Ask HN: ChatGPT’s Impact on Headcount – What’s Your Experience? | Dark Hacker News
Ask HN: ChatGPT’s Impact on Headcount – What’s Your Experience?
20 points by simonmesmith2 years ago | 38 comments
I’ve found it interesting that ChatGPT sparked fresh concern about AI’s impact on jobs and yet many months later we have historically low unemployment in the US.
Obviously, such impact wouldn’t be instantaneous, but I’ve wondered how companies are thinking about this in terms of headcount planning.
From conversations with friends it seems that a common approach is attrition with non-replacement. Companies are using ChatGPT and similar tools to increase productivity, but don’t want to fire a lot of people due to the negative impact on morale on those that remain—and the fact, I assume, it’s not yet clear how much of a productivity gain will occur where, so they don’t even know exactly who to fire yet.
I was wondering what others on HN are seeing. How are your companies approaching this? Is attrition with non-replacement a common strategy? If so, what might that mean for the future job market and unemployment rate?
Worth noting that copilot has been in active use by many devs I know before ChatGPT and that doesn't appear to have eliminated jobs. Fundamentally your job is to understand - not just type characters into a terminal. I think you might be shocked how bad things would get for a person who lacks any understanding of what they're doing if they just concatenate ChatGPT output indefinitely.
I'm not sure how other people use Copilot, but the only way I use it is "Really good autocomplete". Like if I'm typing out 3 repetitive lines, the fact that it figures out the next one for me is amazing. I don't have it writing entire unit test suites for me or designing approaches like with ChatGPT - so I wouldn't even call them in the same league in terms of comparison.
Agreed, but I’m not just talking about engineers. From conversations I’ve had, I think it’s more a question of what will be possible with generative AI over the next year, and whether that uncertainty makes it prudent to hold headcount steady or let it decline. The uncertainty includes the possible impact of the tools on automation, and also how much the tools might solve problems that entire companies now exist to solve, negatively affecting sales.
For Copilot specifically, if it enables developers to be even 25% more productive, and you’re planning headcount, do you reduce your targets by 25%, or raise your ambitions?
I started a company with a combination of no-code tools and LLMs to help me program some of the rest.
I’ve got a degree in CS, but haven’t programmed professionally in 15 years.
These new tools have allowed me to create something even larger teams would struggle with on my own in a matter of months.
I’m astonished at the progress I’m able to make without much assistance.
There is often something I’ll need to sit and debug or figure out the name for if I want to ask Google or the LLM helping me code, but that’s about it.
As far as I’m concerned, that means I haven’t needed to hire at least 2-3 full time engineers, be fully bootstrapped, and profitable from day 1 at launch.
Maybe I’ll hire people in the future, but so far the LLMs have replaced engineers, copywriters, designers, and customer support roles. I can handle it all on my own or have the LLM (in the case of customer support) do a decent enough job that I don’t even have to do it anymore.
But it didn’t really “replace” anyone in this case - you started a new company that might never have existed, and it might only be economically viable because you didn’t need to hire those people. Voila! We have a net increase in employment thanks to ChatGPT.
I think that's a great point. Though now we need to determine whether the result is a company that wouldn't otherwise exist, or a company that would exist but with lower profit margins. Also, would be good to know if this new company can undercut competition, which would likely mean that competitors had to find similar efficiencies.
It was certainly economically viable without LLM help, but the fact that I didn’t have to give up equity, get a loan, or use substantial amounts of my own funds to get it done at what would normally be a longer timeline to boot seems like a loss for some bankers and some freelancers or cofounders I would have needed otherwise.
I do agree that I’m creating something from whole cloth here so it’s not technically replacement as I haven’t let anyone go, but I’m going with the spirit of the question here.
I find it difficult to combine no-code tools with chatgpt. No-code tools are often no-code and point and click. Then chatgpt has to generate point and click instructions. What is your workflow?
I don’t use LLMs for this currently. Maybe one day. I’d certainly be happy if that was the case.
I’ve used tools like https://www.replo.app/ to replicate Ecommerce stores and use LLMs to rewrite copy though. That’s automating a huge amount of work.
I’m using LLMs to help me code backend coding tasks that I need done for backend services and API generation and interfacing.
So in this case, it's negation of 2-3 hires, and likely negation of using freelancers as well. I'm curious to know which LLMs you're relying on for which tasks.
Varies at the moment as I’m onboarding users slowly, testing, and iterating. About $15-30k MRR usually. Only been running a few months so it’s hard to extrapolate too far.
It has been masked by layoffs. It's really hard to actually tell. Companies are still in lean hiring mode. Is that because of lack of work to do or productivity rising to the point that no extra people are needed?
For sure. I’ve also read about companies now, like consultancies, retaining people despite not having enough work, with the expectation that work will pick up in future. So there are many complicating factors.
ChatGPT works really well if you are doing something that is mainstream and well understood.
For example - I've been writing a presentation on image processing and needed a whole bunch of examples. OpenCV has a ton of blog posts, documentation, stack overflow content for it to have learnt from. So it's been great. I can just ask it for some python code to demonstrate some particular algorithm and it can give me really helpful code and suggestions.
I'm also working on a Rails app at work, again, very mainstream, lots of source material for it to have learnt and generalised from. It's a fantastic pair programmer to talk to about how to approach things. And to be honest, I could not use things like ActiveAdmin without it.
Also, I can't stress enough, if you are using 3.5 and not 4 then you really can't comment on how good ChatGPT is.
Senior devs profit heavily from ChatGPT since they can tell bullshit apart from correct answers and know how and what to ask. Otherwise i don’t see much of an impact on the headcount. It helps being already an expert in the subject matter.
Exactly this - I get immense benefit from ChatGPT because I know what I'm trying to do and I know how it should be implemented. I can tell when it's on the right track and nudge it to the right solution.
And because I've got a lot of experience I can "smell" when things are not right.
If I was a junior or fresh developer, I'm not sure I'd be able to do this. But also, as a junior developer, would I know if a senior developer actually knew what s/he was doing?
At least at my company (~100 engineers), we're trialing copilot and have steered away from using GPT at all over legal and compliance concerns. Haven't really spoken to anyone who's made it part of their core workflow. Most of what I've heard outside the company is that the stuff is very useful for senior engineers but if you don't know what you're doing it can lead to some pretty embarrassing issues.
I am excited about AI / machine learning, and I do think ChatGPT is a really cool engineering feat, but in it's current state I don't see it displacing too many jobs.
Basically right now it's a fairly accurate assistant that you can communicate with via text only. I just don't know how many jobs can be replaced with that.
I find it can be a faster way to retrieve information than using Google for inconsequential tasks like, generating a template of a job description which I plan on editing quite heavily anyway.
I would say it makes me roughly 4x as productive as a software engineer. It's all about knowing how to use it. I do agree with the sibling comment that it helps a lot if you are already senior since you can immediately tell what is wrong/garbage that it is outputting.
I wonder what kind of stuff you're working on, that this makes you 4x more productive. For my work (stream processing, analytics, in Clojure), it's useless, AFAICT.
It's only really worthwhile on mainstream coding patterns and paradigms. If you're in a niche then generally the results tend to be very hit and miss.
This is not that surprising. There's just not enough source material for it to have learnt from and generalise.
I have a friend who does embedding programming in C++ - and he finds it occasionally useful but mostly it fails. He has the double whammy of being in a small niche and a lot of embedded people still writing in C (and doing it in very odd ways).
I would mostly agree for this, if you need it for some general patterns it's really good for that. Things like "how do I implement Singleton properly in Python contextually in this codebase". I can't use it as much on things post 2021, though, for instance.
I think in general it's just really good at Python, so it's a boon if you code in Python. I was using it for Kotlin the other day and while it helps, it's not as immediately good in my experience.