What I don't think the HN bubble is fully familiar with is the fact that probably 70% of programming happening today is boring low paid drudgery (or maybe 50-90%). It's creating dashboards in BI tools, it's tweaking Wordpress themes, it's writing custom business processes in ABAP for SAP deployments, or it's 3-10 engineers doing the work of 1 so that a outsourcing company can bulk up its bills.
This is the sort of work that AI is coming for. It's low impact, low effort, low paid, and importantly it doesn't get automated (often). This means there is a lot of low hanging fruit for automation, whether that's by regular tooling or LLMs.
On the other hand, if you're a highly paid software engineer, it's likely you're already working with highly automated and abstracted systems, so there's much less opportunity for LLMs. If your deployment process is a Word doc specifying the 120 steps you have to run in bash to deploy that takes 1 eng day, that's going to be automated, but if you're git-pushing and having Kubernetes roll it out in your CI/CD flow, automation has already raised the level of abstraction that you need to work at and there are fewer opportunities.
Yes there's Copilot/etc which will reduce boilerplate, but the only way a company achieves "70% less coders" is by wholesale automation of large parts of their process that must be horribly basic already.
> boring low paid drudgery
> the sort of work that AI is coming for
This is NOT what AI is coming for. Its too far from even barely capable for that sort of work. BI tools, Wordpress themes, layers and layers of internal tools for custom business processes are all incredibly fragile and intricately balanced. It will have to learn systems where extra empty tags in the xml response get accounts flagged for removal, systems where the output must be a zip file archived within a .cab file for the downstream systems to be able to process it. Its an insanely stupid house of cards held together by bottom feeding groups of human intelligence. There's no budget to fix it. No incentive to improve it. It will never change, till its replaced with garbage from an even lower bottom feeding group.
AI is hopeless, because of the sheer lack of anything to accurately train itself for this task. It is trying to sell itself to the OEMs and platforms with shiny demos as an even cheaper alternative to the bottom feeding group of outsourced labor.
The joke is going to be real when the $$$ are spent on AI and the outcome is outsourced to the bottom feeding group to maintain.
They literally said in the earnings call that they now need fewer third party developers because of AI. And this is one of the biggest legacy companies on the planet.
Strange that the people in the trenches are saying that AI is helping them, but people on HN are screaming that AI is useless…
You're unironically believing an MBA who's bonus likely depends on optimizing processes (which is a fancy way of saying "firing whoever they can get away with") over people on this forum, half of which likely code themselves?
AI has made big splashes and some industries are in the process of getting annihilated (ie. Designers working for marketing agencies and similar roles). I'd expect things to get more spicy over the coming decade for programming, too... But right now, at this very moment: LLMs can't code good enough to save time unless the programmer is a total novice.
Might be different in SV or at a FAANG, but speaking from my nearly 4 decades of experience in tech, many to most programmers that have come up in the last 20 years are either lifetime novices, or at least just barely capable of productivity. I see 80% of the work being done by the top 20%. I think AI is definitely going to devastate the vocation because it’s going to replace that 80% barely productive group. I can see a future where most of the code is written to a 90% completion level by AI, then finished by the best of the best of us ugly bags of mostly water.
> Strange that the people in the trenches are saying that AI is helping them, but people on HN are screaming that AI is useless…
I mean, it sounds a lot like a standard management failure mode, which must be about half a century old at this stage: pretend you can get by on X% of staff due to [faddish pretext], then a year later when the cracks start showing quietly hire more people again. The nature of the fad changes, but the basic story is generally about the same.
Actually, I’m not sure it’s even that; at this point the claim seems to be that they _can_ get by on 30% of the staff, not that they _are doing so_. There’s a decent chance this is just a case of “oh, shit, earnings call coming up, we have to mention AI because the markets like that”, tbh.
So true. What it did come for was my part time job of taking a png in a folder, create 5 sizes with the size appended to the file names, put these in a subfolder, convert all sizes to avif and webp and generate the picture html block for the image (avif with fallback webp. png sizes are deleted, original png archived). It took one prompt to get that script and I only changed the name of it to dothis.sh it works with one or 258 .png files in the folder.
> What I don't think the HN bubble is fully familiar with is the fact that probably 70% of programming happening today is boring low paid drudgery (or maybe 50-90%).
What I find odd is that AI tools are used to create new code but the bulk of the work is in rewriting and changing existing software,and fixing issues by touching single lines of code spread throughout the project.
Perhaps AI tools will improve, but today you need an awful lot of prompting to get a single decent usable starting point. In some applications you end up spending far more time prompting the AI tool than what you would end up doing by simply writing the thing.
My usual dialogues with ChatGPT now 4o: I need a SQL to do 1, 2, 3, 4, and 5. GPT: sure thing. Me: umm you forgot 4. GPT: oops sorry, here. Me: now you forgot 3. GPT oops sorry, here. Me: you forgot 4 again. Make sure you do both 3 and 4. GPT: oops sorry, here. Me: you forgot 4 again the fuck I'll do it.
>On the other hand, if you're a highly paid software engineer, it's likely you're already working with highly automated and abstracted systems, so there's much less opportunity for LLMs.
I think you're wrong about this. There are people today getting paid lots of money for basic jobs, and people with real difficult jobs that don't get paid much. Pay rate is as much a function of location and your particular company or industry as it is the actual nature of what you are doing.
>Yes there's Copilot/etc which will reduce boilerplate, but the only way a company achieves "70% less coders" is by wholesale automation of large parts of their process that must be horribly basic already.
The basics often only seem basic because we've done them a great deal. When the machines are charged with doing a lot of stuff, will the basics still appear basic? Will you be able to spot the critical yet tricky errors buried inside reams of code for a process that you never have to develop yourself? Furthermore, will you be able to skip all the basics in your career development and work on advanced things? I have my doubts. A lot of "horribly basic" "drudgery" will have to be done manually, because it is too important to leave to the whims of a LLM and people have no other way to develop the necessary skills to sustain the industry.
Fair points, although I think you may be reading my points as more concrete than intended.
When I'm talking about pay I'm thinking location-agnostic, i.e. low to high pay within a comparable region. For every engineer in London earning £100k there's many earning £26k doing more basic work.
And when I say "basic" I don't just mean working with low level systems, I mean working with systems that are un-automated. Working with processes where someone has made the judgement call that it's cheaper to throw large numbers of cheap engineers at the problem rather than fewer highly paid engineers to automate it all away.
That's not to say that highly paid engineers don't need to get into the basics, they definitely do, but that digging into the depths of an automated deployment system is not the same as running repetitive manual tasks.
Here is a question worth answering yet often overlooked: if AI helps the newbs generate useful stuff but which they don't properly understand, and also helps experienced find the issues on AI generated stuff based on their experience, WHERE is the middle ground where newbs can gain the experience to bring them to that senior level? I see no path going from there to there, unless the juniors start doing things by hand again, for which nobody will want to pay them (hey why don't you use AI instead). And seniors don't live forever...
I'm well aware of this and in those sorts of jobs, most of the work isn't programming, it's sitting in meetings and communicating. While AI might help with this simply by shrinking the amount of people needed to do a job, I'm very dubious about a 70% efficiency gain when you look at the entire scope of what these workers do.
The entire concept that if you make coding 70% more efficient that you'll make coders 70% more efficient is simply wrong.
Maybe you need to be more specific about which workers you're talking about.
Given a full time worker vs a contractor, I'm guessing the contractor has far less knowledge of the system as a whole, and ripe for being replaced. Then with freed up funds, maybe they hire more internal developers. It could be more of a reshuffling than anything else.
I wonder how much of this can be actually automated by LLMs. Isn't substantial amount of time be just lazily spend tracking down what ever variable name or mapping to right thing?
Just spend time adding very simple script as cronjob by CI/CD... And everything was copy paste. But it still took time to set up and name things. To get the small inconsequential details right.
LLMs are fairly good at understanding how a bunch of naming in slightly different forms maps around a system. This is basically the one trick that transformers/attention does! I think a smart copy/paster is what LLMs will be good at.
What they won't be good at is when there are deeper semantics, when there are performance requirements, when there's a complex data model, or when you need to come up with the data model and consider how that will work over time.
I seems that the current AI is rather good at stomping together projects made of react saas starters with nano/microservices (workers etc) where everything is one shot and complete replace and some api routes. People who cannot code at all without AI are now making products that work and sell.
How about you just tell AI to build you a new site? Presumably your prompt will be better now, and AI will be a newer one, so chances are you'll get a site without that particular bug. This approach to debugging would be quite a new paradigm, and even a bit more adapted to consumer society for better or for worse.
Maybe you'll get a site without that particular bug, but you're almost guaranteed to get a site where the unspecified semantics are slightly different compared to the previous version, along with new exciting bugs, and at some point you're likely going to get to data corruption issues because some invariant of the system won't be preserved properly. You'll run into memory leaks and race conditions that you won't be able to resolve. At some point your AI will likely generate new exciting vulnerabilities, leak all customer data and get you buried in lawsuits too.
All of these issues are nearly impossible to spot and correct unless you're an expert yourself, but then you wouldn't need AI in the first place.
That said, I fully expect such an industry to take off and we'll all be worse off for it. Lowering the barrier to entry is empowering for those who couldn't develop software before, which is good, but if it enters mainstream usage then we're all going to suffer even worse products and services.
I won't disagree with you so just let me add that also today we get buggy software and enshittification all over the place. All what AI has to do is deliver those cheaper.
70%... By lines written? Or by time taken? There is a huge difference. Sometimes you spend days trawling old code, even older dependencies to discover adding a line may fix your problem. In the meantime you may've read hundreds of pages of docs, seen dozens of diagrams. Hunted people that moved between departments and "worked on this before" etc. Then after you fix it yours in charge of organising testing and deployment to cert/prod including representing the change in risk assessment meetings. Good luck using AI to replace this work.
I'm talking about time spent. I completely agree that all the work you've mentioned is stuff that takes time and that AI won't replace any time soon, but I strongly doubt that it's the bulk of the work.
I'm talking about the programmers that a friend of mine works with who take several quarters to produce a dashboard, or the entry level PHP jobs that make barely above minimum wage and sell custom websites and hosting to real-estate companies in small towns, or the outsourced programmers that other friends interact with, who spent ~10 eng over 2 years building a "PaaS" style system in shell scripts that no one in their company actually uses because surprise surprise no one ever thought to ask whether it was necessary or had the right feature set or even worked (but management paid for it).
These people are spending a significant amount of their time coding, but they're not spending time automating, or improving the process for next time, and so the work they're doing is fairly trivial and is most likely to be impacted by AI.
I can tell you from first-hand experience that a wall of source code which only works 50% of the time is precisely the level of quality that you can expect from cheap outsourcing. I believe what AI can do is change managers' expectations so that smaller teams become more accepted. And by kicking out the worst performers, you increase average productivity, no matter what the AI does or doesn't do.
I’d believe if you did nothing with AI and just fired a big chunk of your bottom tier cheap outsourced programming “talent” your productivity would significantly increase as a result of your actual talent no longer being occupied with managing and reacting to low quality code contribution and contributor hand holding.
Agree. If someone's productivity is negative, bringing it closer to 0 is an improvement. That means keeping them busy with a useless AI is actually helpful.
The difference between cheap outsourcing by humans and LLMs is that in the former case this 50% of working code remains more or less constant, while with LLMs, I would assume, it gets exponentially lower with every iteration. I use LLMs for my daily work and personal matters a lot and I certainly admire the value they bring but the shortcomings cannot be unseen.
My experience is that the more experienced you are the less benefit you will see from coding asistants. If you are at the beginning of your career and writing a simple for loop requires serious brain power, an AI coding assistant will seriously speed you up. The same goes for task complexity. Mainstream tasks will work better than specialized ones simply due to the lack of adequate training material.
> Due to GenAI and in-house built developer productivity tools, we have increased the
output of our software developers by around 70% year-over-year. In parallel, we have been
carrying out a large-scale digital talent transformation, accelerated by our in-house AI +
human-powered code-reviews-as-a-service capability. As a part of this we have been
reducing 3rd party software engineers in our organization by around 60% percent in the
last 6 months.
They have increased in-house developers' productivity, which means they are less reliant on 3rd party developers.
The thing about outsourcing is that you need to break tasks into tiny well-specified units for your contractors to do, not because they're idiots but because they have little context on anything and often need to be able to work from a different time zone without much back-and-forth. Tiny well-specified units are exactly what copilot speeds up, but they come at the price of increased costs elsewhere: more PMs/business analysts, more managers, onshore engineers whose job is solely to support offshore staff, etc.
Most big corps just need a lot of decent coders to pump out a lot of business logic and UI. AI is definitely going to be able to do that with some basic human validation along the way.
The problem with the validation part is that the human has to go through all the code written by AI. Guess what, understanding someone's code takes almost as much effort as writing the code yourself. Therefore you require just about the same amount of human beings verify AI's code.
Not enough time to even understand the requirements of a 40h job.
10h would be
minimum.
Unless you are just checking that the style guide is correct and obvious
missed null checks.
For this reason I feel code
reviews are a bit silly and pair programming is way better (pairing with a mix of sync and async work) but that is an aside.
In addition a code review being short is usually because the coder and reviewer are both very competent. Once AI
enters the chat the reviewer needs to look very closely at every line.
Code reviews are more of a formality, a pointless one. There's little to no value in one code reviewing a piece of code without the background business problem being solved by it.
Sure but a “superficial” review is what most corps currently run on. Certain crucial code gets looked at more throughly. If there are any bugs QA catches it.
It’s going to be the same with AI. It pumps out a lot of code. Human does spot checks and probably uses other AI to help with the code review. A real human does the really hard, critical code directly. A more robust QA team makes sure it all works.
I think you are right in the near future but not today. I don’t think we are quite there yet with AI but the pace of improvement is breathtaking and producing code (and more generally “logic” which could be code but also machine code, system design, LLM prompts and so on) is such a pot of gold the AI companies will go for it.
Yeah thats because then they don‘t really look at it and certainly don‘t really understand it.
In some way reviewing code is more complicated and hard than actually writing it.
I really like the „reverse centaur“ metaphor by Doctorow regarding the automation AI will bring us - humans having to double check the stuff AI wrote for correctness in AI pace.
At a big corp: 90% of my job is talking to various people to figure out what we want to code, who has to approve it, who I have to coordinate changes with, etc. AI can currently _maybe_ do half of the 10% that's left, with good amounts of supervision.
The supervision is also difficult to do unless people get enough experience actually doing the job, so expect it to take even longer to get people to verify the code once it's written by AI. This is the same concern people have about autonomous driving. If you let the machine take over, then you'll be out of touch when you're called on to do the real tricky bits of work.
lol, on an earnings call. Someone is lying and being lied to.
Unless they were paying a whole bunch of programmers to do not much useful at all or just incredibly simple repetitive things, this is obviously not true to anyone who has tried to use LLMs to solve problems.
I’m guessing some mid management sold the idea of firing 2/3 of the development staff and somebody is declaring victory a bit early.
AI is StackOverflow on steroids. That's how I see it from a programmer's perspective.
It's excellent at getting code snippets, suggesting possible ways on how to handle something, etc. I recently had to help out my son with Arduino programming, which I'm totally new at, and thanks to ChatGPT this was super fast and easy.
But to handle an actual codebase, even a single developer codebase, it just doesn't have enough context for that.
I'm sure some day computers will be smarter than us, but some people are still seriously overestimating this.
My impression is that AI has done more for the accessibility of coding than the efficiency of coding.
Sure, it's much more possible for bob from accounting to throw together some trivial python script than it was a few years ago, but it seems the main benefit more skilled programmers get from it seems to be helping with syntax and pumping out boilerplate. Which is quite a bit of programming work but well less than 70% of it.
When a problem is tricky, it seems people turn off the autopilot.
Bob from accounting will not even know how to run a Python script, how to deploy it, install required modules. Programming is in a significant part about knowing how to break a solution to a problem into small, testable, reusable chunks of code and knowing when a one-file script is fine or when a framework (3rd-party or written in house) may be a better solution. Programming is also about knowing the tooling for your language, the operating system, protocols, formats. AI won't help Bob with that.
If this number is anywhere near accurate across companies, this might end up having far ranging consequences that we're not ready for. Aside from the obvious unemployment of those replaced by AI, why would anyond still want become CS/IT graduate? Might as well close down 70% of CS/IT departments.
I believe only people with a serious lack of skill and lack of drive to change that are in danger here.
When AI starts to produce compilable non-trivial code, I'll consider changing my mind, but that has yet to happen.
Visual Studio can auto suggest me stuff while I'm refactoring, that easily saves me an hour a week of work during a busy week. The rest of my work can't be automated by any current AI, because most of my work is thinking, debugging and testing.
LLMs can only go "Oh, i'm so sorry, let me just try brute forcing the problem! Heres another variation of the solution, please spend 5-15 minutes testing it to exhaustion!" - "Ah, my bad, heres another fix that doesnt work: ..."...
etc.
I believe it's the other way around - most people who have computer science degrees have attended university with a rigorous math, engineering and theoretical computer science curriculum. And despite that, they struggle to use any of it in any meaningful capacity.
Months can go by before I get to solve a good DS & A problem in production. Most of my job consits of fighting with the mundane drudgery of fighting with cloud providers or getting crappy frameworks to do what I want.
I think the value of people who have a deep understanding is about to skyrocket.
They are probably right about the coders. Extrapolating from my brushes with big corp IT efforts and confidently generalising it to BP too, it is likely that the standard of what they have today is so underwhelming that an AI will do no worse.
This is because the corporate culture and reward structure incentivise using the cheapest labour to spend as much money as possible delaying delivery for so long that accountability doesn't come home to roost and the system can instead be scheduled to be replaced by a bigger worse new project before it ever gets finished.
Yeap I'm cynical but there are precious few 'wow they are doing it right' 'efficient and effective' stories from big corp IT.
They are probably right about their help desks too. And so on.
I like how all comments here are "surely it's all just bad coders and not a threat for my big brain" and not like "all our source is being used against us-- today it's juniors, tomorrow it's me".
AI can help those of us on the edges of what the "coder" of an environment is. Things like getting the right syntax for a given set of pseudocode instructions.
Background is EE, inherited an airgapped DCS with everything from things running Fortran from the early 1980s to modern controllers in a heavily regulated industry.
Time savings do matter. And for low end little things we could never get an external contract on we are enabling those that take more of the system owner type role to have more tools in their bag for response.
"Second things like call centers, the language models have become so sophisticated now. They can operate in multiple languages, 14, 15 languages easily. In the past, that hasn't been something we can do. So we can redeploy people off that given that the AI can do it."
For the customers this is going to be a horrible experience and frustration of going in a loop with an AI model. Anyone facing a problem not listed in the 10-15 predefined problems and resolutions is going to have an excruciating time trying to get any meaningful resolution.
Some company's call centers are designed to fuck around and frustrate you until you give up trying to get what you want (e.g. compensation). Here's looking at you AirBnB. Hah, imagine making something that evil with AI, you don't even need to destroy your outsourced call-center workers' souls...
7/10 devs being obsoleted by AI must mean that the company employs about 3 actual engineers per 10 people it hires, and the rest just push around copy pasted code
This is not the first time programmers became more efficient. We've had that happen before; usually because of new languages, tools, and other things that help people become more productive. AI is just another productivity enhancing tool.
Some studies suggest that the lines of codes produced per time unit is more or less constant across different languages. What matters is what those lines do. If you are programming in assembly, it's going to take ages of very verbose programming to get anything done. Do the same in Haskell and you might reduce it to a couple of lines of code.
Based on what happened in the past, we might expect programming work to outgrow the efficiency gains. That's what has happened every time there was a step change reduction in cost for producing software. The community of developers keeps expanding and isn't shrinking.
> We need 70% less coders from third parties to code as the AI handles most of the coding, the human only needs to look at the final 30% to validate it, that's a big savings for the company moving forward.
Reading code is sometimes much harder than writing it yourself. If you still have to validate the code, I don't understand how it's a net gain.
This isn't telling me much. They are less reliant on contractors. I'm guessing these contractors have far greater overhead and handle things which require the least knowledge of the systems they are working on. This could create more full time jobs for BP, by freeing up funds to put more internal people behind AI assisted dev.
Contract agencies have strong incentives to pad work and payrolls and use process controls to coordinate production. It's perfect for AI disruption because you can test replacements at process boundaries, and there's plenty of excess.
Honestly I believe that should wholesale replacement of people come about, programmers will be one of the last ones to be replaced, with the first ones being probably translators, then customer service, help desk clerks etc.
Basically almost all intellectually challenging problems have a major component of programming, to the point that most engineering professions are thought of a special case of 'programmer'. Data Scientists', ML engineers', Electrical engineers' valuable outputs are usually some kind of computer code.
This is the sort of work that AI is coming for. It's low impact, low effort, low paid, and importantly it doesn't get automated (often). This means there is a lot of low hanging fruit for automation, whether that's by regular tooling or LLMs.
On the other hand, if you're a highly paid software engineer, it's likely you're already working with highly automated and abstracted systems, so there's much less opportunity for LLMs. If your deployment process is a Word doc specifying the 120 steps you have to run in bash to deploy that takes 1 eng day, that's going to be automated, but if you're git-pushing and having Kubernetes roll it out in your CI/CD flow, automation has already raised the level of abstraction that you need to work at and there are fewer opportunities.
Yes there's Copilot/etc which will reduce boilerplate, but the only way a company achieves "70% less coders" is by wholesale automation of large parts of their process that must be horribly basic already.