Compared to my salary, the current cost of the models and tokens to do the work I normally would, is around 10%-25% of it.
Obviously, you still need someone to prevent the models from going insane and messing everything up, but in my experience (webdev projects, DevOps stuff, local software, well known domains), it is very much a force multiplier, as long as you acknowledge that you really need tests and various prebuild scripts.
So I predict one of two things happening:
A) de-valuation of software development work in well-explored domains (and perhaps some changes in regards to outsourcing, as long as cultural and communication differences can be compensated for); with the implications for those learning programming now
B) the squeeze coming in the other direction, making inference 3-5x more expensive, though maybe not with how every big org out there is trying to be a loss leader
Either way, it's an interesting direction - instead of ever becoming "proper" engineering (outside of RFCs and foundational stuff), we went from React/Vue/Angular/Svelte/Express.js/Laravel/Django/Rails/ASP.NET/Spring wild west and frameworks of the day (never being able to nail down what "good practices" are and stick to them for decades, but chasing the new thing forevermore), to even closer to producing non-deterministic slop, except that the slop kinda sorta works. Wild times.
But it is true, the cost is effectively zero. There will be, for a long time, free models available and any one of them will give you code back, always!
They never refuse. Worst case scenario the good models ask for clarification.
The cost for producing code is zero and code producers are in a really bad spot.
I beg to differ. Let's say you're right. Code producers should turn to agriculture and let their managers and product owners prompt AI to produce code. How about code maintainers? Ever heard the mantra "You build it, you run it"? Lets say that AI can build it. Can it run it though? All alone, safely, securely and reliably? No. It can't. We can keep dreaming though, and when will AI code production services turn profitable? Is there a single one which turned profitable?
Agents are monitoring metrics and logs.
A bug is introduced into the system.
Error rates go up and the agents notify diagnostic agents.
These agents look at recent commits and figure out the problem.
They instruct another agent about how to fix the issue and deploy the change.
The problem is fixed before an engineer even has time to start looking at logs.
If you aren’t seeing this, you’re not keeping up with what others are already doing. It’s not just people vibe coding ToDo apps.
Calm down buddy, maybe you're confusing code producers with something else. It's 2026 we don't bother with maintenance no more, we /new to keep context clean and start over.
Just don't forget to comment - never delete - old code. Always keep dead code around to please shareholders, line numbers up always.
We produce code, that is the main thing, never forget.
One could argue we could achieve the same goals by appending \n to a file in a loop, but this is inefficient nowadays with generous token offerings (but could change in the future than I highly suggest just outputting \n to a file an call it productivity increase)
I didn't understand your point about product owners. Who the fuck would ever need one when code produces itself?
Zero as long as your time is worth nothing, and bad code and security issues cost you nothing maybe.
"Getting code" has always been dead simple and cheap. Getting actually good code that works and doesn't turn into a problem for you down the road is the expensive part
i can't remember who said it but a long time ago i remember reading "Linux is free if your time is worthless". Now we all use Linux one way or the other.
That's still very much true, but at least in the case of Linux the cost is getting lower and lower all the time. The time investment for many has reached about the same as the cost needed for Windows and as a result we see more and more people using linux. At this point it's a perfectly viable gaming platform!
Maybe one day LLMs will eventually make good code at a low cost, and that will allow non-programmers to write programs with few problems but the cost will never be zero, and I think we're a long long way from making human programers obsolete.
All of the intelligence that LLMs mimic came directly from the work of human minds which got fed into them, but what LLMs output is a lossy conversion filled with error and hallucination.
My guess is that the LLMs producing code will improve for a
short time, but as they start to slurp up more and more of their own slop they'll start performing worse.
I love this approach. Great work. Building helpful, accurate has been the second hardest part of building my employer’s internal app. (The most difficult thing has been reaching consensus on processes.)
Commenter:
> What % of the code is written by you and what % is written by ai
OP:
> Good question!
>
> All the code, architecture, logic, and design in minikv were written by me, 100% by hand. I did use AI tools only for a small part of the documentation—specifically the README, LEARNING.md, and RAM_COMMUNITY.md files—to help structure the content and improve clarity.
>
> But for all the source code (Rust), tests, and implementation, I wrote everything myself, reviewing and designing every part.
>
> Let me know if you want details or want to look at a specific part of the code!
Oof. That is pretty damning.
———
It’s unfortunate that em-dashes have become a shibboleth for AI-generated text. I love em-dashes, and iPhones automatically turn a double dash ( -- ) into an em dash.
The author felt pressure to build an explosive startup with tons of early funding. For those feeling similar pressure, look into the concept of “slow burn startups.” These are startups that stick around and make long-term impact.
reply