People who are saying they're not seeing productivity boost, can you please share where is it failing?
Because, I am terrified by the output I am getting while working on huge legacy codebases, it works. I described one of my workflow changes here: https://news.ycombinator.com/item?id=47271168 but in general compared to old way of working I am saving half of the steps consistently, whether its researching the codebase, or integrating new things, or even making fixes. I have stopped writing code, occasionally I jump into the changes proposed by LLM and make manual edits if it is feasible, otherwise I revert changes and ask it to generate again but based on my learnings from the past rejected output
The companies laying off people have no vision. My company is a successful not for profit and we are hiring like crazy. It’s not a software company, but we have always effectively unlimited work. Why would anyone downsize because work is getting done faster? Just do more work, get more done, get better than the competition, get better at delivering your vision. We put profits back in the community and actually make life better for people. What a crazy fucking concept right?
I suspect it depends partly on how locked each individual is into a particular type of work, both skill-wise and temperamentally.
To give an example from a field where LLMs started causing employment worries earlier than software development: translation. Some translators made their living doing the equivalent of routine, repetitive coding tasks: translating patents, manuals, text strings for localized software, etc. Some of that work was already threatened by pre-LLM machine translation, despite its poor quality; context-aware LLMs have pretty much taken over the rest. Translators who were specialized in that type of work and too old or inflexible to move into other areas were hurt badly.
The potential demand for translation between languages has always been immense, and until the past few years only a tiny portion of that demand was being met. Now that translation is practically free, much more of that demand is being met, though not always well. Few people using an app or browser extension to translate between languages have much sense of what makes a good translation or of how translation can go bad. Professional translators who are able to apply their higher-level knowledge and language skills to facilitate intercultural communication in various ways can still make good money. But it requires a mindset change that can be difficult.
That was my insight also. As a manager, you already have the headcount approved, and your people just allegedly got some significant percentage more productive. The first thought shouldn't be, great let's cut costs, it should be great now we finally have the bandwidth to deliver faster.
On a macro level, if you were in a rising economic tide, you would still be hiring, and turning those productivity gains into more business.
I wonder what the parallels are to past automations. When part producing companies moved from manual mills to CNC mills, did they fire a bunch of people or did they make more parts?
This is exactly right IMO. I have never worked for a company where the bottleneck was "we've run out of things to do". That said, plenty of companies run out of actual software engineering work when their product isn't competitive. But it usually isn't competitive because they haven't been able to move fast enough
These are words without weights. At some point the put money into software option will max out. Perhaps what we should all be doing is hiring more lawyers, there's always more legal work to be done. When you don't have weights then you can reason like this.
I don’t know what kind of software your used to but software is pretty much universally dog shit these days. I could probably count on one hand the number of programs that I actually like using. There is an astronomical room for improvement. I don’t think we are hitting diminishing returns any time soon.
I talk about this at length in one of my previous posts here: https://news.ycombinator.com/item?id=39963058. I definitely share your opinion and I think this will be exacerbated by vibe coding and having LoC as the main KPI for engineering teams.
That's one of the reasons why I am terrified, because it can lead to burn out, and I personally don't like to babysit bunch of agents, because the output doesn't feel "mine", when its not "mine" I don't feel ownership.
And I am deliberately hitting the brake from time to time not to increase expectations, because I feel like driving someone else's car while not understanding fully how they tuned their car (even though I did those tunings by prompting)
It feels very much like leading a team of junior engineers or even interns who are very fast but have no idea about why we're doing anything. You have to understand the problems you're trying to solve and describe the solutions in a way they can be implemented.
It's not going to be written exactly like you would do it, but that's ok - because you care about the results of the solution and not its precise implementation. At some point you have to make an engineering decision whether to write it yourself for critical bits or allow the agent/junior to get a good enough result.
You're reviewing the code and hand editing anyway, right? You understand the specs even if your agent/junior doesn't, so you can take credit even if you didn't physically write the code. It's the same thing.
> It feels very much like leading a team of junior engineers or even interns who are very fast but have no idea about why we're doing anything
Yes, yes!
And this is problem for me, because of the pace, my brain muscles are not developing enough compared to when I was doing those things myself.
before, I was changing my mind while implementing the code, because I see more things while typing, and digging deeper, but now, because juniors are doing things they don't offer me a refactoring or improvements while typing the code quickly, because they obey my command instead of having "aha" moment to suggest better ways
I don’t think the objections are not necessarily in terms of lack of productivity although my personal experience is not that of massive productivity increases. The fact that you are producing code much faster is likely just to push the bottleneck somewhere else. Software value cycles are long and complicated. What if you run into an issue in 5 years the LLM fails to diagnose or fix due to complex system interactions? How often would that happen? Would it be feasible to just generate the whole thing anew matching functionality precisely? Are you making the right architecture choices from the perspective of what the preferred modus operandi of an llm is in 5 years? We don’t know. The more experienced folks tend to be conservative as they have experienced how badly things can age. Maybe this time it’ll be different?
> People who are saying they're not seeing productivity boost, can you please share where is it failing?
At review time.
There are simply too many software industries that can't delegate both authorship _and_ review to non-humans because the maintenance/use of such software, especially in libraries and backwards-compat-concerning environments, cannot justify an "ends justifies the means" approach (yet).
I work as an ML engineer/researcher. When I implement a change in an experiment it usually takes at least an hour to get the results. I can use this time to implement a different experiment. Doesn't matter if I do it by hand or if I let an agent do it for me, I have enough time. Code isn't the bottleneck.
I also heard an opinion that since writing code is cheap, people implement things that have no economic value without really thinking it through.
+1 on the economic value line. Not everything needs to be about money but if you get paid to ship code it's about money. And now we have coworkers shipping insane amounts of "features" because it's all free to ship and being an engineer, it ends there.
Only it doesn't, there's product positioning, UX, information architecture, onboarding and training, support, QA, change management, analytics, reporting… sigh
It’s been my experience as of recently. I point it at an issue tracker and ask it to investigate, write a test to reproduce the problem and plan a fix together. There’s lots of hand holding from me but it saves me a lot of work and I’ve been surprised by its comfort with legacy code bases. For now I feel empowered, and I’m actually working more intensively, but I was wondering to myself if I’m going run out of work this year. Interestingly, our metrics show that output is slowed by increased workload on reviewers.
AI dramatically increases velocity. But is velocity productivity? Productivity relative to which scope: you, the team, the department, the company?
The question is really, velocity _of what_?
I got this from a HN comment. It really hit for me because the default mentality for engineers is to build. The more you build the better. That's not "wrong" but in a business setting it is very much necessary but not sufficient. And so whenever we think about productivity, impact, velocity, whatever measure of output, the real question is _of what_? More code? More product surface area? That was never really the problem. In fact it makes life worse majority of the time.
The real question is, is it increasing their velocity?
They've already admitted they just 'throw the code away and start again'.
I think we've got another victim of perceived productivity gains vs actual productivity drop.
People sitting around watching Claude churn out poor code at a slower rate than if they just wrote it themselves.
Don't get me wrong, great for getting you started or writing a little prototype.
But the code is bad, riddled with subtle bugs and if you're not rewriting it and shoving large amounts of AI code into your codebase, good luck in 6-12 months time.
Exactly, the $20 codex is so good value it’s irresponsible to not give it to everyone. Claude code $20 is otoh pointless, the limits are good enough for 10 mins of work twice per business day.
Every business that's taking AI seriously is giving their team enterprise accounts to AI services. Otherwise you have no control over where your code, data, company info, etc is going.
Someone deciding to drop a spreadsheet of customer data into their personal AI account to increase their productivity would be catastrophic for business, so you need rules. And rules means paying for enterprise AI tooling.
Not everyone has the capability to rent out data center tier hardware to just do their job. These things require so much damn compute you need some serious heft to actually daisy chain enough stages either in parallel or deep to get enough tokens/sec for the experience to go ham. If you're making bags o' coke money, and deciding to fund Altman's, Zuckernut's or Amazon/Google's/Microsoft's datacenter build out, that's on you. Rest of us are just trying to get by on bits and bobs we've kept limping over the years. If opencode is anything to judge the vibecoded scene by, I'm fairly sure at some point the vibe crowd will learn the lesson of isolating the most expensive computation ever from the hot loop, then maybe find one day all they needed was maybe something to let the model build a context, and a text editor.
I asked opus 4.6 how to administer an A/B test when data is sparse. My options are to look at conversion rate, look at revenue per customer, or something else. I will get about 10-20k samples, less than that will add to cart, less than that will begin checkout, and even less than that will convert. Opus says I should look at revenue per customers. I don't know the right answer, but I know it is not to look at revenue per customers -- that will have high variance due to outlier customers who put in a large order. To be fair, I do use opus frequently, and it often gives good enough answers. But you do have to be suspicious of its responses for important decisions.
Edit: Ha, and the report claims it's relatively good at business and finance...
Edit 2: After discussion in this thread, I went back to opus and asked it to link to articles about how to handle non-normally distributed data, and it actually did link to some useful articles, and an online calculator that I believe works for my data. So I'll eat some humble pie and say my initial take was at least partially wrong here. At the same time, it was important to know the correct question to ask, and honestly if it wasn't for this thread I'm not sure I would have gotten there.
A/B tests are a statistical tool, and outliers will mess with any statistical measure. If your data is especially prone to that you should be using something that accounts for them, and your prompt to Opus should tell it to account for that.
A good way to use AI is to treat it like a brilliant junior. It knows a lot about how things work in general but very little about your specific domain. If your data has a particular shape (e.g lots of orders with a few large orders as outliers) you have to tell it that to improve the results you get back.
I did tell it that I expect to see something like a power-law distribution in order value, so I think I pretty much followed your instructions here. Btw, if you do know the right thing to do in my scenario, I'd love to figure it out. This is not my area of expertise, and just figuring it out through articles so far.
I recommend reading Wikipedia and talking to LLMs to get this one. Order values do follow power-law distributions (you're probably looking for an exponential or a Zipf distribution.) You want to ask how to perform a statistical test using these distributions. I'm a fan of Bayesian techniques here, but it's up to you if you want to use a frequentist approach. If you can follow some basic calculus you can follow the math for constructing these statistical tests, if not some searching will help you find the formulas you need.
Thanks for the suggestions! I didn't want to do the math myself, but I did take your suggestion and found some articles discussing ways to make it work even with a non-normal distribution:
I'm not checking their math, but the articles make sense to me, and I trust they did implement it correctly. In the end the LLM did get me to the correct answer by suggesting the articles, so I guess I should eat some humble pie and say it _did_ help me. At the same time, if I didn't have the intuition that using rpv as-is in a t-test would be noisy, and the suggestions from this comment thread, I think I could have gone down the wrong path. So I'm not sure what my conclusion is -- maybe something like LLMs are helpful once you ask the right question.
I'm with you. The project I'm working on is moving at phenomenal velocity. I'm basically spending my time writing specs and performing code reviews. As long as my code review comments and design docs are clear I get a secure, scalable, and resilient system.
Tests were always important, but now they are the gatekeepers to velocity.
I feel like this might be heavily dependent on both your task and the AI you're using? What language do you code in and what AI do you use? And are your tasks pretty typical/boilerplate-y with prior art to go off of, or novel/at-the-edge-of-tech?
Because its failure rate is too high. Beyond boilerplate code and CRUD apps, if I let AI run freely on the projects I maintain, I spend more time fixing its changes than if I just did it myself. It hallucinates functionally, it designs itself into corners, it does not follow my instructions, it writes too much code for simple features.
It’s fine at replacing what stack overflow did nearly a decade ago, but that isn’t really an improvement from my baseline.
That's my experience too. It's okay at a few things that save me some typing, but it isn't really going to do the hard work for me. I also still need to spend significant amounts of time figuring out what it did wrong and correcting it. And that's frustrating. I don't make those mistakes, and I really dislike being led down bad paths. If "code smells" are bad, then "AI" is a rotting corpse.
Outside of coding/non-physical areas, the impact can be quite muted. I haven't seen much impact on surgical procedures, for example (but maybe others have?).
I’m currently working across like 5 projects (was 4 last week but you know how it is). I now do more in days than others might in a week.
Yesterday a colleague didn’t quite manage to implement a loading container with a Vue directive instead of DOM hacks, it was easier for me to just throw AI at the problem and produced a working and tested solution and developer docs than to have a similarly long meeting and have them iterate for hours.
Then I got back to training a CNN to recognize crops from space (ploughing and mowing will need to be estimated alongside inference, since no markers in training data but can look at BSI changes for example), deployed a new version of an Ollama/OpenAI/Anthropic proxy that can work with AWS Bedrock and updated the docs site instructions, deployed a new app that will have a standup bot and on-demand AI code review (LiteLLM and Django) and am working on codegen to migrate some Oracle forms that have been stagnating otherwise.
It’s not funny how overworked I am and sure I still have to babysit parallel Claude Code sessions and sometimes test things manually and write out changes, but this is a completely different work compared to two or three years ago.
Maybe the problem spaces I’m dealing with are nothing novel, but I assume most devs are like that - and I’d be surprised at people’s productivity not increasing.
When people nag in meetings about needing to change something in a codebase, or not knowing how to implement something and its value add, I’ll often have something working shortly after the meeting is over (due to starting during it).
Instead of sending adding Vitest to the backlog graveyard, I had it integrated and running in one or two evenings with about 1200 tests (and fixed some bugs). Instead of talking about hypothetical Oxlint and Oxfmt performance improvements, I had both benchmarked against ESLint and Prettier within the hour.
Same for making server config changes with Ansible that I previously didn’t due to additional friction - it is mostly just gone (as long as I allow some free time planned in case things vet fucked up and I need to fix them).
Edit: oh and in my free time I built a Whisper + VLM + LLM pipeline based on OpenVINO so that I can feed it hours long stream VODs and get an EDL cut to desired length that I can then import in DaVinci Resolve and work on video editing after the first basic editing prepass is done (also PyScene detect and some audio alignment to prevent bad cuts). And then I integrated it with subscription Claude Code, not just LiteLLM and cloud providers with per-token costs for the actual cuts making part (scene description and audio transcriptions stay local since those don't need a complex LLM, but can use cloud for cuts).
Oh and I'm moving from my Contabo VPSes to running stuff inside of a Hetzner Server Auction server that now has Proxmox and VMs in that, except this time around I'm moving over to Ansible for managing it instead of manual scripts as well, and also I'm migrating over from Docker Swarm to regular Docker Compose + Tailscale networks (maybe Headscale later) and also using more upstream containers where needed instead of trying to build all of mine myself, since storage isn't a problem and consistency isn't that important. At the same time I also migrated from Drone CI to Woodpecker CI and from Nexus to Gitea Packages, since I'm already using Gitea and since Nexus is a maintenance burden.
If this becomes the new “normal” in regards to everyone’s productivity though, there will be an insane amount of burnout and devaluation of work.
> When people nag in meetings about needing to change something in a codebase, or not knowing how to implement something and its value add, I’ll often have something working shortly after the meeting is over (due to starting during it).
We've started building harnesses to allow people who don't understand code to create PRs to implement their little nags. We rely on an engineer to review, merge, and steward the change but it means that non-eng folks do not rely on us as a gate. (We're a startup and can't really afford "teams" to do this hand-holding and triage for us.)
As you say we're all a bit overworked and burned out. I've been context switching so much that on days when I'm very productive I've started just getting headaches. I'm achieving a lot more than before but holding the various threads on my head and context switching is just a lot.
I can only explain it by people not having used Agentic tools and or only having tried it 9 months ago for a day before giving up or having such strict coding style preferences they burn time adjusting generated code to their preferences and blaming the AI even though they’re non-functional changes and they didn’t bother to encode them into rules.
The productivity gains are blatantly obvious at this point. Even in large distributed code bases. From jr to senior engineer.
I can see someone who is very particular about their way being the right way having issues with it. I’m very much the kind of person who believes that if I can’t write a failing test, I don’t have a very serious case. A lot of devs aren’t like that.
A terminology tangent because it's an econ publication: Notice that the article doesn't talk about productivity.
Productivity is a term of art in economics and means you generate more units of output (for example per person, per input, per wages paid) but doesn't take quality or otherwise desireability into account. It's best suited for commodities and industrial outputs (and maybe slop?).
You were probably deficient in RESEARCH skills before. No offense to you, since I was also like this once. LLMs research and put the results on the plate. Yes, for people who were deficient in research skills, I can see 2-3x improvements.
Note1: I have "expert" level research skills. But LLMs still help me in research, but the boost is probably 1.2x max. But
Note2: By research, I mean googling, github search, forum search, etc. And quickly testing using jsfiddle/codepen, etc.
But I also think you are overestimating your RESEARCH skills, even if you are very good at research, I am sure you can't read 25 files in parallel, summarize them (even if its missing some details) in 1 minute and then come up with somewhat working solution in the next 2 minutes.
I am pretty sure, humans can't comprehend reading 25 code files with each having at least 400 lines of non-boilerplate code in 2 minutes. LLM can do it and its very very good at summarizing.
I can even steer its summarizing skills by prompting where to focus on when its reading files (because now I can iterate 2-3 times for each RESEARCH task and improve my next attempt based on shortcomings in the previous attempt)
please don't change your comment constantly (or at maybe add UPDATE 1/2/3), because you had different words before, like you were saying something in this fashion:
* you probably lack good RESEARCH skills
* I can see at most 1.25x improvements - now it is 2-3x
By updating your comment you are making my reply irrelevant to your past response
I don't write code for a living but I administer and maintain it.
Every time I say this people get really angry, but: so far AI has had almost no impact on my job. Neither my dev team nor my vendors are getting me software faster than they were two years ago. Docker had a bigger impact on the pipeline to me than AI has.
Maybe this will change, but until it does I'm mostly watching bemusedly.
Yep. All AI has done for me is give me the power of how good search engines were 10+ years ago, where I could search for something and find actually relevant and helpful info quickly.
I've seen lots of people say AI can basically code a project for them. Maybe it can, but that seems to heavily depend on the field. Other than boilerplate code or very generic projects, it's a step above useless imo when it comes to gamedev. It's about as useful as a guy who read some documentation for an engine a couple years ago and kind of remembers it but not quite and makes lots of mistakes. The best it can do is point me in the general direction I need to go, but it'll hallucinate basic functions and mess up any sort of logic.
My experience is the same. There are modest gains compensating for lack of good documentation and the like, but the human bottlenecks in the process aren't useless bureaucracy. Whether or not a feature or a particular UX implementation of it makes sense, these things can't be skipped, sped up or handed off to any AI.
For me this is a huge boost in productivity. If I remember how I was working in the past (example of Google integration):
Before:
* go through docs to understand how to start (quick start) and things to know
* start boilerplate (e.g. install the scripts/libs)
* figure out configs to enable in GCP console
* integrate basic API and test
* of course it fails, because its Google API, so difficult to work with
* along the way figure out why Python lib is failing to install, oh version mismatch, ohh gcc not installed, ohh libffmpeg is required,...
* somehow copy paste and integrate first basic API
* prepare for production, ohhh production requires different type of Auth flow
* deploy, redeploy, fix, deploy, redeploy
* 3 days later -> finally hello world is working
Now:
* Hey my LLM buddy, I want to integrate Google API, where do I start, come up with a plan
* Enable things which requires manual intervention
* In the meantime LLM integrates the code, install lib, asks me to approve installation of libpg, libffmpeg,....
* test, if fails, feed the error back to LLM + prompt to fix it
* deploy
That's only certain if investments in tech infrastructure always led to productivity increases. But sometimes they just don't. Lots of firms spent a lot of money on blockchain five years ago, for instance, and that money is just gone now.
I find it odd the universal assumption that AI is going to be good for productivity
The loss of skills, complete loss of visibility and experience with the codebase, and the complete lack of software architecture design, seems like a massive killer in the long term
I have a feeling that we're going to see productivity with AI drop through the floor
I'd claim the opposite. Better models design better software, and quickly better software than what most software developers were writing.
Just yesterday I asked Opus 4.6 what I could do to make an old macOS AppKit project more testable, too lazy to even encumber the question with my own preferences like I usually do, and it pitched a refactor into Elm architecture. And then it did the refactor while I took a piss.
The idea that AI writes bad software or can't improve existing software in substantial ways is really outdated. Just consider how most human-written software is untested despite everyone agreeing testing is a good idea simply because test-friendly arch takes a lot of thought and test maintenance slow you down. AI will do all of that, just mention something about 'testability' in AGENTS.md.
OK so this comes back to the question I started this subthread with: where is this better software? Why isn't someone selling it to me? I've been told for a year it's coming any day now (though invariably the next month I'm told last month's tools were in fact crap and useless compared to the new generation so I just have to wait for this round to kick in) and at some point I do have to actually see it if you expect me to believe it's real.
I'm happy to sell it to you, though it is also free. I guided Claude to write this in three weeks, after never having written a line of JavaScript or set up a server before. I'm sure a better JavaScript programmer than I could do this in three weeks, but there's no way I could. I just had a cool idea for making advertising a force for good, and now I have a working version in beta.
I'd say it is better software, but better is doing a lot of heavy lifting there. Claude's execution is average and always will be, that's a function of being a prediction engine. But I genuinely think the idea is better than how advertising works today, and this product would not exist at all if I had to write it myself. And I'm someone who has written code before, enough that I was probably a somewhat early adopter to this whole thing. Multiply that by all the people whose ideas get to live now, and I'm sure some ideas will prove to be better even with average execution. Like an llm, that's a function of statistics.
In glad you made something with it you wanted to make, and as a fan of Aristotle I'm always happy to see the word eudaimonia out there. Best of luck. That said I don't understand what this does or why I would want the tokens it mentions.
Yeah, I gotta make a video walkthrough. Its basically a goal tracker combined with an ad filter - write what you want out of life and block ads, it replaces them with ads that actually align with your long term goals instead of distracting from them. The tokens let you add ads to the network, though you also get some for using the goal tracker.
Though this does suggest one possible answer to me: the new software is largely web applications, and the web is just a space I don't spend much time anymore other than a few retro sites like this
How would you know if all software written in the last six months shipped X% faster and was Y% better?
Why would you think you have your finger on the pulse of general software trends like that when you use the same, what, dozen apps every week?
Just looking at my own productivity, as mere sideprojects this month, I've shipped my own terminal app (replaced iTerm2), btrfs+luks NAS system manager, overhauled my macOS gamepad mapper for the app store, and more. All fully tested and really polished, yet I didn't write any code by hand. I would have done none of that this month without AI.
You'd need some real empirics to pick up productivity stories like mine across the software world, not vibes.
Right, I'm sympathetic to the idea that LLMs facilitate the creation of software that people previously weren't willing to pay for, but then kind of by definition that's not going to have a big topline economic impact.
It's on the people pushing AI as the panacea that has changed things to show workings. Not someone saying "I've not seen evidence of it". Otherwise it's "vibes" as you put it.
Having the productivity "drop through the floor" is a bit hyperbolic, no? Humans are still reviewing the PRs before code merge at least at my company (for the most part, for now).
I don't know that it's likely but it's certainly a plausible outcome. If tooling keeps getting built for this and the financial music stops it's going to take a while for everybody to get back up to speed
Remember this famously happened before, in the 1970s
There's an actual working product now, albeit one which is currently loss leading. In software world at least there is definitely enough value for it to be used even if it's just better search engine. I'm not sure why it would disappear if the financial music stops as opposed to being commoditised.
Because there's cheaper ways to get an equally good search engine? But yes I imagine some amount of inference will continue even in an AI Winter 3.0 scenario.
Ironically, abstraction bloat eats away any infra gains. We trade more compute to allow people less in tune with the machine to get things done, usually at the cost of the implementation being eh... Suboptimal, shall we say.
I think there's a broad category error where people see that every gain has been an abstraction (true) but conclude from that that every abstraction will be a gain (dubious)
Same here, more or less, in the ops world. Yeah, I use AI but I can't honestly say it's massively improved my productivity or drastically changed my job in any way other than the emails I get from the other managers at my work are now clearly written by AI.
I can turn out some scripts a little bit quicker, or find an answer to something a little quicker than googling, but I'm still waiting on others most of the time, the overall company processes haven't improved or gotten more efficient. The same blockers as always still exist.
Like you said, there has been other tech that has changed my job over time more than AI has. The move to the cloud, Docker, Terraform, Ansible, etc. have all had far more of an impact on my job. I see literally zero change in the output of others, both internally and externally.
So either this is a massively overblown bubble, or I'm just missing something.
> ... but I'm still waiting on others most of the time, the overall company processes haven't improved or gotten more efficient. The same blockers as always still exist.
And that's the key problem, isn't it? I maintain current organizations have the "wrong shape" to fully leverage AI. Imagine instead of the scope of your current ownership, you own everything your team or your whole department owns. Consider what that would do to the meetings and dependencies and processes and tickets and blockers and other bureaucracy, something I call "Conway Overhead."
Now imagine that playing out across multiple roles, i.e. you also take on product and design. Imagine what that would do to your company org chart.
Imagine if your grandma had wheels! She'd be a bicycle. Now imagine she had an engine. She could be a motorcycle! Unfortunately for grandma, she lives in reality and is not actually a motorcycle, which would be cool as hell. Our imagination can only take us so far.
To more substantively reply to your longer linked comment: your hypothesis is that people spend as little as 10% of time coding and the other 90% of time in meetings, but that if they could code more, they wouldn't need to meet other people because they could do all the work of an entire team themselves[1]. The problem with your hypothesis is that you take for granted that LLMs actually allow people to do the work of an entire team themselves, and that it is merely bureacracy holding them back. There have been absolutely zero indicators that this is true. No productivity studies of individual developers tackling tasks show a 10x speedup; results tend to be anywhere from +20% to minus 20%. We aren't seeing amazing software being built by individual developers using LLMs. There is still only one Fabrice Bellard in the world, even though if your premise could escape the containment zone of imagination anyone should be able to be a Bellard on their own time with the help of LLMs.
[1] Also, this is basically already true without LLMs. It is the reason startups are able to disrupt corporate behemoths. If you have just a small handful of people who spend the majority of their work time writing code (by hand! No LLMs required!), they can build amazing new products that outcompete products funded by trillion-dollar entities. Your observation of more coding = less meetings required in the first place has an element of truth to it, but not because LLMs are related to it in any particular way.
> Imagine if your grandma had wheels! She'd be a bicycle.
I always took this to be a sharp jab saying the entire village is riding your grandma, giving it a very aggressive undertone. It's pretty funny nonetheless.
Too early to say what AI brings to the efficiency table I think. In some major things I do it's a 1000x speed up. In others it is more a different way of approaching a problem than a speed up. In yet others, it is a bit of an impediment. It works best when you learn to quickly recognize patterns and whether it will help. I don't know how people who are raised with ai will navigate and leverage it, which is the real long-term question (just as the difference between pre- and post-smartphone generations is a thing).
> No productivity studies of individual developers tackling tasks show a 10x speedup; results tend to be anywhere from +20% to minus 20%.
The only study showing a -20% came back and said, "we now think it's +9% - +38%, but we can't prove rigorously because developers don't want to work without AI anymore": https://news.ycombinator.com/item?id=47142078
Even at the time of the original study, most other rigorous studies showed -5% (for legacy projects, obsolete languages) to 30% (more typical greenfield AND brownfield projects) way back in 2024. Today I hear numbers up to 60% from reports like DX.
But this is exactly missing the point. Most of them are still doing things the old way, including the very process of writing code. Which brings me to this point:
> There have been absolutely zero indicators that this is true.
I could tell you my personal experience, or link various comments on HN, or point you to blogs like https://ghuntley.com/real/ (which also talks about the origanizational impedance mismatch for AI), but actual code would be a better data point.
So there are some open-source projects worth looking at, but they are typically dismissed because they look so weird to us. Here's two mostly vibe-coded (as in, minimal code review, apparently) projects that people shredded for having weird code, but is already used by 10s of 1000s of people, up to 11 - 18K stars now. Look at the commit volume and patterns for O(300K) LoC in a couple of months, mostly from one guy and his agent:
It's like nothing we've seen before, almost equal number of LoC additions and deletions, in the 100s of Ks! It's still not clear how this will pan out long term, but the volume of code and apparent utility (based purely on popularity) is undeniable.
> they are typically dismissed because they look so weird to us.
I dismiss them because Yegge's work (if it can even be called his work, given that he doesn't look at the code) is steaming garbage with zero real-world utility, not "because they look weird". You suggest the apparent utility is undeniable, while saying "based purely on popularity" -- but popularity is in no way a measure of utility. Yegge is a conman who profited hundreds of thousands of dollars shilling a memecoin rugpull tied to these projects. The actual thousands of users are people joining the hypetrain, looking to get in on the promised pyramid scheme of free money where AI will build the next million dollar software for you, if only you have the right combination of .md files to make it work. None of these software are actually materialising, so all the people in this bubble can do is make more AI wrappers that promise to make other AI wrappers that will totally make them money.
I am completely open to being proven wrong by a vibe-coded open source application that is actually useful, but I haven't seen a single one. Literally not even one. I would count literally anything where the end-product is not an AI wrapper itself, which has tens to hundreds of thousands of users, and which was written entirely by agents. One example of that would be great. Just one. There have been a couple of attempts at a web browser, and Claude's C compiler, but neither are actually useful or have any real users; they are just proofs of concept and I have seen nothing that convinces me they are a solid foundation from which you could actually build useful software from, or that models will ever be on a trajectory to make them actually useful.
This isn't the counter you think it is. It's too much to expect existing behemoths to reshape their orgs substantially on a quick enough timeline. The gains will be first seen in new companies and new organizations, and they will be able to stay flat a longer and outcompete the behemoths.
I've been in ops for 30 years, Claude Code has changed how I work. Ops-related scripting seems to be a real sweet spot for the LLMs, especially as they tend to be smaller tools working together. It can convert a few sentences into working code in 15-30 minutes while you do something else. I've given it access to my apache logs Elastic cluster, and it does a great job at analyzing them ("We suspect this user has been compromised, can you find evidence of that?"). It's quite startling, actually, what it's able to do.
Yeah, it's useful for scripting, but it's still only marginally faster. It certainly hasn't been "groundbreaking productivity" like it's being sold.
The problem with analyzing logs is determinism. If I ask Claude to look for evidence of compromise, I can't trust the output without also going and verifying myself. It's now an extra step, for what? I still have to go into Elastic and run the actual queries to verify what Claude said. A saved Kibana search is faster, and more importantly, deterministic. I'm not going to leave something like finding evidence of compromise up to an LLM that can, and does, hallucinate especially when you fill the context up with a ton of logs.
An auditor isn't going to buy "But Claude said everything was fine."
Is AI actually finding things your SIEM rules were missing? Because otherwise, I just don't see the value in having a natural language interface for queries I already know how to run, it's less intuitive for me and non deterministic.
It's certainly a useful tool, there's no arguing that. I wouldn't want to go back to working with out it. But, I don't buy that it's already this huge labor market transformation force that's magically 100x everyone's productivity. That part is 100% pure hype, not reality.
The tolerance for indeterminacy is I think a generational marker; people ~20 years younger than me just kind of think of all software as indeterminate to begin with (because it's always been ridiculously complicated and event-driven for them), and it makes talking about this difficult.
People younger than me are not even adults. I grew up during the dial up era and then the transition to broadband. I don't think software is indeterminate.
I shudder to think of how many layers of dependency we will one day sit upon. But when you think about it, aren’t biological systems kind of like this too? Fallible, indeterminable, massive, labyrinthine, and capable of immensely complex and awe inspiring things at the same time…
Is it? A couple days ago I had it build tooling for a one-off task I need to run, it wrote ~800 lines of Python to accomplish this, in <30m. I found it was too slow, so I got it to convert it to run multiple tasks in parallel in another prompt. Would have taken a couple days for me to build from hand, given the number of interruptions I have in the average day. This isn't a one-off, it's happening all the time.
Imo it's only a matter of time as companies start to figure out how to use ai. Companies don't seem to have real plans yet and everyone is figuring out ai in general out.
Soon though I will think agents start popping up, things like first line response to pages, executing automation
To deal with the problems where there is ambiguity in the problem and the approach to solving it. Not everything is a basic decision tree. Humans aren't deterministic either, the way we woukd approach a problem is probably different. Is one of us right or wrong? We're generally just focused on end results.
Maybe 2 years ago Ai was doing random stuff and we got all those funny screenshots of dumb gemini answers. The indeterminism leading to random stuff isn't really an issue any more.
Humans are funny. But most cant seem to understand that the tool is a mirage and they are putting false expectations on it. E.g. management of firms cutting back on hiring under the expectation that LLMs will do magic - with many cheering 'this is the worst itll be bro!!".
I just hope more people realise before Anthropic and OAI can IPO. I would wager they are in the process of cleaning up their financials for it.
My employer is pretty advanced in its use of these tools for development and it’s absolutely accelerated everything we do to the point we are exhausting roadmaps for six months in a few weeks. However I think very few companies are operating like this yet. It takes time for tools and techniques to make it out and Claude code alone isn’t enough. They are basically planning to let go of most of the product managers and Eng managers, and I expect they’re measuring who is using the AI tools most effectively and everyone else will be let go, likely before years end. Unlike prior iterations I saw at Salesforce this time I am convinced they’re actually going to do it and pull it off. This is the biggest change I’ve seen in my 35 year career, and I have to say I’m pretty excited to be going through it even though the collateral damage will be immense to peoples lives. I plan to retire after this as well, I think this part is sort of interesting but I can see clearly what comes next is not.
Why are you excited for this? They’re not going to give YOU those peoples’ salaries. You will get none of it. In fact, it will drag your salary through the floor because of all the available talent.
The dev team is committing more than they used to. A lot, in fact, judging from the logs. But it's not showing up as a faster cadence of getting me software to administer. Again, maybe that will change.
I’m at a big tech company. They proudly stated more productivity measures in commits (already nonsense). 47% more commits, 17% less time per commit. Meaning 128% more time spent coding. Burning us out and acting like the AI slop is “unlocking” productivity.
There’s some neat stuff, don’t get me wrong. But every additional tool so far has started strong but then always falls over. Always.
Right now there’s this “orchestrator” nonsense. Cool in principle, but as someone who made scripts to automate with all the time before it’s not impressive. Spent $200 to automate doing some bug finding and fixing. It found and fixed the easy stuff (still pretty neat), and then “partially verified” it fixed the other stuff.
The “partial verification” was it justifying why it was okay it was broken.
The company has mandated we use this technology. I have an “AI Native” rating. We’re being told to put out at least 28 commits a month. It’s nonsense.
They’re letting me play with an expensive, super-high-level, probabilistic language. So I’m having a lot of fun. But I’m not going to lie, I’m very disappointed. Got this job a year ago. 12 years programming experience. First big tech job. Was hoping to learn a lot. Know my use of data to prioritize work could be better. Was sold on their use of data. I’m sure some teams here use data really well, but I’m just not impressed.
And I’m not even getting into the people gaming the metrics to look good while actually making more work for everyone else.
Lol its gonna take longer than it should for this to play out.
Sunk cost fallacy is very real, for all involved. Especially the model producers and their investors.
Sunk cost fallacy is also real for dev's who are now giving up how they used to work - they've made a sunk investment in learning to use LLMs etc. Hence the 'there's no going back' comments that crop up on here.
As I said in this thread - anyone who can think straight - Im referring to those who adhere to fundamental economic principles - can see what's going on from a mile away.
Management is just stupid sometimes. We had a similar metric at my last company and my manager's response was "well how else are we supposed to measure productivity?", and that was supposed to be a legitimate answer.
A tool with a mediocre level of skill in everything looks mediocre when the backdrop is our own area of expertise and game changing when the backdrop is an unfamiliar one. But I suspect the real game changer will be that everyone is suddenly a polymath.
1) Not a throaway, can't remember what my old account is called
2) Feel free to screen shot. Stick it on your desktop and set a reminder and check the state of the world in 12 months time.
For some of us, the world has already changed drastically. I am shipping more code, better code, less buggy code WAY faster than ever before. Big systemic changes for the better to our infra as well. There are days where I easily do 2 weeks worth of my best work ever.
I totally understand that not everyone is having that experience. And yet until people live it, it seems they just discount the experience others are having.
Cool, and you're doing it on top of the single largest IP hijacking in the history of the world, a massive uptick in infra spend and energy burn to "just throw more compute" at it instead of figuring out how to throw "the right compute at it", cannibalization of the onboarding graduates, and losing having enough friction to keep you from running off after what's probably a bad idea on further analysis, because you can crank this out in a weekend. Last time somewhat did that, we got fucking JS. We still haven't rid ourselves of it.
12 months I won't be surprised if there's not much change. But in 5 years? 10? Anything can happen. It is presumptuous to think you can project the future capabilities of this technology and confidently state that labour markets will never be affected.
Guys like you dont get it. You think OAI, Amazon etc can freely put large amounts of money into this for 5-10 years? Lmao - delusional. Investors are impatient. Show huge jumps in revenue this year or you no longer have permission to put monumental amounts of money into this anymore.
Short of that they'll just destroy the stock price by selling off; leaving employees who get paid via SBC very unhappy.
Whatever you want to say about other companies, Amazon (and Meta) is quite willing to spend many years pouring billions into technology they think will pay off later.
I dont use LLMs much. When I do, the experience always feels like search 2.0. Information at your fingertips. But you need to know exactly what you're looking for to get exactly what you need. The more complicated the problem, the more fractal / divergent outcomes there are. (Im forming the opinion that this is going to be the real limitations of LLMs).
I recently used copilot.com to help solve a tricky problem for me (which uses GPT 5.1):
I have an arbitrary width rectangle that needs to be broken into smaller
random width rectangles (maintaining depth) within a given min/max range.
The first solution merged the remainder (if less than min) into the last rectangle created (regardless if it exceeded the max).
So I poked the machine.
The next result used dynamic programming and generated every possible output combination. With a sufficiently large (yet small) rectangle, this is a factorial explosion and stalled the software.
So I poked the machine.
I realized this problem was essentially finding the distinct multisets of numbers that sum to some value. The next result used dynamic programming and only calculated the distinct sets (order is ignored). That way I could choose a random width from the set and then remove that value. (The LLM did not suggest this). However, even this was slow with a large enough rectangle.
So I poked my brain.
I realized I could start off with a greedy solution: Choose a random width within range, subtract from remaining width. Once remaining width is small enough, use dynamic programming. Then I had to handle the edges cases (no sets, when it's okay to break the rules.. etc)
So the LLMs are useful, but this took 2-3 hours IIRC (thinking, implementation, testing in an environment). Pretty sure I would have landed on a solution within the same time frame. Probably greedy with back tracking to force-fit the output.
Most of these are new features, but then they have to integrate with the existing software so it's not really greenfield. (Not to mention that our clients aren't getting any faster at approving new features, either.)
Did you train a self-hosted/open source LLM on your existing software and documentation? That should make it far more useful. It's not claude code, but some of those models are 80% there. In 6 months they'll be today's claude code.
Its this kind of thinking that tells me people cant be trusted with their comments on here re. "Omg I can produce code faster and it'll do this and that".
No simply 'producing a feature' aint it bud. That's one piece of the puzzle.
I've taken to calling LLMs processors. A "Hello World" in assembly is about 20 lines and on par with most unskilled prompting. It took a while to get from there to Rust, or Firefox, or 1T parameter transformers running on powerful vector processors. We're a notch past Hello World with this processor.
The specific way it applies to your specific situation, if it exists, either hasn't been found or hasn't made its way to you. It really is early days.
I'm working on a project right now, that is heavily informed by AI. I wouldn't even try it, if I didn't have the help. It's a big job.
However, I can't imagine vibe-coders actually shipping anything.
I really have to ride herd on the output from the LLM. Sometimes, the error is PEBCAK, because I erred, when I prompted, and that can lead to very subtle issues.
I no longer review every line, but I also have not yet gotten to the point, where I can just "trust" the LLM. I assume there's going to be problems, and haven't been disappointed, yet. The good news is, the LLM is pretty good at figuring out where we messed up.
I'm afraid to turn on SwiftLint. The LLM code is ... prolix ...
All that said, it has enormously accelerated the project. I've been working on a rewrite (server and native client) that took a couple of years to write, the first time, and it's only been a month. I'm more than half done, already.
To be fair, the slow part is still ahead. I can work alone (at high speed) on the backend and communication stuff, but once the rest of the team (especially shudder the graphic designer) gets on board, things are going to slow to a crawl.
> However, I can't imagine vibe-coders actually shipping anything.
I'm a vibe-coder, and I've shipped lots! The key is to vibe-code apps that has a single user (me). Haven't coded anything for 15 years prior to January too.
>> I no longer review every line, but I also have not yet gotten to the point, where I can just "trust" the LLM.
Same here. This is also why I haven't been able to switch to Claude Code, despite trying to multiple times. I feel like its mode of operation is much more "just trust to generated code" than Cursor, which let's you review and accept/reject diffs with a very obvious and easy to use UX.
Most of the folks I work with who uninstalled Cursor in favor of Claude Code switched back to VSCode for reviewing stuff before pushing PRs. Which... doesn't actually feel like a big change from just using Cursor, personally. I tried Claude Code recently, but like you preferred the Cursor integration.
I don't have the bandwidth to juggle four independent things being worked on by agents in parallel so the single-IDE "bottleneck" is not slowing me down. That seems to work a lot better for heavy-boilerplate or heavy-greenfield stuff.
I am curious about if we refactored our codebase the right way, would more small/isolatable subtasks be parallelizable with lower cognitive load? But I haven't found it yet.
From my experience as a software engineer, doubling my productivity hasn’t reduced my workload. My output per hour has gone up, but expectations and requirements have gone up just as fast. Software development is effectively endless work, and AI has mostly compressed timelines rather than reduced total demand.
This seems unlikely. My company is in competition with a number of other startups. If AI removes one of my co-workers, our competitors will keep the co-worker and out-compete us.
It depends on the "shape" of the company. Larger companies have a lot more of what I call "Conway Overhead", basically a mix of legit coordination overhead and bureaucracy. Startups by necessity have a lot less of that, and so are better "shaped" to fully harness AI.
This is already happening. Fewer people are getting hired. Companies are quietly (sometimes not, like Block) letting people go. At a personal level all the leaders in my company are sounding the “catch up or you’ll be left behind” alarm. People are going to be let go at an accelerated pace in the future (1-3 years).
I don’t think that addresses my point. I understand a lot of companies are firing under the guise of AI, but it’s unclear to me whether AI is actually driving this - especially when the article we are both responding to says:
> We find no systematic increase in unemployment for highly exposed workers since late 2022
> If AI removes one of my co-workers, our competitors will keep the co-worker and out-compete us.
This assumes that the companies' business growth is a function of the amount of code written, but that would not make much sense for a software company.
Many companies (including mine) are building our product with an engineering team 1/4 the size of what would have been required a few years ago. The whole idea is that we can build the machine to scale our business with far fewer workers.
How many companies have you worked at in the past where the backlog dried up and the engineering team sat around doing nothing?
Even in companies that are no longer growing I've always seen the roadmap only ever get larger (at that point you get desperate to try to catch back up, or expand into new markets, while also laying people off to cut costs).
Will we finally out-write the backlog of ideas to try and of feature requests? Or will the market get more fragmented as more smaller competitors can carve out different niches in different markets, each with more-complex offerings than they could've offered 5 years ago?
That's not necessarily a result of AI, you also have to consider the broader economic environment. I mean, it was also difficult to get a job as a graduate in 2008, whereas it's typically been easier to get a job when credit is cheap.
Isn't it, for something like 70-80% of families? Just in slow-motion?
How long have we been hearing about crushing affordability problems for property? And how long ago did that start moving into essentials? The COVID-era bullwhip-effect inflation waves triggered a lot of price ratcheting that has slowed but never really reversed. Asset prices are doing great, as people with money continue to need somewhere to put it, and have been very effective at capturing greater and greater shares of productivity increases. But how's the average waiter, cleaning-business sole-proprietor, uber driver, schoolteacher, or pet supply shopowner doing? How's their debt load trending? How's their savings trending?
There’s a difference between a collapse and a slowdown. We don’t need a collapse for hiring to slow down [1,2]. I think we’re finally just seeing the maturation of software development. Software is increasingly a commodity, so maybe the era of crazy growth and hiring is over. I don’t think that we need AI to explain this either, although possibly AI will simply commodify more kinds of software.
FAANG realizing that they can't make infinite money by expanding into every possible market while paying FAANG salaries for low-scale-CRUD-prototyping roles has a lot to do with this, and that started a bit earlier than the AI wave.
Lots going on right now in the market, but IMO that retreat is the biggest one still.
Many companies were basically on a path of infinite hiring between ~2011 and ~2022 until the rapid COVID-era whiplash really drove home "maybe we've been overhiring" and caused the reaction and slowdown that many had been predicting annually since, oh, 2015.
Manager gigs at FAANG are pretty rough right now in my network, you can't be a manager when the higher-ups notice your group isn't a big revenue generator and so doesn't justify new hires and bigger org charts, and cutting the middlemen is the easiest way to juice the ROI numbers. If the ICs that now have 1/3 the managerial structure and have to wear more hats don't turn things around, oh well, it's not a critical area anyway, just nuke it.
Erm its been fucked for many years across many professions, it was just less so for software engineering in particular. Now entry into the S-E profession is taking a hit.
Also dont forget theres only so many viable revenue-generating and cost-saving projects to take. And said above - overhiring in COVID.
There's definitely tone deaf statements from managers/leaders like "AI will allow us to do more with less headcount!" As if the end worker is supposed to be excited about that, knuckleheads, lol.
Yeah I’ve been scratching my head about this too. Like, if my boss said this, I would basically start looking for a new job right then and there. Seems like a good way to drive off your own talent.
In a bear market in a bloated company, maybe. We’re still actively hiring at my startup, even with going all-in on AI across the company. My PM is currently shipping major features (with my review) faster and with higher-quality code than any engineer did last year.
It’s hard to compare, honestly. Last year, my PM didn’t have the AI tools to do any of this, and engineers were spread thin. Now, the PM (with a specialized Claude Code environment) has the enthusiasm of a new software engineer and the product instincts of a senior PM.
Then any company that was staffed at levels needed prior to the arrival of current-level LLM coding assistants is bloated.
If the company was person-hour starved before, a significant amount of that demand is being satisfied by LLMs now.
It all depends on where the company is in the arc of its technology and business development, and where it was when powerful coding agents became viable.
When you work long enough you'll find it. Places where changing software is risky you can end up waiting for approvals. Places where another company purchased yours or you are getting shutdown soon and there is no new work. Sometimes you end up on a system that they want to replace but they never get around to it.
Being overworked is sometimes better than being underworked. Sometimes the reserve is better. They both have challenges.
Outside of purchased-and-being-shutdown, these are still frequently "we want to do things but we're scared of breaking things" situations, not "we don't want to do anything." Even if the things they want to do are just "we want to move off this 90s codebase before everyone who knows how it works is dead."
In that sort of high-fear, change-adverse environment "get rid of all the devs and let the AI do it" may not be the most compelling sales pitch to leadership. ("Use it to port the code faster so we can spend more time on the migration plan and manual testing" might have better luck.)
That’s the economy in general. Labor saving innovations increase productivity but do not usually reduce work very much, though they can shift it around pretty dramatically. There are game theoretic reasons for this, as well as phenomena like the hedonic treadmill.
Ideal state for every company is to have minimum input costs with maximum output costs. Labor always gets cut out of the loop because it’s one of the most expensive input costs.
I don't think there's been much of an impact, really. Those who know how to use AI just got tangentially more productive (because why would you reveal your fake 10x productivity boost so your boss hands you 10x more tasks to finish?), and those w/o AI knowledge stayed the way they were.
The real impact is for indie-devs or freelancers but that usually doesn't account for much of the GDP.
Management often has a perverse short-term incentive to make labor feel insecure. It’s a quick way to make people feel insecure and work harder ... for a while.
Also, “AI makes us more productive so we can cut our labor costs” sounds so much better to investors than some variation of “layoffs because we fucked up / business is down / etc”
You should look into the concepts of skepticism, materialism, and cynicism. Maybe don't trust the leadership of where you work, the leadership that sees you as a number and not a human.
I also thought it was hilarious that they invented a brand new metric that (surprise) makes their product’s long term projection look really good (financially).
the numbers they show are barely distinguishable from noise as far as I can interpret them.
For me, the impact is absolutely in hiring juniors. We basically just stopped considering it. There's almost no work a junior can do that now I would look at and think it isn't easier to hand off in some form (possibly different to what the junior would do) to an AI.
It's a bit illusory though. It was always the case that handing off work to a junior person was often more work than doing it yourself. It's an investment in the future to hire someone and get their productivity up to a point of net gain. As much as anything it's a pause while we reassess what the shape of expertise now looks like. I know what juniors did before is now less valuable than it used to be, but I don't know what the value proposition of the future looks like. So until we know, we pause and hold - and the efficiency gains from using AI currently are mostly being invested in that "hold" - they are keeping us viable from a workload perspective long enough to restructure work around AI. Once we do that, I think there will be a reset and hiring of juniors will kick back in.
If AI increases productivity, and juniors are cheaper to hire, but is just as able to hand off tasks to ai as a senior, then it makes more sense to hire more juniors to get them working with an AI as soon as possible. This produces output faster, for which more revenue could be derived.
So the only limiting factor is the possibility of not deriving more revenue - which is not related to the AI issue, but broader, macroeconomic issue(s).
Juniors are not as capable of delegating to AI as seniors are. Delegation to AI requires code review, catching the AI when it doesn’t follow good engineering practices, and catching the AI in semantic mistakes due to the AIs lack of broader context. Those things are all hard for juniors.
I think this is the crux of it. Someone who doesn't know the right thing to do just isn't in a position to hand off anything. Accelerating their work will just make them do the wrong thing faster.
One of the more interesting takes I heard from a colleague, who’s in the marketing department, is that he uses the corporate approved LLM (Gemini) for “pretend work” or very basic tasks. At the same time he uses Claude on his personal account to seriously augment his job.
His rationale is he won’t let the company log his prompts and responses so they can’t build an agentic replacement for him. Corporate rules about shadow it be damned.
I'm not really concerned about the availability of SW dev jobs, but I am concerned about the quality of them. For many companies the velocity (and quality, much to my chagrin) of the code you can produce doesn't really matter. What matters more is whether or not you're building the right thing, and too often you're not. These companies also tend to keep more headcount than seems justified, I think because they are gambling that a few employees are going to do something awesome but they don't know which ones. As AI gets better what will these companies do? I don't think they will fire a bunch of SW devs. I think instead they will embrace the slop and just take more shots, and crazier shots. It doesn't just give us something to do, it also gives a bunch of PHBs something to do.
My day to day is even busier now with agents all over the place making code changes. The Security landscape is even more complex now overnight. The only negative impact I see is that there’s not much need for junior devs right now. The agent fills that role in a way. But we’ll have to backfill some way or another.
I know multiple devs who would have a very large productivity increase but instead choose to slow down their output on purpose and play video games instead.
I get it.
The problem with using unemployment as a metric is hiring is driving by perception. You're making an educated guess as to how many people you need in the future.
Between 2004 and 2008 I did many things in computing as a company that offered my services, one of these was information gathering automation. It almost never immediately lead to decreases in employment. The systems had a to be in place for a while, people had to get used to them, people had to stop making common mistakes with them.
Then the 2008 crash happened and those people were gone in a blink of an eye and never replaced. The companies grew in staff after that, but it was in things like sales and marketing.
I'm an SDE with 1 YOE using AI tools heavily (doing "day's work" in ~2 hrs, perfect reviews). Spending most time on specs/review vs. raw coding. Worried I'm optimising short-term output over long-term skill development. Should I consider pivoting to AI/ML roles? Would love advice from anyone who's hired juniors in the current era.
Data point of 1: Having hired juniors as a startup founder, I need more generalists than AI/ML specialists. AI application work right now is basically standard software engineering - you’re finding clever ways to supply the right context to a model within certain constraints.
No one knows what’s going to happen in the future. Yes there already are fewer SWE jobs than before because of AI, and yes the days of companies hiring new grads in droves at $300k+ packages are likely over. IMO all you can really do is study what you’re interested in, learn it deeply, and do good work with cool people. If unsure, it’s possible to go back to what you were doing before if the new path doesn’t work out.
Shipping speed never/is was the issue. Most companies are terrible at figuring out what exactly they should be allocating resources behind.
Speeding up does not solve the problem that most humans who are at the top of the hierarchy are poor thinkers. In fact it compounds it. More noise, nice.
Apple has already shown this decades ago - they got the iPhone and iPod developed and out the door in relatively short-time scales given the impact of the products on the world.
Once you know what you want, exactly what you want, things moves fast - really fast.
Or worse. I’ve heard stories from friends where leadership expects huge boosts in productivity due to LLMs, and perceive anything but an order of magnitude boost as incompetence or a refusal to adapt.
this keeps me up at night. i’m in a role that is essentially deployment management for LLMs at faang esque company. very little coding or need to code, mostly navigating guis, pipelines, and docker to get deployments updated with a new venting or model version or some patch
> There's suggestive evidence that hiring of young workers (ages 22–25) into exposed occupations has slowed — roughly a 14% drop in the job-finding rate
There goes my excuse of not finding a job in this market.
A possible outcome of AI: domestic technical employment goes up because the economics of outsourcing change. Domestic technical workers working with AI tools can replace outsourcing shops, eliminating time-shift issues, etc at similar or lower costs.
This rhymes with another recent study from the Dallas Fed: https://www.dallasfed.org/research/economics/2026/0224 - suggests AI is displacing younger workers but boosting experienced ones. This matches what we see discussed here, as well as the couple similar other studies we've seen discussed here.
I think the underlying reason is simply because companies are "shaped wrong" to absorb AI fully. I always harp on how there's a learning curve (and significant self-adaptation) to really use AI well. Companies face the same challenge.
Let's focus on software. By many estimates code-related activities are only 20 - 60%, maybe even as low as 11%, of software engineers' time (e.g. https://medium.com/@vikpoca/developers-spend-only-11-of-thei...) But consider where the rest of the time goes. Largely coordination overhead. Meetings etc. drain a lot of time (and more the more senior you get), and those are mostly getting a bunch of people across the company along the dependency web to align on technical directions and roadmaps.
I call this "Conway Overhead."
This is inevitable because the only way to scale cognitive work was to distribute it across a lot of people with narrow, specialized knowledge and domain ownership. It's effectively the overhead of distributed systems applied to organizations. Hence each team owned a couple of products / services / platforms / projects, with each member working on an even smaller part of it at a time. Coordination happened along the heirarchicy of the org chart because that is most efficient.
Now imagine, a single AI-assisted person competently owns everything a team used to own.
Suddenly the team at the leaf layer is reduced to 1 from about... 5? This instantly gets rid of a lot of overhead like daily standups, regular 1:1s and intra-team blockers. And inter-team coordination is reduced to a couple of devs hashing it out over Slack instead of meetings and tickets and timelines and backlog grooming and blockers.
So not only has the speed of coding increased, the amount of time spent coding has also gone up. The acceleration is super-linear.
But, this headcount reduction ripples up the org tree. This means the middle management layers, and the total headcount, are thinned out by the same factor that the bottom-most layer is!
And this focused only on the engineering aspect. Imagine the same dynamic playing out across departments when all kinds of adjacent roles are rolled up into the same person: product, design, reliability...
These are radical changes to workflows and organizations. However, at this stage we're simply shoe-horning AI into the old, now-obsolete ticket-driven way of doing things.
So of course AI has a "capability overhang" and is going to take time to have broad impact... but when it does, it's not going to be pretty.
The TL;DR is that there is little measurable impact (and I'd personally add "yet").
To quote:
"We find no systematic increase in unemployment for highly exposed workers since late 2022, though we find suggestive evidence that hiring of younger workers has slowed in exposed occupations"
My belief based on personal experience is that in software engineering it wasn't until November/December 2025 that AI had enough impact to measurably accelerate delivery throughout the whole software development lifecycle.
I have doubts that this impact is measurable yet - there is a lag between hiring intention and impact on jobs, and outside Silicon Valley large scale hiring decisions are rarely made in a 3 month timeframe.
The most interesting part is the radar plot showing the lack of usage of AI in many industries where the capability is there!
> My belief based on personal experience is that in software engineering it wasn't until November/December 2025 that AI had enough impact to measurably accelerate delivery throughout the whole software development lifecycle.
Gemini 3 and Opus 4.6 were the "woah, they're actually useful now!" moment for me.
I keep saying to colleagues that it's like a rising tide. Initially the AIs were lapping around our ankles, now the level of capability is at waist height.
Many people have commented that 50% of developers think AI-generated code is "Great!" and 50% think its trash. That's a sign that AI code quality is that of the median developer. This will likely improve to 60%-40%, then 70%-30%, etc...
I don’t see definitive evidence that there is some kind of Moore’s law for model improvement though. Just because this year’s model performs better than last year’s model doesn’t mean next year’s model will be another leap. Most of the big improvements this year seem to be around tooling - I still see Opus 4.6 (which is my daily driver at work) making lots of mistakes.
I really hate to say it, but this article in particular needs a tldr. The author does a web recipe take. Don't put the actual factual info upfront and require parsing through everything to find anything important.
Kinda done with this.
If you have something important to say, say it up front and back it up with literature later.
My alternative? Nationalize the company and implement a workplace democracy to replace the executive team + board.
I trust the workers more to dictate the direction of a company than most executives.
They can't do worse.
edit: or what another commentator said, fucking academia. Public universities have done more for humanity than nearly anything to come out of SV. Surveillance capitalism, mass misery + psychosis; it's very telling what our society values when mass amounts of the Earth are desperately trying to ban these very same services to protect children.
Just like anything else, you either think "that's definitely wrong" or "huh, I guess that's probably it." If its really serious, you have to pause and make of a judgement call of course.
There was a recent anecdote from the head of radiology, Mayo Clinic I believe, that went something like this:
- AI has allowed radiologists to review a much higher rate of x-rays
- The above has led to a dramatic increase in need for faster processing, more storage of scans etc
- which in turn led to needing a bigger IT department to manage all of the additional workload
There was a similar anecdote about the IRS where the claim is they went from having N accountants to having much fewer accountants but now they need N IT people to manage the new systems.
Because, I am terrified by the output I am getting while working on huge legacy codebases, it works. I described one of my workflow changes here: https://news.ycombinator.com/item?id=47271168 but in general compared to old way of working I am saving half of the steps consistently, whether its researching the codebase, or integrating new things, or even making fixes. I have stopped writing code, occasionally I jump into the changes proposed by LLM and make manual edits if it is feasible, otherwise I revert changes and ask it to generate again but based on my learnings from the past rejected output
I am terrified about what's coming
To give an example from a field where LLMs started causing employment worries earlier than software development: translation. Some translators made their living doing the equivalent of routine, repetitive coding tasks: translating patents, manuals, text strings for localized software, etc. Some of that work was already threatened by pre-LLM machine translation, despite its poor quality; context-aware LLMs have pretty much taken over the rest. Translators who were specialized in that type of work and too old or inflexible to move into other areas were hurt badly.
The potential demand for translation between languages has always been immense, and until the past few years only a tiny portion of that demand was being met. Now that translation is practically free, much more of that demand is being met, though not always well. Few people using an app or browser extension to translate between languages have much sense of what makes a good translation or of how translation can go bad. Professional translators who are able to apply their higher-level knowledge and language skills to facilitate intercultural communication in various ways can still make good money. But it requires a mindset change that can be difficult.
On a macro level, if you were in a rising economic tide, you would still be hiring, and turning those productivity gains into more business.
I wonder what the parallels are to past automations. When part producing companies moved from manual mills to CNC mills, did they fire a bunch of people or did they make more parts?
It’s refreshing to see the same sentiment from so many other people independently here.
That's one of the reasons why I am terrified, because it can lead to burn out, and I personally don't like to babysit bunch of agents, because the output doesn't feel "mine", when its not "mine" I don't feel ownership.
And I am deliberately hitting the brake from time to time not to increase expectations, because I feel like driving someone else's car while not understanding fully how they tuned their car (even though I did those tunings by prompting)
It's not going to be written exactly like you would do it, but that's ok - because you care about the results of the solution and not its precise implementation. At some point you have to make an engineering decision whether to write it yourself for critical bits or allow the agent/junior to get a good enough result.
You're reviewing the code and hand editing anyway, right? You understand the specs even if your agent/junior doesn't, so you can take credit even if you didn't physically write the code. It's the same thing.
Yes, yes!
And this is problem for me, because of the pace, my brain muscles are not developing enough compared to when I was doing those things myself.
before, I was changing my mind while implementing the code, because I see more things while typing, and digging deeper, but now, because juniors are doing things they don't offer me a refactoring or improvements while typing the code quickly, because they obey my command instead of having "aha" moment to suggest better ways
At review time.
There are simply too many software industries that can't delegate both authorship _and_ review to non-humans because the maintenance/use of such software, especially in libraries and backwards-compat-concerning environments, cannot justify an "ends justifies the means" approach (yet).
I also heard an opinion that since writing code is cheap, people implement things that have no economic value without really thinking it through.
Only it doesn't, there's product positioning, UX, information architecture, onboarding and training, support, QA, change management, analytics, reporting… sigh
The question is really, velocity _of what_?
I got this from a HN comment. It really hit for me because the default mentality for engineers is to build. The more you build the better. That's not "wrong" but in a business setting it is very much necessary but not sufficient. And so whenever we think about productivity, impact, velocity, whatever measure of output, the real question is _of what_? More code? More product surface area? That was never really the problem. In fact it makes life worse majority of the time.
They've already admitted they just 'throw the code away and start again'.
I think we've got another victim of perceived productivity gains vs actual productivity drop.
People sitting around watching Claude churn out poor code at a slower rate than if they just wrote it themselves.
Don't get me wrong, great for getting you started or writing a little prototype.
But the code is bad, riddled with subtle bugs and if you're not rewriting it and shoving large amounts of AI code into your codebase, good luck in 6-12 months time.
I'm going to guess that many people on HN are also on the "free ChatGPT isn't that good at programming" train.
Probably that's the reason why some people are sure their job is still safe.
Nature of job is changing rapidly
I'm not a tech CEO but people who are anti-LLM for programming have no place on my team.
There’s obviously a benefit of paying higher rates for US programmers, but does that benefit change when llms are thrown into the mix
Someone deciding to drop a spreadsheet of customer data into their personal AI account to increase their productivity would be catastrophic for business, so you need rules. And rules means paying for enterprise AI tooling.
Til then wtf_are_these_abstractions.jpg
Edit: Ha, and the report claims it's relatively good at business and finance...
Edit 2: After discussion in this thread, I went back to opus and asked it to link to articles about how to handle non-normally distributed data, and it actually did link to some useful articles, and an online calculator that I believe works for my data. So I'll eat some humble pie and say my initial take was at least partially wrong here. At the same time, it was important to know the correct question to ask, and honestly if it wasn't for this thread I'm not sure I would have gotten there.
A good way to use AI is to treat it like a brilliant junior. It knows a lot about how things work in general but very little about your specific domain. If your data has a particular shape (e.g lots of orders with a few large orders as outliers) you have to tell it that to improve the results you get back.
- https://cxl.com/blog/outliers/
- https://www.blastx.com/insights/the-best-revenue-significanc...
- (online tool to calculate significance) https://www.blastx.com/rpv-calculator
I'm not checking their math, but the articles make sense to me, and I trust they did implement it correctly. In the end the LLM did get me to the correct answer by suggesting the articles, so I guess I should eat some humble pie and say it _did_ help me. At the same time, if I didn't have the intuition that using rpv as-is in a t-test would be noisy, and the suggestions from this comment thread, I think I could have gone down the wrong path. So I'm not sure what my conclusion is -- maybe something like LLMs are helpful once you ask the right question.
Tests were always important, but now they are the gatekeepers to velocity.
It’s fine at replacing what stack overflow did nearly a decade ago, but that isn’t really an improvement from my baseline.
Yesterday a colleague didn’t quite manage to implement a loading container with a Vue directive instead of DOM hacks, it was easier for me to just throw AI at the problem and produced a working and tested solution and developer docs than to have a similarly long meeting and have them iterate for hours.
Then I got back to training a CNN to recognize crops from space (ploughing and mowing will need to be estimated alongside inference, since no markers in training data but can look at BSI changes for example), deployed a new version of an Ollama/OpenAI/Anthropic proxy that can work with AWS Bedrock and updated the docs site instructions, deployed a new app that will have a standup bot and on-demand AI code review (LiteLLM and Django) and am working on codegen to migrate some Oracle forms that have been stagnating otherwise.
It’s not funny how overworked I am and sure I still have to babysit parallel Claude Code sessions and sometimes test things manually and write out changes, but this is a completely different work compared to two or three years ago.
Maybe the problem spaces I’m dealing with are nothing novel, but I assume most devs are like that - and I’d be surprised at people’s productivity not increasing.
When people nag in meetings about needing to change something in a codebase, or not knowing how to implement something and its value add, I’ll often have something working shortly after the meeting is over (due to starting during it).
Instead of sending adding Vitest to the backlog graveyard, I had it integrated and running in one or two evenings with about 1200 tests (and fixed some bugs). Instead of talking about hypothetical Oxlint and Oxfmt performance improvements, I had both benchmarked against ESLint and Prettier within the hour.
Same for making server config changes with Ansible that I previously didn’t due to additional friction - it is mostly just gone (as long as I allow some free time planned in case things vet fucked up and I need to fix them).
Edit: oh and in my free time I built a Whisper + VLM + LLM pipeline based on OpenVINO so that I can feed it hours long stream VODs and get an EDL cut to desired length that I can then import in DaVinci Resolve and work on video editing after the first basic editing prepass is done (also PyScene detect and some audio alignment to prevent bad cuts). And then I integrated it with subscription Claude Code, not just LiteLLM and cloud providers with per-token costs for the actual cuts making part (scene description and audio transcriptions stay local since those don't need a complex LLM, but can use cloud for cuts).
Oh and I'm moving from my Contabo VPSes to running stuff inside of a Hetzner Server Auction server that now has Proxmox and VMs in that, except this time around I'm moving over to Ansible for managing it instead of manual scripts as well, and also I'm migrating over from Docker Swarm to regular Docker Compose + Tailscale networks (maybe Headscale later) and also using more upstream containers where needed instead of trying to build all of mine myself, since storage isn't a problem and consistency isn't that important. At the same time I also migrated from Drone CI to Woodpecker CI and from Nexus to Gitea Packages, since I'm already using Gitea and since Nexus is a maintenance burden.
If this becomes the new “normal” in regards to everyone’s productivity though, there will be an insane amount of burnout and devaluation of work.
We've started building harnesses to allow people who don't understand code to create PRs to implement their little nags. We rely on an engineer to review, merge, and steward the change but it means that non-eng folks do not rely on us as a gate. (We're a startup and can't really afford "teams" to do this hand-holding and triage for us.)
As you say we're all a bit overworked and burned out. I've been context switching so much that on days when I'm very productive I've started just getting headaches. I'm achieving a lot more than before but holding the various threads on my head and context switching is just a lot.
I've always done more in days than others might in a week. YMMV.
The productivity gains are blatantly obvious at this point. Even in large distributed code bases. From jr to senior engineer.
Productivity is a term of art in economics and means you generate more units of output (for example per person, per input, per wages paid) but doesn't take quality or otherwise desireability into account. It's best suited for commodities and industrial outputs (and maybe slop?).
Note1: I have "expert" level research skills. But LLMs still help me in research, but the boost is probably 1.2x max. But
Note2: By research, I mean googling, github search, forum search, etc. And quickly testing using jsfiddle/codepen, etc.
But I also think you are overestimating your RESEARCH skills, even if you are very good at research, I am sure you can't read 25 files in parallel, summarize them (even if its missing some details) in 1 minute and then come up with somewhat working solution in the next 2 minutes.
I am pretty sure, humans can't comprehend reading 25 code files with each having at least 400 lines of non-boilerplate code in 2 minutes. LLM can do it and its very very good at summarizing.
I can even steer its summarizing skills by prompting where to focus on when its reading files (because now I can iterate 2-3 times for each RESEARCH task and improve my next attempt based on shortcomings in the previous attempt)
* you probably lack good RESEARCH skills
* I can see at most 1.25x improvements - now it is 2-3x
By updating your comment you are making my reply irrelevant to your past response
Every time I say this people get really angry, but: so far AI has had almost no impact on my job. Neither my dev team nor my vendors are getting me software faster than they were two years ago. Docker had a bigger impact on the pipeline to me than AI has.
Maybe this will change, but until it does I'm mostly watching bemusedly.
I've seen lots of people say AI can basically code a project for them. Maybe it can, but that seems to heavily depend on the field. Other than boilerplate code or very generic projects, it's a step above useless imo when it comes to gamedev. It's about as useful as a guy who read some documentation for an engine a couple years ago and kind of remembers it but not quite and makes lots of mistakes. The best it can do is point me in the general direction I need to go, but it'll hallucinate basic functions and mess up any sort of logic.
Am trying to compare this to reports that people are not reviewing code any more.
1) Do that inside their IDEs, which is less funny
2) Generate blog post about it instead of memes
For me this is a huge boost in productivity. If I remember how I was working in the past (example of Google integration):
Before:
Now:The docs used to be good enough that there would be an example which did exactly what you needed more often than the llm gets it right today.
A famous economist once said, "You can see the computer age everywhere but in the productivity statistics."
There are many reasons for the lag in productivity gain but it certainly will come.
https://en.wikipedia.org/wiki/Productivity_paradox
The loss of skills, complete loss of visibility and experience with the codebase, and the complete lack of software architecture design, seems like a massive killer in the long term
I have a feeling that we're going to see productivity with AI drop through the floor
Just yesterday I asked Opus 4.6 what I could do to make an old macOS AppKit project more testable, too lazy to even encumber the question with my own preferences like I usually do, and it pitched a refactor into Elm architecture. And then it did the refactor while I took a piss.
The idea that AI writes bad software or can't improve existing software in substantial ways is really outdated. Just consider how most human-written software is untested despite everyone agreeing testing is a good idea simply because test-friendly arch takes a lot of thought and test maintenance slow you down. AI will do all of that, just mention something about 'testability' in AGENTS.md.
I'm happy to sell it to you, though it is also free. I guided Claude to write this in three weeks, after never having written a line of JavaScript or set up a server before. I'm sure a better JavaScript programmer than I could do this in three weeks, but there's no way I could. I just had a cool idea for making advertising a force for good, and now I have a working version in beta.
I'd say it is better software, but better is doing a lot of heavy lifting there. Claude's execution is average and always will be, that's a function of being a prediction engine. But I genuinely think the idea is better than how advertising works today, and this product would not exist at all if I had to write it myself. And I'm someone who has written code before, enough that I was probably a somewhat early adopter to this whole thing. Multiply that by all the people whose ideas get to live now, and I'm sure some ideas will prove to be better even with average execution. Like an llm, that's a function of statistics.
Why would you think you have your finger on the pulse of general software trends like that when you use the same, what, dozen apps every week?
Just looking at my own productivity, as mere sideprojects this month, I've shipped my own terminal app (replaced iTerm2), btrfs+luks NAS system manager, overhauled my macOS gamepad mapper for the app store, and more. All fully tested and really polished, yet I didn't write any code by hand. I would have done none of that this month without AI.
You'd need some real empirics to pick up productivity stories like mine across the software world, not vibes.
Remember this famously happened before, in the 1970s
Predictions without a deadline are unfalsifiable.
I can turn out some scripts a little bit quicker, or find an answer to something a little quicker than googling, but I'm still waiting on others most of the time, the overall company processes haven't improved or gotten more efficient. The same blockers as always still exist.
Like you said, there has been other tech that has changed my job over time more than AI has. The move to the cloud, Docker, Terraform, Ansible, etc. have all had far more of an impact on my job. I see literally zero change in the output of others, both internally and externally.
So either this is a massively overblown bubble, or I'm just missing something.
And that's the key problem, isn't it? I maintain current organizations have the "wrong shape" to fully leverage AI. Imagine instead of the scope of your current ownership, you own everything your team or your whole department owns. Consider what that would do to the meetings and dependencies and processes and tickets and blockers and other bureaucracy, something I call "Conway Overhead."
Now imagine that playing out across multiple roles, i.e. you also take on product and design. Imagine what that would do to your company org chart.
I added a much more detailed comment here: https://news.ycombinator.com/item?id=47270142
> Now imagine
> Imagine what that would do
Imagine if your grandma had wheels! She'd be a bicycle. Now imagine she had an engine. She could be a motorcycle! Unfortunately for grandma, she lives in reality and is not actually a motorcycle, which would be cool as hell. Our imagination can only take us so far.
To more substantively reply to your longer linked comment: your hypothesis is that people spend as little as 10% of time coding and the other 90% of time in meetings, but that if they could code more, they wouldn't need to meet other people because they could do all the work of an entire team themselves[1]. The problem with your hypothesis is that you take for granted that LLMs actually allow people to do the work of an entire team themselves, and that it is merely bureacracy holding them back. There have been absolutely zero indicators that this is true. No productivity studies of individual developers tackling tasks show a 10x speedup; results tend to be anywhere from +20% to minus 20%. We aren't seeing amazing software being built by individual developers using LLMs. There is still only one Fabrice Bellard in the world, even though if your premise could escape the containment zone of imagination anyone should be able to be a Bellard on their own time with the help of LLMs.
[1] Also, this is basically already true without LLMs. It is the reason startups are able to disrupt corporate behemoths. If you have just a small handful of people who spend the majority of their work time writing code (by hand! No LLMs required!), they can build amazing new products that outcompete products funded by trillion-dollar entities. Your observation of more coding = less meetings required in the first place has an element of truth to it, but not because LLMs are related to it in any particular way.
Too early to say what AI brings to the efficiency table I think. In some major things I do it's a 1000x speed up. In others it is more a different way of approaching a problem than a speed up. In yet others, it is a bit of an impediment. It works best when you learn to quickly recognize patterns and whether it will help. I don't know how people who are raised with ai will navigate and leverage it, which is the real long-term question (just as the difference between pre- and post-smartphone generations is a thing).
The only study showing a -20% came back and said, "we now think it's +9% - +38%, but we can't prove rigorously because developers don't want to work without AI anymore": https://news.ycombinator.com/item?id=47142078
Even at the time of the original study, most other rigorous studies showed -5% (for legacy projects, obsolete languages) to 30% (more typical greenfield AND brownfield projects) way back in 2024. Today I hear numbers up to 60% from reports like DX.
But this is exactly missing the point. Most of them are still doing things the old way, including the very process of writing code. Which brings me to this point:
> There have been absolutely zero indicators that this is true.
I could tell you my personal experience, or link various comments on HN, or point you to blogs like https://ghuntley.com/real/ (which also talks about the origanizational impedance mismatch for AI), but actual code would be a better data point.
So there are some open-source projects worth looking at, but they are typically dismissed because they look so weird to us. Here's two mostly vibe-coded (as in, minimal code review, apparently) projects that people shredded for having weird code, but is already used by 10s of 1000s of people, up to 11 - 18K stars now. Look at the commit volume and patterns for O(300K) LoC in a couple of months, mostly from one guy and his agent:
https://github.com/steveyegge/beads/graphs/commit-activity
https://github.com/steveyegge/gastown/graphs/commit-activity
It's like nothing we've seen before, almost equal number of LoC additions and deletions, in the 100s of Ks! It's still not clear how this will pan out long term, but the volume of code and apparent utility (based purely on popularity) is undeniable.
I dismiss them because Yegge's work (if it can even be called his work, given that he doesn't look at the code) is steaming garbage with zero real-world utility, not "because they look weird". You suggest the apparent utility is undeniable, while saying "based purely on popularity" -- but popularity is in no way a measure of utility. Yegge is a conman who profited hundreds of thousands of dollars shilling a memecoin rugpull tied to these projects. The actual thousands of users are people joining the hypetrain, looking to get in on the promised pyramid scheme of free money where AI will build the next million dollar software for you, if only you have the right combination of .md files to make it work. None of these software are actually materialising, so all the people in this bubble can do is make more AI wrappers that promise to make other AI wrappers that will totally make them money.
I am completely open to being proven wrong by a vibe-coded open source application that is actually useful, but I haven't seen a single one. Literally not even one. I would count literally anything where the end-product is not an AI wrapper itself, which has tens to hundreds of thousands of users, and which was written entirely by agents. One example of that would be great. Just one. There have been a couple of attempts at a web browser, and Claude's C compiler, but neither are actually useful or have any real users; they are just proofs of concept and I have seen nothing that convinces me they are a solid foundation from which you could actually build useful software from, or that models will ever be on a trajectory to make them actually useful.
I've been in ops for 30 years, Claude Code has changed how I work. Ops-related scripting seems to be a real sweet spot for the LLMs, especially as they tend to be smaller tools working together. It can convert a few sentences into working code in 15-30 minutes while you do something else. I've given it access to my apache logs Elastic cluster, and it does a great job at analyzing them ("We suspect this user has been compromised, can you find evidence of that?"). It's quite startling, actually, what it's able to do.
The problem with analyzing logs is determinism. If I ask Claude to look for evidence of compromise, I can't trust the output without also going and verifying myself. It's now an extra step, for what? I still have to go into Elastic and run the actual queries to verify what Claude said. A saved Kibana search is faster, and more importantly, deterministic. I'm not going to leave something like finding evidence of compromise up to an LLM that can, and does, hallucinate especially when you fill the context up with a ton of logs.
An auditor isn't going to buy "But Claude said everything was fine."
Is AI actually finding things your SIEM rules were missing? Because otherwise, I just don't see the value in having a natural language interface for queries I already know how to run, it's less intuitive for me and non deterministic.
It's certainly a useful tool, there's no arguing that. I wouldn't want to go back to working with out it. But, I don't buy that it's already this huge labor market transformation force that's magically 100x everyone's productivity. That part is 100% pure hype, not reality.
Is it? A couple days ago I had it build tooling for a one-off task I need to run, it wrote ~800 lines of Python to accomplish this, in <30m. I found it was too slow, so I got it to convert it to run multiple tasks in parallel in another prompt. Would have taken a couple days for me to build from hand, given the number of interruptions I have in the average day. This isn't a one-off, it's happening all the time.
Imo it's only a matter of time as companies start to figure out how to use ai. Companies don't seem to have real plans yet and everyone is figuring out ai in general out.
Soon though I will think agents start popping up, things like first line response to pages, executing automation
Maybe 2 years ago Ai was doing random stuff and we got all those funny screenshots of dumb gemini answers. The indeterminism leading to random stuff isn't really an issue any more.
The way it thinks keeps it on track.
Humans are funny. But most cant seem to understand that the tool is a mirage and they are putting false expectations on it. E.g. management of firms cutting back on hiring under the expectation that LLMs will do magic - with many cheering 'this is the worst itll be bro!!".
I just hope more people realise before Anthropic and OAI can IPO. I would wager they are in the process of cleaning up their financials for it.
There’s some neat stuff, don’t get me wrong. But every additional tool so far has started strong but then always falls over. Always.
Right now there’s this “orchestrator” nonsense. Cool in principle, but as someone who made scripts to automate with all the time before it’s not impressive. Spent $200 to automate doing some bug finding and fixing. It found and fixed the easy stuff (still pretty neat), and then “partially verified” it fixed the other stuff.
The “partial verification” was it justifying why it was okay it was broken.
The company has mandated we use this technology. I have an “AI Native” rating. We’re being told to put out at least 28 commits a month. It’s nonsense.
They’re letting me play with an expensive, super-high-level, probabilistic language. So I’m having a lot of fun. But I’m not going to lie, I’m very disappointed. Got this job a year ago. 12 years programming experience. First big tech job. Was hoping to learn a lot. Know my use of data to prioritize work could be better. Was sold on their use of data. I’m sure some teams here use data really well, but I’m just not impressed.
And I’m not even getting into the people gaming the metrics to look good while actually making more work for everyone else.
Sunk cost fallacy is very real, for all involved. Especially the model producers and their investors.
Sunk cost fallacy is also real for dev's who are now giving up how they used to work - they've made a sunk investment in learning to use LLMs etc. Hence the 'there's no going back' comments that crop up on here.
As I said in this thread - anyone who can think straight - Im referring to those who adhere to fundamental economic principles - can see what's going on from a mile away.
Are you hiring?
IMO AI will make 70-80% job obsolete for sure.
People who actually know how to think can see it a mile away.
Job done fella.
I totally understand that not everyone is having that experience. And yet until people live it, it seems they just discount the experience others are having.
I'll take the 12 month bet.
Let us not lose sight of how we got here.
Guys like you dont get it. You think OAI, Amazon etc can freely put large amounts of money into this for 5-10 years? Lmao - delusional. Investors are impatient. Show huge jumps in revenue this year or you no longer have permission to put monumental amounts of money into this anymore.
Short of that they'll just destroy the stock price by selling off; leaving employees who get paid via SBC very unhappy.
Won't matter. The Chinese models will be running on potatoes by then and be better than ever.
I recently used copilot.com to help solve a tricky problem for me (which uses GPT 5.1):
The first solution merged the remainder (if less than min) into the last rectangle created (regardless if it exceeded the max).So I poked the machine.
The next result used dynamic programming and generated every possible output combination. With a sufficiently large (yet small) rectangle, this is a factorial explosion and stalled the software.
So I poked the machine.
I realized this problem was essentially finding the distinct multisets of numbers that sum to some value. The next result used dynamic programming and only calculated the distinct sets (order is ignored). That way I could choose a random width from the set and then remove that value. (The LLM did not suggest this). However, even this was slow with a large enough rectangle.
So I poked my brain.
I realized I could start off with a greedy solution: Choose a random width within range, subtract from remaining width. Once remaining width is small enough, use dynamic programming. Then I had to handle the edges cases (no sets, when it's okay to break the rules.. etc)
So the LLMs are useful, but this took 2-3 hours IIRC (thinking, implementation, testing in an environment). Pretty sure I would have landed on a solution within the same time frame. Probably greedy with back tracking to force-fit the output.
No simply 'producing a feature' aint it bud. That's one piece of the puzzle.
The specific way it applies to your specific situation, if it exists, either hasn't been found or hasn't made its way to you. It really is early days.
However, I can't imagine vibe-coders actually shipping anything.
I really have to ride herd on the output from the LLM. Sometimes, the error is PEBCAK, because I erred, when I prompted, and that can lead to very subtle issues.
I no longer review every line, but I also have not yet gotten to the point, where I can just "trust" the LLM. I assume there's going to be problems, and haven't been disappointed, yet. The good news is, the LLM is pretty good at figuring out where we messed up.
I'm afraid to turn on SwiftLint. The LLM code is ... prolix ...
All that said, it has enormously accelerated the project. I've been working on a rewrite (server and native client) that took a couple of years to write, the first time, and it's only been a month. I'm more than half done, already.
To be fair, the slow part is still ahead. I can work alone (at high speed) on the backend and communication stuff, but once the rest of the team (especially shudder the graphic designer) gets on board, things are going to slow to a crawl.
I'm a vibe-coder, and I've shipped lots! The key is to vibe-code apps that has a single user (me). Haven't coded anything for 15 years prior to January too.
Same here. This is also why I haven't been able to switch to Claude Code, despite trying to multiple times. I feel like its mode of operation is much more "just trust to generated code" than Cursor, which let's you review and accept/reject diffs with a very obvious and easy to use UX.
I don't have the bandwidth to juggle four independent things being worked on by agents in parallel so the single-IDE "bottleneck" is not slowing me down. That seems to work a lot better for heavy-boilerplate or heavy-greenfield stuff.
I am curious about if we refactored our codebase the right way, would more small/isolatable subtasks be parallelizable with lower cognitive load? But I haven't found it yet.
This is already happening. Fewer people are getting hired. Companies are quietly (sometimes not, like Block) letting people go. At a personal level all the leaders in my company are sounding the “catch up or you’ll be left behind” alarm. People are going to be let go at an accelerated pace in the future (1-3 years).
> We find no systematic increase in unemployment for highly exposed workers since late 2022
This assumes that the companies' business growth is a function of the amount of code written, but that would not make much sense for a software company.
Many companies (including mine) are building our product with an engineering team 1/4 the size of what would have been required a few years ago. The whole idea is that we can build the machine to scale our business with far fewer workers.
Even in companies that are no longer growing I've always seen the roadmap only ever get larger (at that point you get desperate to try to catch back up, or expand into new markets, while also laying people off to cut costs).
Will we finally out-write the backlog of ideas to try and of feature requests? Or will the market get more fragmented as more smaller competitors can carve out different niches in different markets, each with more-complex offerings than they could've offered 5 years ago?
It is absolutely likely. The hiring market for juniors is fucked atm.
(And if it is, what is the cause?)
How long have we been hearing about crushing affordability problems for property? And how long ago did that start moving into essentials? The COVID-era bullwhip-effect inflation waves triggered a lot of price ratcheting that has slowed but never really reversed. Asset prices are doing great, as people with money continue to need somewhere to put it, and have been very effective at capturing greater and greater shares of productivity increases. But how's the average waiter, cleaning-business sole-proprietor, uber driver, schoolteacher, or pet supply shopowner doing? How's their debt load trending? How's their savings trending?
[1] https://www.npr.org/2026/02/12/nx-s1-5711455/revised-labor-d...
[2] https://www.marketplace.org/story/2025/12/18/expect-more-of-...
Lots going on right now in the market, but IMO that retreat is the biggest one still.
Many companies were basically on a path of infinite hiring between ~2011 and ~2022 until the rapid COVID-era whiplash really drove home "maybe we've been overhiring" and caused the reaction and slowdown that many had been predicting annually since, oh, 2015.
There's a lot of perverse interests and incentives at play.
> We find no systematic increase in unemployment for highly exposed workers since late 2022
Also dont forget theres only so many viable revenue-generating and cost-saving projects to take. And said above - overhiring in COVID.
That's... not a good look for your engineers?
Then any company that was staffed at levels needed prior to the arrival of current-level LLM coding assistants is bloated.
If the company was person-hour starved before, a significant amount of that demand is being satisfied by LLMs now.
It all depends on where the company is in the arc of its technology and business development, and where it was when powerful coding agents became viable.
Being overworked is sometimes better than being underworked. Sometimes the reserve is better. They both have challenges.
In that sort of high-fear, change-adverse environment "get rid of all the devs and let the AI do it" may not be the most compelling sales pitch to leadership. ("Use it to port the code faster so we can spend more time on the migration plan and manual testing" might have better luck.)
Best time to be a solo founder in underserved markets :)
The real impact is for indie-devs or freelancers but that usually doesn't account for much of the GDP.
Don't know if this is effective and I don't think management knows either, but it's what they're doing
Doesn't mean the two are related.
Is AI just the excuse? We've got tariffs, war, uncertainty and other drama non stop.
Management often has a perverse short-term incentive to make labor feel insecure. It’s a quick way to make people feel insecure and work harder ... for a while.
Also, “AI makes us more productive so we can cut our labor costs” sounds so much better to investors than some variation of “layoffs because we fucked up / business is down / etc”
"We've frozen hiring because our growth potential is tapped out."
"We've frozen hiring because AI can replace employees."
Instead they are using Electron and calling it a day. Very ironic isn't it? If AI is so good then why don't we get native software from Anthropic?
For me, the impact is absolutely in hiring juniors. We basically just stopped considering it. There's almost no work a junior can do that now I would look at and think it isn't easier to hand off in some form (possibly different to what the junior would do) to an AI.
It's a bit illusory though. It was always the case that handing off work to a junior person was often more work than doing it yourself. It's an investment in the future to hire someone and get their productivity up to a point of net gain. As much as anything it's a pause while we reassess what the shape of expertise now looks like. I know what juniors did before is now less valuable than it used to be, but I don't know what the value proposition of the future looks like. So until we know, we pause and hold - and the efficiency gains from using AI currently are mostly being invested in that "hold" - they are keeping us viable from a workload perspective long enough to restructure work around AI. Once we do that, I think there will be a reset and hiring of juniors will kick back in.
If AI increases productivity, and juniors are cheaper to hire, but is just as able to hand off tasks to ai as a senior, then it makes more sense to hire more juniors to get them working with an AI as soon as possible. This produces output faster, for which more revenue could be derived.
So the only limiting factor is the possibility of not deriving more revenue - which is not related to the AI issue, but broader, macroeconomic issue(s).
I think this is the crux of it. Someone who doesn't know the right thing to do just isn't in a position to hand off anything. Accelerating their work will just make them do the wrong thing faster.
His rationale is he won’t let the company log his prompts and responses so they can’t build an agentic replacement for him. Corporate rules about shadow it be damned.
Only the paranoid survive I guess
Anthropic can cause layoffs through pure marketing. People were crediting an Anthropic statement in causing a drop in IBM's stock value, which may genuinely lead to layoffs: https://finance.yahoo.com/news/ibm-stock-plunges-ai-threat-1...
We'll probably have to wait for the hype to wear off to get a better idea, but that might take a long while.
Then the 2008 crash happened and those people were gone in a blink of an eye and never replaced. The companies grew in staff after that, but it was in things like sales and marketing.
No one knows what’s going to happen in the future. Yes there already are fewer SWE jobs than before because of AI, and yes the days of companies hiring new grads in droves at $300k+ packages are likely over. IMO all you can really do is study what you’re interested in, learn it deeply, and do good work with cool people. If unsure, it’s possible to go back to what you were doing before if the new path doesn’t work out.
Shipping speed never/is was the issue. Most companies are terrible at figuring out what exactly they should be allocating resources behind.
Speeding up does not solve the problem that most humans who are at the top of the hierarchy are poor thinkers. In fact it compounds it. More noise, nice.
Writing code is lesser problem than figuring out what we want when we want, and to get stakeholders at one place.
Apple has already shown this decades ago - they got the iPhone and iPod developed and out the door in relatively short-time scales given the impact of the products on the world. Once you know what you want, exactly what you want, things moves fast - really fast.
There goes my excuse of not finding a job in this market.
Also, it seems to me the concept of "observed exposure" is analogous to OpenAI's concept of "capability overhang" - https://cdn.openai.com/pdf/openai-ending-the-capability-over...
I think the underlying reason is simply because companies are "shaped wrong" to absorb AI fully. I always harp on how there's a learning curve (and significant self-adaptation) to really use AI well. Companies face the same challenge.
Let's focus on software. By many estimates code-related activities are only 20 - 60%, maybe even as low as 11%, of software engineers' time (e.g. https://medium.com/@vikpoca/developers-spend-only-11-of-thei...) But consider where the rest of the time goes. Largely coordination overhead. Meetings etc. drain a lot of time (and more the more senior you get), and those are mostly getting a bunch of people across the company along the dependency web to align on technical directions and roadmaps.
I call this "Conway Overhead."
This is inevitable because the only way to scale cognitive work was to distribute it across a lot of people with narrow, specialized knowledge and domain ownership. It's effectively the overhead of distributed systems applied to organizations. Hence each team owned a couple of products / services / platforms / projects, with each member working on an even smaller part of it at a time. Coordination happened along the heirarchicy of the org chart because that is most efficient.
Now imagine, a single AI-assisted person competently owns everything a team used to own.
Suddenly the team at the leaf layer is reduced to 1 from about... 5? This instantly gets rid of a lot of overhead like daily standups, regular 1:1s and intra-team blockers. And inter-team coordination is reduced to a couple of devs hashing it out over Slack instead of meetings and tickets and timelines and backlog grooming and blockers.
So not only has the speed of coding increased, the amount of time spent coding has also gone up. The acceleration is super-linear.
But, this headcount reduction ripples up the org tree. This means the middle management layers, and the total headcount, are thinned out by the same factor that the bottom-most layer is!
And this focused only on the engineering aspect. Imagine the same dynamic playing out across departments when all kinds of adjacent roles are rolled up into the same person: product, design, reliability...
These are radical changes to workflows and organizations. However, at this stage we're simply shoe-horning AI into the old, now-obsolete ticket-driven way of doing things.
So of course AI has a "capability overhang" and is going to take time to have broad impact... but when it does, it's not going to be pretty.
The TL;DR is that there is little measurable impact (and I'd personally add "yet").
To quote:
"We find no systematic increase in unemployment for highly exposed workers since late 2022, though we find suggestive evidence that hiring of younger workers has slowed in exposed occupations"
My belief based on personal experience is that in software engineering it wasn't until November/December 2025 that AI had enough impact to measurably accelerate delivery throughout the whole software development lifecycle.
I have doubts that this impact is measurable yet - there is a lag between hiring intention and impact on jobs, and outside Silicon Valley large scale hiring decisions are rarely made in a 3 month timeframe.
The most interesting part is the radar plot showing the lack of usage of AI in many industries where the capability is there!
Gemini 3 and Opus 4.6 were the "woah, they're actually useful now!" moment for me.
I keep saying to colleagues that it's like a rising tide. Initially the AIs were lapping around our ankles, now the level of capability is at waist height.
Many people have commented that 50% of developers think AI-generated code is "Great!" and 50% think its trash. That's a sign that AI code quality is that of the median developer. This will likely improve to 60%-40%, then 70%-30%, etc...
Kinda done with this.
If you have something important to say, say it up front and back it up with literature later.
They do nothing?
I trust the workers more to dictate the direction of a company than most executives.
They can't do worse.
edit: or what another commentator said, fucking academia. Public universities have done more for humanity than nearly anything to come out of SV. Surveillance capitalism, mass misery + psychosis; it's very telling what our society values when mass amounts of the Earth are desperately trying to ban these very same services to protect children.
I hope they know what they're doing.
- AI has allowed radiologists to review a much higher rate of x-rays
- The above has led to a dramatic increase in need for faster processing, more storage of scans etc
- which in turn led to needing a bigger IT department to manage all of the additional workload
There was a similar anecdote about the IRS where the claim is they went from having N accountants to having much fewer accountants but now they need N IT people to manage the new systems.