Coding Jobs and GPT-4 Do you believe that GPT-4 will significantly reduce the demand for coders? If not, why not? I am curious to get your sense of this. I am not a professional coder. I write code for scientific work. But I am really curious. |
Coding Jobs and GPT-4 Do you believe that GPT-4 will significantly reduce the demand for coders? If not, why not? I am curious to get your sense of this. I am not a professional coder. I write code for scientific work. But I am really curious. |
It will also increase the opportunities. - Small companies and individuals will learn to be 10x more efficient. - Many, many newer business and side hustles will pop up in matter of days.
Define "coders".
A slightly cynical view: yes, it will reduce the demand for "writers of code in a single language and framework, following the docs and without much other understanding". No, it will not reduce the demand for proper software engineers, who understand the inside-out of the systems they work with, from hardware to kernel to libs and whatever stack they're using, and who will become more productive and will likely be able to deliver more from their huge backlogs.
Explanation: my impression (totally an opinion, but I've heard some people roughly agreeing with me) is that in the last 10-12 years a whole lot of people entered the software development profession attracted by salary, hype, easy money (crypto?), or whatever else, maybe through a dev bootcamp or just by learning on the internet. That's not a problem per se, some of the most skilled devs I know are self taught. But there're a number of developers that, even after a lot of years, have only a limited understanding of how their systems work.
Python programmers that know Django but have no idea what a syscall is, or how a process works in their operating system of choice, or how the Django ORM translates those classes into some SQL. They live in the IDE, somebody else is taking care of the deployment after they push, they just make changes and addition to an existing app. That's it.
Frontend programmers that know some toolkit like React or Angular with a few libraries, that can write some piece of Typescript, but that often have no idea about some Javascript basics, the browser, http. Again, they contribute to an existing project, adding files, changing things, but somebody else is in charge of deployment, and they won't be able to setup a new project from scratch, if needed.
A lot of those people still had/have good salaries, maybe not the best, but the market was tilted in their favour; if top tier talent flew towards big tech, there were positions to be filled in other sectors. But I think that even big tech wasn't immune to the coding bootcamp effect.
It seems to me that when money was cheap and the work to be done was a lot, it made total sense to add whatever kind of help could be found on the market. Now, this could change; if software engineers can move quicker with the help of AIs, follow-the-docs-only programmers may be become far less useful.
Just my 2c.
We expect the economy to get way worse as de-dollarization seems to be looming on the horizon, and preparing now.
Coding (physically typing a solution) is a small percentage of a professional software engineer's time. Tools already exist to generate code, to find answers (stack overflow), etc. GPT will be another tool to help developers but much like stack overflow answers, you have to do some diverse research to make sure you're not getting BS or tangential but useless information. GPT can be confidently and completely wrong.
Then there are the legality concerns of sending Intellectual Property to a giant void that may output that IP to other parties. Its not a concern of mine, personally, but its a concern for the legal entity (aka corporation) that employs me.
Also, re de-dollarization. This pops up every 5 years or so when some autocratic regime gets tired of operating under the us dollar hegemony, but I don't see any other options that don't have their own major issues. The euro? That currency is fraying at the seams from trying to stuff high and low performing countries (i.e. Germany and Italy) together without enough control over each country's finances. China? They've got to free float and stop manipulating the yuan first.
Literally nobody is as trustworthy as the USD, no matter how badly Moscow, Beijing or Tehran might wish it to not be so.
What pipelines were you recruiting from?
I see AI decimating "become an engineer in 10 weeks bootcamp" types of coders, but proper engineers? The demand is still exceptionally strong, especially now that Twitter and Meta released a lot of talent on the market at the same time.
> We expect the economy to get way worse as de-dollarization seems to be looming on the horizon, and preparing now.
Economists and computer scientists confirmed it: The de-dollarization will happen the same year as the year of the Linux desktop!
American social media has been surprisingly loud in their predictions about the total imminent collapse of the US dollar and economy since February 2022.
As a European I would like cheaper American products and services caused by a weakened US dollar. But I don't see it happening yet. The US dollar is still strong, American companies are still competitive, and the S&P500 is a lucrative ingredient in most European pension funds.
1. For now, not by much. It will make us more productive but the world is not short of projects with endless backlogs so huge nobody can ever drain them. Maybe some of those tickets are low value or obsolete, but a lot of them represent real bugs or ideas for improvement that the cost of software development makes prohibitive to reach.
2. It may change the nature of the job a lot. It could hurt developers who are very heads down and ideas constrained. If what you love most about coding is the thrill of finding the perfect algorithmic implementation, then, well, AI may reduce the enjoyment of the job. If what you love about coding is seeing your ideas come alive, then AI can increase your enjoyment of the job.
3. Whilst some people claim they ideate and get inspired by talking to AI, my own experience has been that GPT4 is a rather conservative and predictable sort of personality. Perhaps it's the training and maybe via the API you can ramp up the temperature and get more creativity out of it, but whenever I've asked it for ideas or tried to bounce ideas around, I tend to get milquetoast yes-man type results, or the ideas it comes up with are generic and obvious. Also with each day that passes it's getting easier to notice the lack of "AI invents/solves something new" type stories. There was the one where it came up with a word game, which was pretty cool, and we know it can invent stories. But it doesn't yet seem to be producing an abundance of new ideas for things like new features, business innovations, etc. Maybe it does for other people. Maybe it will start doing it for me soon. But for now, it doesn't seem able to do that.
So - programmers who just want to play code golf, be ready for that to become more of a hobby than a job. But for programmers who always wished they had more hours in the day to get through all their ideas, it will enable them to create a lot more value and that value will in turn create demand for yet more value-add on top of that. So it should be a virtuous circle, in theory.
But to me, this also implies a vicious circle of yet more crappy, barely-good-enough, bloated, inefficient software. I wish this wasn't the case. I recognize the upside; more real-world problems will be solved. But we really need to do something about the excesses that lead to things like the old (Electron-based) Microsoft Teams.
Then ask GPT-4 to convert it to SwiftUI. It actually does it. Or try Jetpack Compose: same thing.
This leads to an obvious thought - part of why people use heavy abstractions is because the cost of native development is too high. You could pay development teams to write a WPF UI, a Cocoa UI, a GTK UI, an Android UI and a UIKit UI for your app but even VC funded firms flush with cash can't justify that level of extravagance. It's too hard and expensive to hire devs, so when you do, the opportunity cost of them porting stuff between endless UI native APIs just can't make sense. AI changes that equation completely. GPT-4 is cheap for what it does. Suddenly, auto-porting your app into 5 different native APIs simultaneously and then assigning the resulting bugs to a bug-fixing bot doesn't seem unimaginable.
Now, for better or worse, even if AI reaches the point where it can do that, it doesn't mean Teams won't be slow. People use Electron for a lot of reasons, not just cost of duplicated development (skills, libraries, Windows lacking a competitive and non-deprecated UI toolkit etc). But it's interesting to think about ideas that we instinctively write off today due to the massive cost of software developer time.
Folks building software can be more ambitious with the things they create, and it can become easier to plug in another software engineer to your team if tooling can make them more predictable in their output.
I see people here saying that increased efficiency is going to reduce their team's headcount which makes zero sense to me. I've never been on a team where there wasn't enough work to be done, it's usually a problem of too much work and needing to choose which is most important.
That being said, as someone who's been heavily using GPT-4 in their work and personal projects I feel like the gains are minimal. It's really great when I'm exploring an API I was unfamiliar with before, but for stuff I've been doing daily for years it's really not much help at all. Even places where ChatGPT shines, it takes some discernment to sift through the bullshit.
The value I have found from it has been quite high, I have been more confident about being able to offer ideas, knowing I wouldn't have to spend quite as much time refreshing my language skills for languages i have not used in a while.
I agree with your last point,when I first started playing with chat GPT I thought I'd throw some of my already fixed problems and see how it handled them. It didn't have the best of answers (tripping over problems I also had initially) or wasn't much better than what I had already made.
For big, complex, programs I work on (multiple C programs up to 100k LoC) it's hard to frame a question in such a way that it doesn't need to know details of the codebase. For peripheral stuff it's pretty useful as a quick way to get some cut'n'paste code ('write me code to read the load average on linux and log it to a file every second' - I know exactly how to do that, but it's a faster typer).
So initial take is it's a very useful new tool - so sort-of a corollary, people who learn how to use it quickly and well are going to have an advantage in the short term. And even if/when everyone uses it, some people are going to be better at using it (i.e. asking the right question - like I find myself groping around for, when asking it about UK company law earlier today).
For public APIs that might be the killer feature; a better way to look at doc (or rather, rely less on doc) if it can look at different code bases and find snippets of code using a popular API.
Also the quarterly planning process is "figure out the ten things you want to get done out of the hundred in your backlog", and the quarterly execution process is "land the one thing you can get done of the ten commitments you made." So there's no lack additional work to be done even if AI takes over some chunk of it.
This is my use case for GPT-4. It usually does it better than most humans, including those who wrote the code and those who have maintained the codebase for years.
Before Stackoverflow, software research was rather difficult. You had books that acted like tomes of reference materials. You had maybe some tribal knowledge internally on how to do something. But even in the early 2000s something like MSDN was something you might search on a CD ROM! So the time to get a "basic Windows app" up and running was probably twice as long as it would be today. I'd say there's a similar thing with Github, languages with pkg management, with libraries that do every little thing, whereas when I got started in software even the C std lib might be sketchy and even then its very limited.
When Stackoverflow came out, there was a kind of similar panic that we forget about. Every problem would be solved on stackoverflow, and you just copy-paste everything from stackoverflow. It'll get way too easy to launch an app. Yet the reality is there's even more developers, with higher salaries than everywhere, with software in even more niches in our lives.
The current "panic" about software engineering going away is also happening at the same time as layoffs, so its hard to know how that colors things.
The programming profession has always demanded that developers operate at higher and higher levels of abstraction and produce more/bigger/faster. I expect LLMs to continue that trend - a tool to get more done at a bigger scale.
But we don't use punched cards anymore. We don't write low level machine code anymore, unless we must or we are into that. We don't have to use complex and difficult programming languages when we can do the same in a few lines of Python that abstract everything. And now we don't have to write most of the high level code as we can have a loose conversation with the machine.
I'm expecting that the demand for good coders will stay high, to use these new AIs, fix the bugs, check the outputs, and build much more ambitious projects.
However the demand for cheap coders that write simple applications will disappear and be replaced by machines. The humans will do other activities instead, in software development or not.
I CLI
II GUI
III Human language
Ergo, I suspect that not only will GPT et. al. drastically reduce demand for coders, it will reduce demand for e.g. computer languages, frameworks, stacks, etc. What I mean is that questions like "which programming language to use?" will become obsolete.
As for why, my best theory is that writing code isn’t the hard part, but in fact one of the easier parts of the job, much like how drafting isn’t really the hard part when it comes to engineering or architecture.
If so, I imagine engineers will be more productive with AI (and wages will go down, or at least stop growing somewhat), but the demand for software engineers will stay strong.
If not, then engineers being more productive would mean fewer engineers can meet the global demand for software engineering work and I would expect to see the demand for software engineers reduced.
And yet developers have always been in demand. We just have a lot more programs now, and a lot of very-similar and/or niche programs. All the random libraries on npm, Rust, Haskell, etc. the various crypto-currencies, static/dynamic webpage builders, 3 separate JavaScript runtimes. People start businesses and actually get funding for these libraries. And it seems like a lot of companies want almost the exact same product, some "business solution" or "cloud solution", but they have specific reasons existing solutions aren't acceptable (performance? security? some feature?), so they pay $200k+ salaries for developers to build them.
But will this always be the case? Even ignoring GPT4, we don't really need these developers who are working on niche and similar projects; we are having a harder time getting hired right now with the economic downturn. And still, it's entirely possible this growth has a limit, and GPT4 will make developers turn ideas into working products faster than people can come up with them.
Then, how much time do we collectively spend on boilerplate? Stackoverflow + the rich collection of open source libraries and frameworks have been doing a great job minimizing boilerplate for typical architectures. E.g., building a Flask app today vs building with Apache mod_perl in 2000 is like a >90% reduction in boilerplate.
So I don't think LLMs are much of a threat to engineers.
I don't do a lot of functional programming myself but would prefer to. I'm guessing that the output might be be one arbitrarily selected from the described domain but could just as well have been composed completely differently. If it's able to output multiple distinct compositions and we can choose one that suits the team's way of thinking/working that would be pretty cool.
What would be awesome is if it could find the shared characteristics across domains and synthesize a solution in the posed domain that was only ever used in other ones.
1) The dichotomy between frontend vs backend dev evaporates entirely. Because GPT-x can churn out code so quickly, all devs will be expected to be fullstack and the job breakdown will be more like 20-30% coding (assisted by these tools) with the rest of the time spent on higher level concerns like architecture, security (and perhaps lots of meetings...) etc.
2) The vast majority of dev jobs simply get folded into business functions. So your product managers, accountants, etc will increasingly be expected to do prompt engineering work as part of the job. This change will probably take a while to happen and needs more work to occur on hooking up these tools to other tools like compilers and whatnot, but I'm pretty sure startups will emerge to tackle this. Companies that aren't developing any real tech as their main product/service simply won't need many devs anymore, although they might keep a few on staff here and there just in case the others need someone to step in from time to time. The only places that will need sizable numbers of actual devs will be the ones building brand new software - think new database sytems, compilers, device drivers, OS kernel work etc. - basically the really cutting edge stuff for which original research + thinking is needed. Since these tools tend to be trained on existing knowledge out there I'm not sure if they can be used to build brand new technologies from scratch... at least for now. But who knows what future iterations of these tools will lead to!
The reason is fairly simple - software isn't done eating the world.
Make it easier and cheaper to write software and two things happen 1) more problems are worth solving with software and 2) all existing software must improve or get eaten by a competitor who uses productivity gains to improve.
But that's on a technical level.
On the cognitive and reflective matter it will not be as interesting or job-overtaking.
Writing code is 1/6 of the job. The rest is communication, debugging and testing. That's not for GPT.
It has made me a better programmer on the technical level but that was the easy part :)
It's totally possible to make LLMs write actually good code that can be maintained using normal software engineering practices.
Of course there will still be a need for a human touch, such as knowing limitations of LLMs and possible workarounds and understanding business requirements. Such productivity boost may even create more jobs as products will become cheaper and faster to create, hence more new markets will be explored. The pie might also get bigger not smaller.
In an ideal world, you even not coding to achieve things that help others. Today it's required but it's not the point of a valuable product.
I feel like this is more programmers realizing they have to become the 10x they were told was a myth to survive. Which is kinda true! BUT what also is true using AI tools will make them that if they learn to use them, now they need to know security inside and out, systems design etc. Higher level concepts way above a code monkey's pay grade, now will be expectations, all the theory side is going to be much more important.