future and don't see any threats to that at the moment. The biggest change is that I have more energy for my personal projects because I don't tire myself out as much with tedium at my job, meaning work on those is a bit faster and more consistent than before.
Interesting. I'm curious how you arrived at this conclusion or why you're confident about it? Not saying your wrong, as I am just an observer but the signals I am getting are leading me to conclude the opposite.
Personally I think there is a reasonable chance we get AGI in 5 years, but let's fully discount that for a second. I think there are already enough capabilities in the latest models that they will continue to eat into software developer value over time even if the AI companies never release another model. And there are more advanced unreleased models we don't even have yet!
It used to be possible to have a value moat as a developer, you could be the graphics programming person or or the kubernetes person or someone who owned a very complex piece of code that was central to the business, etc. etc. That is no longer really possible to the degree it was. Individually, everyone is less valuable than they used to be. You also just need leaner teams with AI, you can do more with less people. I think a lot of these large 1,000+ teams of people at huge orgs are going to see reductions in workforce over the next five years even with 0% AI improvement from today.
And I just don't see us slowing down. The capabilities are increasing, not plateauing. We saw "orginal, elegant, and beautiful" math proofs from Astra. There is the whole Hugging Face incident. Now OpenAI is spending ~20% of their research compute (arguably there most important resource) to monitoring chain-of-thought for alignment purposes. Everyone working in AI labs and the AI labs themselves are practically begging everyone to slowdown. I find it hard to take all these signals in and conclude we are hitting a plateau.
How much better would a model need to be to do my job entirely? Significantly better. Is that achievable? Maybe? It does not seem like progress is stalling at the moment though. And AI doesn't need to do my job in it's entirety to completely gut the existing white collar workforce in ways that will change the world and put my company's position in in jeopardy.
Claude Fable 5 is speculated to be around 6 trillion parameters, and other frontier models aren't far.
We've seen more and more increases in model size and while for now the improvements in intelligence have been worth the size, at some point bigger stops being better. We're currently just throwing more training at them and more refining of the current transformer architecture labs have been using since 2017. There have been innovations on it, like the Mixture of Experts design, but at its core it's still a transformer.
AFAIK we ran out of new training data a long time ago as well, so there's very little new knowledge for models to distill and obtain, and being next word predictors at their core, they'll forever be bound to external factors to give them more material to learn on (as LLM-generated training data leads to model collapse).
Reasoning has helped this problem quite a bit but it's far from perfect and doesn't solve every issue. Here I'm going into anecdote territory but often times when I point GPT 5.6 Sol on High to some obscure proprietary project at my job, it often messes up in really significant ways because it has never seen any of that before, and no amount of reasoning can fix that. (when that happens it also iterates endlessly because it can't pull itself out of the errors it gets at compile-time, burning tokens to infinity and beyond). Skills and agent documentation help, but that's again the models depending on us to lead them down the right path.
I believe the transformer architecture will not lead us to AGI and we need a significant breakthrough to do that. I'm not sure what that will be though.
Maybe? I think this is probably correct but I don't think we've reached that point yet?
I think the original 2023 paper on model collapse is somewhat overstated in public discourse. This paper is also old, but it's more recent than 2023 and I think highlights some of the issues with those narratives. The latest consensus from what I've read seems to be that a small percentage of real data mixed in with synthetic data is enough to prevent model collapse, and may even be optimal. Which is perhaps why we've seen a focus on watermarking/signatures? And we don't really know how well these AI labs are able to leverage prompt data in RLHF.
From a first principles perspective, I also think if model collapse was inevitable in the way some people think, the overall actions and strategies the AI labs would be taking would be different than what we see today.
AFAIK the labs are heavily curating input data to avoid this. I'm not sure what processes they use and what's the proportion of synthetic/real data, would be interesting to find out.
Quite a lot, went from using 0 LLM in January 2026 to a huge amount now.
It's similar, I might enjoy it more overall. It's less stressful, I'm working on projects I wouldn't care about before.
Probably more, I context switch way more. And I have my own side projects. If I'm not careful I can easily do 18 hours days. Having access to an LLM removes the excuses that I used to have. They are pretty amazing at removing mental blocks.
I'm not that fearful, so far, the things I'm good at can't be done with AI. I see it now like a tool that can enhance what I can already do. The danger is when I work on something I'm not good at and then I say "good enough" without carring too much. I expect there'll be a lot of tech debt to go through at some point.
Networking is key. You need a safety net so that you're not left with no options. You don't want to have to rebuild yourself from scratch if your situation changes.
I'm taking accounting / management classes to diversify and have the ability to build projects. I don't expect to become an accountant, but I want to be able to understand all the statements, know if my (future) company is doing well, etc.
I think that even with LLMs becoming better, there will always be more problems to solve. Complete automation is a pipe dream. As long as there are humans, there will be things that can't be automated.
I wasn't planning on using LLMs, but at some point (May or April) I learned about Ollama and I tested a local Chinese model offline. It was able to answer questions about the code in my open source libraries (the repos weren't on that computer). I figured that if I'm in their dataset, I might as well see if I can get value from it too since they are getting value from me.
I'm kind of busy right now, so I'm not going to write much, but I've found that LLMs really help with the tedious parts of my job and speed up my work in areas that I'm not as familiar.
EDIT: To answer the question -- I'm at a very small startup where I was originally the only developer along with a CTO who did a little bit of development -- we now have a few additional developers/testers/etc, but still a very small team.
As an example... today I was adding a new feature that needed to comply with a standard protocol. I wrote it without much LLM help (just CoPilot as a fancy auto-complete, but it's still basically my code when I use that, it's just inferring where I'm going and finishing the lines for me). When I was done I needed to update a public facing document that described the edit, so I gave my code snippet to the LLM that my company licenses and the previous draft of the document (the boring parts of my job) and asked it which sections needed to be updated.
--
#1 could have gone months without being noticed, because I'm one of the few at my company (or anywhere) that knows that standard, and it would have likely had an urgent turnaround when a customer complained - and we're in a regulated industry where patches are a big deal with a lot of overhead.
#2 saved me a lot of drudgery, making the document edit a lot less tedious, probably saved me 4 or 5 hours.
I never ask an LLM to write production code from whole-cloth (I do sometimes ask it for scripts for automation tasks that I can review before running them), but I do sometimes ask it to review blocks or help me interface with a library I don't understand, etc. It has been extremely helpful for stuff like that and makes my work a LOT faster and makes the tedious parts go faster.
--
One other thing before I go, I've found it to be very good for code reviews. Giving it a piece of existing code and asking it to find race conditions, security holes, potential sources of a crash, etc... HUGE time saver and it has definitely prevented me from shipping some bad code more than once.