We kind of have a different outlook, as a software development company. Certainly affected significantly.
First, reduction of total number of projects across the board. The projects we do take are completed way more effectively by our senior devs with AI assistance. Basically, it is at the point where it is literally easier to explain to AI what you want / how to do something, than it is to a junior. Add very fast execution on top of that, and it is not close. Yeah, a junior learns and AI doesn't, but the models are kind of improving faster than juniors do, even if it is not domain-specific knowledge. And it is hard to estimate where any given junior will plateau at, too. So even the long term investment into training is very questionable now.
A senior reviewing and correcting AI work is much faster than a team lead doing that for a team of less experienced developers.
We're letting people go both because we don't have work for them, and because it isn't effective to keep a lot of them, even if we did. And I'm writing it from a perspective of a company that kept up the hiring and training of juniors for a very long time when the market wasn't doing that anymore.
If I had to make a parallel to BW, basically anyone below A rank for our field became a liability. (Simplifying of course, soft skills and ongoing relationships of trust between engineers and clients & so on still exist)
Mids don't do that well with AI either, because a mid that hit his skill ceiling at that level is good at doing tasks, not planning, and so is AI.
Only thing that might change it is token economy worsening, but generally speaking what is frontier currently is already at that level, and the cheaper or self-hosted options will catch up to the current level eventually.
And then you end up in a situation down the line when there are no more seniors to oversee the AI because they all retire or get poached by other companies that offer better salaries and there's no one to replace them because no juniors/mids are being trained that could become seniors with experience.
Then it will really be cooked.
Edit: There's also this.
If AI displaces human workers faster than the economy can reabsorb them, it risks eroding the very consumer demand firms depend on. We show that knowing this is not enough for firms to stop it. In a competitive task-based model, demand externalities trap rational firms in an automation arms race, displacing workers well beyond what is collectively optimal. The resulting loss harms both workers and firm owners. More competition and "better" AI amplify the excess; wage adjustments and free entry cannot eliminate it. Neither can capital income taxes, worker equity participation, universal basic income, upskilling, or Coasian bargaining. Only a Pigouvian automation tax can. The results suggest that policy should address not only the aftermath of AI labor displacement but also the competitive incentives that drive it.
While that is generally correct, that is not going to be a problem within 5-10 years, the field is oversupplied, and god knows where the technology will be at that point.
That's one thing.
Another is that not every junior/mid is going to become a senior even you pour a lot of effort into training them, and it just got way more expensive compared to the alternative of not doing it at all.
The balance doesn't add up, and I am not surprised to see the industry go the way it's going now.
Many people expect the whole rebound of rehiring people after AI makes a mess, but I genuinely do not see that happening in software development specifically. Not for an average developer, at least, and those are the ones getting axed disproportionally.
On June 10 2026 21:02 Soulforged wrote: While that is generally correct, that is not going to be a problem within 5-10 years, the field is oversupplied, and god knows where the technology will be at that point.
That's one thing.
Another is that not every junior/mid is going to become a senior even you pour a lot of effort into training them, and it just got way more expensive compared to the alternative of not doing it at all.
The balance doesn't add up, and I am not surprised to see the industry go the way it's going now.
Many people expect the whole rebound of rehiring people after AI makes a mess, but I genuinely do not see that happening in software development. Not for an average developer, at least, and those are the ones getting axed disproportionally.
Maybe for some simpler software that's true. I work on bigger and more complex systems and AI isn't much help there, especially that with big bucks come demanding clients and the fallout of AI hallucinating would be a disaster. Can't really take the approach of "it's not perfect but we'll fix it over time" for most things. This leads to a state where even the smallest things AI does needs to be scrutinized carefully.
Not saying it's all bad. Sometimes the AI can point you in the right direction but ultimately all the work has to be done by the devs anyway. In the past 2 months I think I've scrapped 99% of the code that AI produced for me and had to re-write it by hand.
Hmm, might be dependent on the specific domain, models, setup, etc. We're dealing with pretty complex, enterprise/B2B systems at scale. Not rocket science, but complex enough. That said, the web has always had a lot of materials for this, and agents seemed to have picked up best practices decently well. Then there's gated validation and so on reducing some oversight labor, although human in the loop still very necessary.
If I had to sum up my overall experience so far: I had to scrap less AI-produced code than mid/junior dev produced code. Our most profitable streams are 1 top tier dev + AI, not 1 top tier dev + team. The result is shippable quality.
I'll stop at this point since I don't want to derail the thread into an AI efficiency in software development. In the end, it's a subjective experience of one company.
On June 10 2026 19:26 Soulforged wrote: We kind of have a different outlook, as a software development company. Certainly affected significantly.
First, reduction of total number of projects across the board. The projects we do take are completed way more effectively by our senior devs with AI assistance. Basically, it is at the point where it is literally easier to explain to AI what you want / how to do something, than it is to a junior. Add very fast execution on top of that, and it is not close. Yeah, a junior learns and AI doesn't, but the models are kind of improving faster than juniors do, even if it is not domain-specific knowledge. And it is hard to estimate where any given junior will plateau at, too. So even the long term investment into training is very questionable now.
A senior reviewing and correcting AI work is much faster than a team lead doing that for a team of less experienced developers.
We're letting people go both because we don't have work for them, and because it isn't effective to keep a lot of them, even if we did. And I'm writing it from a perspective of a company that kept up the hiring and training of juniors for a very long time when the market wasn't doing that anymore.
If I had to make a parallel to BW, basically anyone below A rank for our field became a liability. (Simplifying of course, soft skills and ongoing relationships of trust between engineers and clients & so on still exist)
Mids don't do that well with AI either, because a mid that hit his skill ceiling at that level is good at doing tasks, not planning, and so is AI.
Only thing that might change it is token economy worsening, but generally speaking what is frontier currently is already at that level, and the cheaper or self-hosted options will catch up to the current level eventually.
It's cooked.
things have been moving like this forever though. Fire up MS Visual Studio, from say 2000, before they moved to the dot net platform. Try making anything in C++ or Visual Basic. Even Intellisense doesn't work most of the time. There is no inheritance in VB6. Then fire up VS 2005. It is a different world in just 5 years. 1 good software engineer can replace 2 mediocre juniors with VS 2005 as his or her #1 tool. Relative to the primitive Visual Studio 6 ... Visual Studio 2005 is a miracle worker. The micro level coding part of the software engineer's job has been getting easier for decades and long before ChatGPT and Claude arrived.
Weaker software engineers have been getting replaced for a very long time. Now, whining by weak employees who are bad at career planning is at an all time high. So the perception becomes "every software engineer is losing his job".
On June 10 2026 19:26 Soulforged wrote: Mids don't do that well with AI either, because a mid that hit his skill ceiling at that level is good at doing tasks, not planning, and so is AI.
i find AI goes off in crazy, silly directions once you get above the ground coding level. AI can sometimes be good at a micro level. At a macro level... its a joke. I love how ChatGPT will say to me "good catch". ChatGPT is like that political, lazy employee with that "good catch" phrase.
Also, more and more of my customers want everything offline. I've got 2 big customers who pulled their data off the cloud. They got burned big time by regulators and former employees getting access to their cloud data. They've gone from "cloud database warehouses"..blah...blah.. blah... to encrypted .dbf files stored locally on a server not connected to any kind of Wifi at all. Its all wires.
To put it in every day layman's terms... I run a database app shop cranking out the equivalent of the HESA Shahed-136 drone. Business is booming and our HESA Shahed-136 drones are getting the job done. and AI does not know how to build and/or maintain a HESA Shahed-136. So we need humans.
The current error rate in the AI models coming out of these data centers makes useing it in any serious independent fashion untenable. Self-driving cars will only work once their error rate is below human error rate + the cost of a human for insurance purposes. As someone who has a degree in applied engineering with an emphasis on robotics, I can tell you it has zero commercial applications outside of desktop level simulation work. Actual commercial robotics relies on working within a margin of error of thousands or ten thousands of an inch for manufacturing. Even when we get some sort of AI software for self driving cars or drones or manufacturing, it will not be reliant on data center LLM's it will be self-contained ladder logic running on smartphone level computations with code that can be stored in text files.
I can't share the numbers exactly for what these AI companies are paying for parts for their data center cooling system buildouts but Its more than double what the oil industry pays for the same standard parts. And the Oil industry has always been a boom or bust customer.
On June 10 2026 22:38 Sermokala wrote: The current error rate in the AI models coming out of these data centers makes useing it in any serious independent fashion untenable. Self-driving cars will only work once their error rate is below human error rate + the cost of a human for insurance purposes. As someone who has a degree in applied engineering with an emphasis on robotics, I can tell you it has zero commercial applications outside of desktop level simulation work. Actual commercial robotics relies on working within a margin of error of thousands or ten thousands of an inch for manufacturing. Even when we get some sort of AI software for self driving cars or drones or manufacturing, it will not be reliant on data center LLM's it will be self-contained ladder logic running on smartphone level computations with code that can be stored in text files.
I can't share the numbers exactly for what these AI companies are paying for parts for their data center cooling system buildouts but Its more than double what the oil industry pays for the same standard parts. And the Oil industry has always been a boom or bust customer.
I think you're confusing and conflating things. Partially the industry and hype train are to blame, because lots of technologies are maturing around the same time: - LLMs - Robotics - Computer vision
What Jimmy, Manit0u and Soulforged are talking about above is about 99% the first. That same is what is currently driving about 99% of the price of GPUs, the frantic data center construction, etc. That is where coding agents live. It's where microsoft shoving copilot into everything lives. It's where AI slop lives too.
Computer vision advancements are most of what is pushing self-driving cars. None of this runs in big datacenters, because that is just far too slow. These are embedded solutions that run inside the car to recognize traffic, street signs, objects, etc. etc. etc. A lot of the fundamental algorithms underpinning LLMs and computer vision are similar, but CNNs work better in vision, and transformers work better in text, so while those are both ANN architectures, they aren't identical. What they share is that they use stochastic gradient descent in backprop through the Adam optimizer. Meaning that algorithmic optimizations in general ANN training/inference are an accelerant for both computer vision and LLM models.
Robotics are where this comes together, sort of. Particularly also with a lot of recent hardware advances. Better servos, better sensors, better materials and better ways of constructing things. Software advances also exist, and putting a faster chip in a better robot definitely allows for far better fine motoric adjustment, leading to crazy awesome videos of kids in China dancing and doing sports with robots. It also gives you drones in Ukraine.
Self-driving cars are basically robotics, but adding in that this must be essentially autonomous. None of the data centers are at all relevant for self-driving, because it all has to be local for inference, you only need the data centers for training (which while far more computationally demanding than inference, has to happen far less often, and isn't where most of the compute is needed). That said, GPUs obviously are (as well as embedded chips): they just need to live inside the car, instead of in a rack in a datacenter.
As you can see, entirely different things, and while they rely partially on the same hardware and partially on the same software, they share very little in terms of business or actual tech stack, regardless of what Elon Musk tries to tell you.
It's a cool idea that now whole companies are emerging with sole purpose of obfuscating publicly-available data so that it's still normal for humans but garbage for LLM training
EPT
