Ask HN: Do you think some devs are in denial regarding coming AI changes?

I echo a lot of what the other people in here are saying.

I'm also pretty convinced that AI won't replace software dev any earlier than it will replace the majority of corporate jobs, and at that point everyone has a problem.

then everyone will have a problem

>Anyone who agrees with any of the 3 sentiments, what is the justification backing up your belief? Do you feel the general cynicism/dismissiveness is justifiable?

>While most junior level dev work is easily automatable, mid-senior level dev work cannot be replaced and in 5 years time will be no sooner able to be replaced

Because companies were not set up with the idea of being able to contextualize and feed all their internal information into an AI. Maybe going forward, this will change... _maybe_ in 10-20 years (rather arbitrary I guess) one or two "managers" could use AI to interface with customers but for the foreseeable future, I just don't see how it could work.

Assume you had an LLM that would never hallucinate or give a "wrong" answer. Someone still has to hook that up and constantly feed it context or its rather useless.

This is something I understand as a criticism and my assumptions for why it won't hold have to do with my assumption (based on product releases and papers I've read) that the following capabilities will be rolled out in the coming years:

1. Fine tuning LLM's with a company's internal information (already on the "coming soon" part of OpenAI ChatGPT enterprise)

2. More agentic models able to use tools to retrieve, process, and record data, then access its own work as a sort of memory (lots of open source tools are working on these problems now)

3. Much larger context windows with lower tradeoff, and "networks" of AI models controlled by higher level agents with rotating/overlapping context windows

We will need a completely new set of models, LLMs are not sufficiently complex enough to deal with what we deal with. You cannot remove hallucinations so will always need humans to double check what the AI outputs. Most of the time, it's faster and cheaper for the human to do it the first time. Make the AI the junior, not the human

We first have to move beyond text token streams, but to do so requires insanely large models. It will probably be 5 years before we can even make enough GPUs to support every company having an OpenAI scale model they can build & run

Large context windows come with their own problems and there are other papers that show the agent pyramids are fools errand (the "liar" checks the "liar" issue)

Generally, I think you are holding onto the positive hype and not looking at the limitation, negative results, and real world, where these programming AI become more of a burden than a boon.

So no, jobs are not going to magically disappear in the next 5 years. There is plenty of economic history and research to more accurately estimate the job changes that will come with a new technology

> You cannot remove hallucinations

What makes you sure that will hold for the next 5 years as tech improves?

> It will probably be 5 years before we can even make enough GPUs to support every company having an OpenAI scale model they can build & run

Not every company needs its own monster model, the point is OpenAI type companies will build the base model and other companies will fork their own versions

>Large context windows come with their own problems and there are other papers that show the agent pyramids are fools errand (the "liar" checks the "liar" issue)

I looked up agent pyramids and "liar" checks the "liar" issue and found nothing. Could you link those papers you're referencing?

I don't imagine jobs are just going to "disappear", what I do imagine is that gradually senior devs will offload more and more of their work to internal AI models. Big companies like Google will have the best, so they'll signal how things will change for the rest. As the models improve, the chance of any task a model is set to has a higher and higher chance of success and more human readable documentation, to the point where senior devs no longer have to check that they did their job correctly, and other AIs will begin to read/use the code and documentation other AIs wrote. This process will inch up the ladder piece by piece until senior devs begin assigning AI tasks that other senior devs would have done, notice that the AI does it better, and the total efficacy of the senior staff erodes. At this point companies will emerge entirely autonomously run, which is the final signal for other companies that human labor is no longer needed the moment something is automatable.

1. I said we will need new models, not that it wasn't possible

2. It's more of a total through out problem, even if OpenAI ran all the models, you still need the hardware capacity. We are already two years behind in supply vs demand vs manufacturing capacity. A lot of people cannot even try to build, this happened to me over the weekend

3. I'm not going to do research for you. Certainly doing simple searches is not enough to understand where things are going. You have to stay tuned in to the progress and have your finger on the pulse. There are plenty of prominent accounts on X to follow and be aware of what I'm pointing out here. You claims about what is to come are likely exaggerated because you are only understanding the surface at this point. Keep learning and researching!

My job is messy. AI isn't about to replace that or figure it out. Humans naturally are inclined to create complexity. Product owners, designers, etc.; the apparatus of business. It all comes down to us.

AI can take the business of shunting flat boring data around in APIs, but everything else isn't quite such low hanging fruit. It can also have a crack at complex data analysis in really constrained problems, say, feature detection. I can't see dealing with the infinitely variable whims of UI designers or approaching new problems to be within its purview however.

It'll probably be more like the chat bots or automated call centers that just annoy people.

If it's on the web it can do it, but if it's new or complex it doesn't seem to work. Mostly, I see it as an interface to reporting and maybe calling some API's.

The conversion of data tests I've tried all failed horribly.

Come to think of it... The main people that are working on it are advertisers. That is probably the future. =P

We could be headed towards a future where 80% of software is written by AI, but the other 20% requires 200 million humans.

AI will likely revolutionize some software development tasks, but it is too hard to predict the long term effects when we don't even know the future demand for software.

I wonder why you believe the remaining 20% will require 200 million humans. What will those humans do that no AI could do or even hope to do?

You completely missed the point.

I'm just illustrating the possiblity that, if there are only tasks that humans can do, there is the possible those segment of the software industry could grow to be bigger that todays software industry.

It's been roughly 75 years since the world had one computer and a handful of programmers. Nobody would have predicted that humans would replace wristwatches with computers.

The AI revolution is just beginning , but the state of technology could be so different that any predictions we make are likely to be worthless.

My point is how do you imagine that software will require 200 million people after it becomes more efficient due to AI? As a field becomes more efficient, unless demand increases, labor decreases. My question is, why would demand for human software labor increase? Not saying it won't happen, but I'd like to know what you're speculating will happen that will lead to that change, even if it's hard to pin down anything specific.

So what do you think the future will look like, there will be a product owner sitting there on their own prompting the "auto-coder" and then having it deploy to prod? Flawlessly, no problems, no issues figuring anything out just writing words and having products spat out the other end?

In my opinion, the more you can do, the more will be done. The business will want more, we're never going to run out of work to do. So it will be you, working with an "AI" or whatever just generating code and trying to deploy it at break neck speeds while the business has an existential crisis about what to do next.

If AI is successful as you're making out, it sounds like everyone will be able to have a business, there will be no value in software creation, so the whole industry will be gone, it will just be about products and product delivery. Any current day developer can use the AI to build the things they want to build. It will just be different.

The current LLM prompt interface is just generation 1. GPT-4 already has visual input that isn't being shared with the public. As these models become more multi-modal, more ways of "sensing out" what their role in the world/company is, the more efficiently product owners can use them in space of a bunch of workers. It won't be an easy, seamless or cheap transition, but once the benefits start weighing up, there will be no going back.

The point isn't whether or not the amount of work will increase, which is something I agree with as a given. The point is whether or not the rate of increase of capabilities of all AI labor will exceed the rate of increase of demand. I believe it will, which means the demand for human labor in that sector will go down. There are fewer farmers today than 50 years ago, both in absolute and relative terms, even though demand for food has risen exponentially, because the rate of efficiency increases due to farming equipment/techniques outpaced increase in demand.

So you can just be a product owner too ?

As long as people don't get wiped out by AGI (who knows), we'll just go do something else. Who cares about this stuff? Personally, I'm a software developer and I really fucking hate the job, I hate being on a computer (destroys your health), I hate zoom calls, I don't really like "tech", I love being outside in the sunshine. If it ends with automation, thank fucking god really. I do my job for money, that is it.

The existence humans have created for ourselves, staring at a box and pecking away at a keyboard is actually fucking idiotic when you think about it, it's the stupidest reality we could've come up with.

I have had nightmares that I will be forced to continue working as a SWE because the number of SWEs who actually understand the full stack is diminishing. These are actual serious nightmares and I actually think there may even be some truth to them.

I know people who are essentially bribed into continuing to work as a "pilots" for huge ships because no one else is qualified to do the job. So it does happen.

Anyway, I think the situation you're describing is kind of moronic, because if companies automate everything with AI and no one has a job, there will be no economy, there will be no businesses, there will be no need to run AIs, there will be really not much left, just poverty. We'll all be fucked basically and then, a new dawn...

I think even if it replace devs, the devs will make up some other bullshit job to do. Most software development is bullshit so it's already this way.