We're Going to Make Out Like Bandits
55 points by stig
55 points by stig
For me this kind of supply and demand optimism is harder to hang onto. In my country there’s a shortage of bus drivers and yet bus driver salaries remain in place. Personally I think a better mental model for the labour market is the one about casinos, “the house always wins” and adages about supply and demand sound to me like those people who think they’ve invented a system to beat the roulette wheel.
I share both the expectation and qualm.
I feel the sphincter of our timeline, too tightly contracted for many with my skills to pop through.
But repeatedly shouldering the consequences of seduced managers and executives shows me they don't see or feel the same.
I know they'll live in my reality, but I'm not sure they'll notice.
I wish I could share the optimism that there will be an economy capable of paying for software development five years from now instead of just a smoking crater.
This and even if you did make a lot of money off of it you've already paid for it by sitting in a chair pulling your hair out for the next 5 years and for what? Maybe a slightly better quality of life than everyone else, but not by a large margin.
I'm actively planning my life as if I won't have a job 5 years from now and am raking in all that I can before this AI bubble consumes us all.
I said this a week or two ago in chat. Within 5 years I will have an insane gravy train just being paid to fix all the trash output...then retire on a yacht somewhere. Its going to be far more severe than any WYSIWYG or offshore team cleanup work from the past.
Is that cleanup fun? Assuming I had the skills to do it, I'm not sure if have the stomach for it ... Despite the truckloads of cash LOL
I can’t speak to the slop cleanup yet, but cleaning up after poor outsourcing decisions in the past was both not fun at all and paid handsomely.
I was just going to say… LLM-generated “slop” is significantly higher quality than what I remember receiving from the off-shore outsourcing teams a decade ago. I very rarely look at 2026-era LLM-generated code and think “I am completely incapable of assessing what exactly this code is actually trying to do”
And while I agree with you that that kind of cleaning was not “type I” fun, I did find a kind of perverse “type II” fun in it and got a fair bit of satisfaction from reshaping such a mess into something nice. And, you know, when it got depressing, I just dried my tears with $100 bills and got back to it (that got harder when Canada moved to plastic bills, they don’t soak up tears very well)
You're an optimist. The cleanup will be done by AI. After which another AI will clean that up.
The article proceeds on the assumption that people are about non-buggy products.
Here is my alternative prediction: software is just going to get worse and worse, with more bugs. Consumers prefer low price over quality, which is a pattern that can be seen in many product categories. Plus, we've basically trained people to accept that "software is buggy".
There will be no incentive to compete on quality so all companies can just ride the water slide to the bottom, where software is a thing that a handful of engineers in a dark basement with the help of an LLM.
What, me, cynical? Why do you ask?
I think putting everything on the web will eventually put a limit to this.
Shodan and Metasploit can coexist with a technological civilisation systemically not caring about robustness only because people who see the scariness of the combo are not yet sincerely going really all-out burn-it-down. Yet.
I've made the same argument quite frequently this year. Not just cleaning up the garbage code, but also de-LLM'ing all the things which should be deterministic and instead you are paying per token for stochastic errors.
"You're paying $20,000 a month to use a chatbot to figure out which of the four options the user chose? For the low-low price of $20,000 I can give you a program that will do that same function with no long tail cost that changes based on token prices. Oh, and yes, I DO have a maintenance plan of only $1,000 a month, if you want."
I also think 5 years is too long a time window. Companies are already realizing that it costs more to LLM-ify many things than to pay the human and many have already let their humans go. I think the harvest is soon to be here.
That one I hadn’t even really thought much about, but now that you mention it I have definitely overheard conversations between people trying to do “prompt engineering” on problems that have known deterministic/optimal algorithms.
Worse than that, I was on projects that were ripping out already built and in-use software that implemented deterministic decision trees just to rewrite them as prompts.
When I asked why anyone thought we were going to get better results than what we already had everyone on the calls except the salespeople got very quiet.
Why not let them burn and start something new? I’d rather not try to save businesses from their lack of vision, if that’s even possible.
I do wonder if this introduces a new class of company alongside the incumbents, or do rapid feature shipping slop shops eventually grind to a halt on reliability, operations, vision, etc. that they fall back in to that class, leaving room for well-executing companies to take their slice of the pie.
I don't like how these posts always compare LLMs to junior developers. There is nothing special about "senior" developers. In my opinion, LLMs are definitely better that almost all "senior" developers at coding.
This feels like it's an assumption that the problems grow faster than the solutions. If tooling creates bad code, then there's a strong incentive for tooling that mitigates that. In the same way that it's easy to see a future where there's lots of people that create slop, it's easy to see a future where tools to fix slop are ubiquitous.
The tools we have today are in their infancy, but they will get better over time and should outpace the slop factories in capability (mainly because they're part of making slop factories less sloppy).
I wouldn't bet on this being the COBOL moment for millenials / gen-z programmers.
If being a person with actual skills at the right point in time allows me to cash out on that like this until my retirement, there's no end to the amount of laughs I'm going to have.
Hyrum's Law is very real, and nuanced interactions in logic and constraints might only be observable under extreme circumstances.
The usual response is, "Don't worry, the AI will get better." However, cleanup work requires experience and risk management. While AI will nickel-and-dime you for every token, software engineers are typically salaried and accountable for the outcome.
Now, we also know that the tendency of the AIs isn’t to go and clean up code, reduce duplication, or focus on maintainability. Context windows (even the large ones!) can’t hold entire modern repos. They miss things. Most AI written code is additive, and is frequently duplicative. This is already happening.
At least in this specific area, I suspect AI agents have a lot of capacity to improve in the future. You don't need a gigantic context window if you have really good semantic search. I can imagine training an embedding model specifically so coding agents can be really good at finding duplicated concepts: "There are five other resources that also track expiration times with priority queues, so let's load only those files into memory and write a single ExpirationQueue class."
(Of course, the flip side is that this hypothetical AI might be too hasty to apply premature abstractions!)