Thoughts on slowing the fuck down
90 points by equeue
90 points by equeue
I chuckled, I feel the pain, the recommendations are absolutely on point and fundamentally yes, we should ideally probably collectively slow the fuck down.
But without any supporting evidence the premise that software as a whole is visibly getting less robust smells of confirmation bias.
In this climate providing first-hand evidence is bad for your career. I wish I could go into details of how a friend has spent the past week cleaning up a mess of bugs created by a slightly overambitious robot, but for now these things are shared privately.
The little public science we do have is not promising though: https://entropicthoughts.com/no-swe-bench-improvement
It's been known for a long time that nothing beats automation when it comes to producing defects at a high rate -- this is why jidoka/autonomation became a big deal in manufacturing. I don't know how wo can achieve that with reasoning robots.
Yeah there's a genuine spike of hiring contractors (who command top rates) to clean up the messes from AI-assisted code triggering outages. Unfortunately the incentives are such that we don't hear about these stories:
"Keep it in the family."
But without any supporting evidence the premise that software as a whole is visibly getting less robust smells of confirmation bias.
Would be pretty difficult to measure, but just based on service availability alone it 'feels' obvious.
Feels like stories such as this and this are becoming all too frequent
GitHub and AWS reliability issues have been a continuous problem since forever, no? Visibility of similar levels of outages might change for fashion / politics reasons. Which is why we need data to distinguish «bad, and always has been bad» from «the trend got significantly worse very recently»
For context, the author wrote the wildly popular pi coding agent: https://mariozechner.at/posts/2025-11-30-pi-coding-agent/
While all of this is anecdotal, it sure feels like software has become a brittle mess, with 98% uptime becoming the norm instead of the exception,
That's not my experience. My local setup is more robust than ever before. And Github and AWS had problems before, and over last 5 years there have been many voices and signals that the technical debt and organizational problems are catching up with them. The slop revolution is probably putting additional strain on Github now just because there is way more volume they need to handle.
there sure is a feeling that Windows is going down the shitter.
25 years now without WIndows, my Linux systems are working better than ever. Again - for at least 15 years now with Windows there have been plenty of voices and signals that it's on a downhill trajectory. Is it really LLMs fault the incentives that had been there from the beginning are coming to inevitable conclusion?
Lots of software was working like crap already 5 years ago, way before slopocalypse.
But you can't blame LLMs for your own choices. In the presence of LLMs and software slopcceleration your technical wisdom to choose robust technologies and products becomes even more important. So far the tech stack I've been investing works robustly in the presence of LLMs. E.g. Open Source/Linux, CLIs, strong types (Rust), declarative systems (NixOS), hypermedia UI.
I personally I do agree that vibecoding doesn't work and producing tons of poor quality code is counterproductive. That's why I don't do it. But also - I don't care if other people do. I will not be using their software anyway. Just like I didn't use Windows. Or bloated complex IDEs.
But you can't blame LLMs for your own choices. In the presence of LLMs and software slopcceleration your technical wisdom to choose robust technologies and products becomes even more important. So far the tech stack I've been investing works robustly in the presence of LLMs. E.g. Open Source/Linux, CLIs, strong types (Rust), declarative systems (NixOS), hypermedia UI.
Sure. I mean, that works for some people, people who have autonomy or are more tech-ish. Outside of our bubble (as developers, sysadmins, and whatnot), most, if not all, people are still relying on whatever folks sell them that is any good. Out of the blue, it feels like every single tool is adding an AI-mode, AI "enhancement", and even when we still pick specific tools that are robust and have been robust for years, they start to add/accept AI slop, such as Vim, NeoVim, Helix, and probably many others too.
And, not only that, it strains more old, reluctant developers to keep the barrier high enough so slop doesn't start to spread (such as, for example, Wayland maintainers). Raising the question of what happens when they stop doing that, or leave the project?
In the end, it is a correct assumption that even if you don't care because you don't use it, not doing anything ourselves enables others to keep doing their part to destroy whatever is left
they start to add/accept AI slop, such as Vim, NeoVim, Helix, and probably many others too.
I have more nuanced view on using LLMs than treating everything as evil-slop. While I believe there's going to be a lot of vibecoded slop that falls apart, and there will be trouble even with non-pure-slop software, and people will need some time to fully figure out when too much LLM is too much, learn to use it responsibly, and how to collaborate in this new environment, I think think it will sort itself out.
Raising the question of what happens when they stop doing that, or leave the project?
Fork button is always there.
it feels like every single tool is adding an AI-mode
That part is actually great. The worse it gets, the higher chance that software consumer will awake from their slumber and accepts more active role in taking responsibility for creation and funding of software they use. Or not. In which case more power and competitive advantage to people who do.
I think you get out what you put in, the tools are an accelerant but (as the article points out) this is resulting in many people abdicating responsibility for the code because they don’t write it anymore. Failures are generally cheap and you can shove in the error messages to get quick fixes, but repeat that process enough without some guiding principles and you naturally end up with a mess.
So, slow down from full YOLO mode. Challenge the agents to produce better tests, find the gaps in your specifications, and do adversarial review. I think I’m producing ~90% of the quality I do from full manual when using agents, but around 4x faster. The new pace has been set, and I don’t think there’s any going back.
The point is: let the agent do the boring stuff
This is sniping a small part from the conclusion. I wonder if we underestimate how much "boring stuff" there is? In my case it turned out to be a lot - more than I thought before I dove in.
In that light, these tools feel quite liberating. I know that’s not universal -- but maybe that’s exactly where the split is?
Personally, I am enjoying watching the trillions of dollars being funneled into the AI black hole of silliness. Sure, it's a bit wasteful, but the amount of opportunity this is creating for the next generation of competent, hardworking people is mind-boggling! This will be the biggest crash (assuming we can't soft-land it) that the economy of the world has ever witnessed, but it will also be the greatest set of "green field" opportunities that our industry has ever seen, dwarfing the current bubble, the Internet / web bubble, the Windows 3.1 bubble, the PC bubble, the Persian Abacus bubble of 532 BC, and all the other bubbles we've ever had the misfortune to be deeply invested in when they popped. (In the meantime, I'm enjoying buying up all of the bored-ape NFTs at a steep discount; these will no doubt be worth quite a bit some day.)
It's not a question of "slowing down". As always, it's simply a question of values and priorities. There are still real problems out there to help address and possibly solve, that can actually benefit humans by being addressed and/or solved. The good news is that there isn't a lot of competition in the space of "actually helping humans". All the competition is in "finding ways to violate people's privacy", and "finding ways to shove more dishonest ads in front of people's eyeballs against their wishes", and "turning natural resources into poison and pollution", and "ripping off old people", etc. So the hard work is almost competition-free. And a new generation of young (with some help from not-so-young) people will hopefully rise up to the opportunity. Yes, we're bequeathing them a world in much worse shape than we found it, despite very few of us supporting that result. So the least that we can do is encourage and support them as they begin their steep climb.
I am not so sure that the result will be an incentive (work , pay. ...) or if it will enable a system that just normalizes that mess in a way. The reason I think so is that a large chunk of "enshittification" became normal and that for example people seem to be pretty fine with AWS and CF being down quite a lot, despite their main claim at least previously being to increase uptime. I could go on with things like Fortinet and so on, but there seems to be a trend where effects and incentives seem off from a classical economical point of view. I know this could be argued in various ways but this is a response to the idea of - assuming there be the loss in quality - there will be an increased market.
I think that product quality is not an important factor for businesses in many cases. If anything then perceived quality (marketing, sales, etc.) is a measure. Great example of that is how phones, apple products etc. are usually put in fancy jewel case looking packaging even though it's basically meant to throw it away. It changes nothing about the product and also isn't keeping it more save than alternatives but it creates the idea of quality.
So I think there is a good chance that "the mess" will become even more normalized and the rest will be selling it and things "looking better" in one way or another without making it overall better.
I'm pretty sure that there was a documentary about this progression a few decades back ... https://www.youtube.com/watch?v=sP2tUW0HDHA
our craft is lost, software development is no longer building clean interfaces and well designed systems, Now it's time to slop it all together