Speaking for myself. Not really. The problem that I had with enjoying programming was that many of my coworkers who didn't enjoy programming have some opinions that were frustrating to need to entertain.
Now all I have to do is make sure they're not peddling an actively bad idea and if not I can just tell the bot to incorporate their dubious ideas.
In the past it was all, I'm not familiar with this construct that solves our problem perfectly, why don't we cludge adapter pattern in here again.
This was my take on it as well. Which in many ways makes a lot of sense and is pragmatic.
The thing which I think is going to be a very sore lesson for everyone at some point is that if anyone uses AI it becomes a burden for everyone. Even in projects that reject all AI usage.
What I can do is create 50 personas that all contribute patches. 40 of them get called out and are distrusted, but 10 make it through. I can give more focus on those 10 to ensure their trust level increases. Then slowly have them all come together to shepard some subsystem where I'm going to be the only real person involved and can sign off on my own work in a that looks like I have oversight from the outside.
In the past this was possible but took a very high cost. You need a bunch of competent software engineers who you can trust not to blab their secret mission. Today anyone can give it a try.
I don't disagree, and AI helps scaling many things: good work, bad work, malware.
However, I remember that a couple of university researchers (PhD candidates maybe?) thought it would be a good idea to introduce security issues in Linux, and managed to do it. Without AI and without a lot of resources.
A consequence of that was that Linux immediately reverted all the patches coming from that university, saying "we don't know if we can trust you guys" (but really I guess as a way to tell universities "if you fuck with us we ban you, now keep your people in line").
And that's what I was thinking about when I said that if too many contributors misbehave, as a maintainer I will just stop accepting contributions from people I don't know. And maybe that will happen. Maybe it will be harder and harder to just be accepted into the club of the people who can send a contribution and have it reviewed at all.
My theory is that the LLMs can only think in terms of data structures of a certain "dimensionality".
So for example I tried to have LLMs write xml parsers from scratch in Q3 2025, but they kept getting confused and midstream would migrate from c# to xml (that is their solution was half c# half xml file). So thinking in a tree data structure was beyond them.
Then Q1 2026 they were able to write parsers including parsers for languages where I specified with a bunch of examples.
But I tried specifying a language that I thought would require storing the AST in a graph structure. Everytime the LLM would instead parse into a tree and then do post processing on the tree to get the right output. So perhaps graph structures are (were?) beyond them.
I haven't investigated forth for a while, but it's been on my to-do list. I'll have to bump it up some because it sounds like I can kill two birds with one stone here.
I do kind of wonder if tree space is a local maximum, though. You can do quite a lot of things in that space and most of the code that I've read that leans on some sort of graphical thinking (like heavy usage of global variables that effect the execution path all over the call graph) is a less than ideal substrate to have to think through.
[One of the problems with investigating this is that I can't just ask it to solve a problem that is obviously just dijkstra because that's in the training data.]
There are perhaps two different modes of communication between people.
On the one hand we have vibing. Everyone just sits around telling each other things that affirm the group to be correct. No actual information is being transferred in the traditional information theory sense because saying something surprising would kill the vibe.
I don't really understand what this is good for but it happens all the time so I'm going to assume it has some social function.
Then there's communication that's more in line with information theory. That is information theory at least in part is dealing with some percentage of being surprised at what you receive. And so some human to human communication seems to be about hearing not what you expect but what you don't.
Personally, I seek the later, but I think most people desire the former.
And so, when the AI (or other people for that matter) tell me that I'm right, I assume that I don't understand the whole picture yet or am otherwise confused about something.
And on the other hand we've got the latent space of everything a human has ever written sitting inside a harness of ever increasing complexity.
Even if you put aside the qualitative nature of the analogy, the quantitative difference is so great that I'm not convinced there's any meat to get out of it.
Edit:
Actually here's a competing analogy. 210 BC emperor Qin Shi Huang drinks mercury in an attempt to live forever.
Then in 1996 Dr. Karen Wetterhahn has a drop of an organic mercury compound soak right through her latex glove, through her skin, and into her bloodstream. Where it kills her after a year of worsening physical symptoms.
Mercury, a tool that was super scary thousands of years ago, but remains unbelievably dangerous nightmare fuel.
I mean, I love rotary sanders too, but I'm not sure I'd call them the latent space of everything a human has ever written sitting in a harness of ever increasing complexity. That's not really my form of woodworking. But LLMs do fit on discs, even spinning HDDs on occasion (though I suspect it would be low tokens for second).
Mercury is incredible useful despite its hazards. Nitric acid is worse in negative affect despite its utility -- acid that rots your bones. When I visited fabs they were careful to point out which vats would burn me and which would make me instead lose the bone in an arm. Fortunately I survived the visits.
A rotary sander can damage plenty if you're not careful.
Yeah, the end was in desperate need of another editing pass. He takes you by the head, smashes your face into the ground, and then drags you for half a mile.
Anyone who gets that far into the book doesn't need to have their hand held that much. I get that it's the entire point of the book, but everything until this point has been a challenging read.
Personally, I think it might have been interesting to rewrite the whole thing such that conclusion is ambiguous.
My pet theory is that there are different types of intelligence that have different pros and cons. Social/cultural, intuition, and structural.
Structural is like step by step reasoning or math or raw compute.
Intuition is statistical from repeated trial and error.
And social is leaning on the wisdom of the crowds. So like high latitude countries where they eat fish for breakfast and get better health outcomes.
So I believe that LLMs have stumbled upon a partial component of our social intelligence. Word distribution, ontologies, jargon, information theory (frequently used symbols should be short). We mutate the language that we speak to be useful to us based on the problems we face. To some extent being able to talk the talk means you can also walk the walk. At least partially.
It's kind of shocking how far they can get, but at the same time it's kind of a surprise how far they don't. The existence of agentic harnesses is sort of an admission of defeat.
While some might be fooled into thinking that they reason, everyone I've met isn't. As a software engineer I'm drowning in work. And if that's not an admission that this isn't a real intelligence then I don't know what is.
But ultimately it looks like we've got all the individual components sorted. The old school 70s era stuff has a lot of the structural intelligence covered. The data science era of statistical ML has the intuition. And LLMs have the intelligence from our culture.
Maybe there are more general or energy efficient or powerful or special purpose techniques out there. And maybe combining everything together requires some additional insight. Regardless it feels like moving forward to something better than our current AI landscape is plausible, albeit with a completely unknown level of effort.
> Once you come to terms with the idea that translating an exact specification into decent code isn't a unique skill
My employer moved out of the "accept spec; produce code" space back in the early aughts. It was clear that early that this was going to be a bad space to operate in. (Although not for LLM reasons, obviously)
I've said for a while that the main reason Rust is so popular is that it has a lot of effort put into the developer experience, with the low-level safety honestly not being all that important to a large portion of the programmers who would be fine with a garbage collector. I used to think that maybe a "Rust with garbage collector" would come along, but at this point it honestly seems more likely that an optional garbage collector would be added to Rust (probably with just the primitives in std and leaving it up to libraries to provide a more full experience, like with async runtimes).
No, a class-based OO language where you need to spend effort crafting build targets by hand or use an IDE to define how to build is not anything close to what I'm talking about. If you think that it's "Rust with GC", I think you're misunderstanding what actually appeals to most people about Rust.
I'd also argue that "runs on all OS" is true, but "is easy to develop without extra work in a cross-platform way" is not. I've never cloned a Rust project and had trouble building out of the box on Linux, but I'd estimate maybe one out of 20 C# projects I clone from Github build for me out of the box with `dotnet build`; the rest either require me manually tweaking the build configs to avoid stuff like hardcoded Windows-style paths or link to system dependencies that don't exist on Linux. I imagine you might argue that this is a property of how people use the language rather than the language itself, but that doesn't really matter from the standpoint of whether it's worth it for developers who don't use Windows to spend any time trying to invest in the ecosystem.
Now all I have to do is make sure they're not peddling an actively bad idea and if not I can just tell the bot to incorporate their dubious ideas.
In the past it was all, I'm not familiar with this construct that solves our problem perfectly, why don't we cludge adapter pattern in here again.
reply