> The code they [LLMs] produce is often fine. It works. It passes tests. It might ship as-is.
I don't disagree, but I've been thinking about this a bit: a lot of _human_ written code was/is less-than-fine. And a lot of human devs didn't understand the context when they wrote it.
I'm not advocating that we fire devs, or evangelizing that LLms are awesome. But I do wish there was a slightly more honest take on the pre-LLM world: it's not just about cost reduction, it's about solving some long-term structural deficiencies of industry.
> a lot of _human_ written code was/is less-than-fine.
At least at my work place this is often the case whenever we bring in contractors who will never be tasked with maintaining their code.
Middle management rarely ever gives us the time to fix this broken mess, and when they do, it's actually in a neat maintainable state.
We recently took 700mb of compiled code and reduced it to under 40mb just by getting rid of useless libraries, and removing unused code, and consolidating all the build objects into one object. This code ran significantly faster too.
The result was that the release team was pissed and made a big stink about how this was causing or going to cause all sorts of issues. We surmised that it was because it reduced their OT.
It's been over a year and none of the issues have ever occurred.
My biggest issue with LLM written code is that it's often submitted by non-programmers, and often fails the first test, which is "does this even work, and solve the issue?". A test that I normally wouldn't even do since no one on my team would submit code that doesn't compile, or doesn't fix the problem.
The problem is that an actual person is submitting this code, and that we have to review it. They don't seem to know how to compile, or run a local version of our framework on docker to deploy it to.
> a lot of _human_ written code was/is less-than-fine. And a lot of human devs didn't understand the context when they wrote it.
I don't think anyone truly disputes this, but I don't think it's a good argument.
If we already had a sort of poor quality output when humans were at the helm, then what can we expect from AI steered by the sort of humans that produced poor quality code?
My guess would be even more poor quality code, at much higher volumes than before. It is impossible for me to see this as an improvement over "some people always made poor quality code"
> then what can we expect from AI steered by the sort of humans that produced poor quality code.
Great point, and I think that's my argument: above-average engineers can now produce more above average code. We don't need as many (any?) below-average developers moving forward.
It's possible, but I think unlikely. For one thing, most managers can't really tell who is above or below average. We still in 2026 have "lines of code" as a usual metric for judging the quality of a dev.
So how do you think that will play out in reality?
I'm not sure, but it might be recursive: above-average managers can tell the difference, and so we won't need as many below-average managers. But this requires above-average leadership, and the market will need to reward those organizations.
> it's not just about cost reduction, it's about solving some long-term structural deficiencies of industry.
You know, I hate that this is a world where I have to ask myself if this is LLM written because it is one of those patterns.
But that is besides the point of what I wanted to say anyway. Those deficiencies aren't going to be solved by LLMs I recon. In fact, they likely will make things worse. As you said, a lot of human devs didn't understand the context when they wrote code previously. True, but LLMs are even worse at context in many areas and still need human prompting for input.
The only thing I really see happening is that the blast radius of people not fully grasping the context and still producing something is going to be larger. More specifically, it is already larger. Previously incompetence limited the damage people could do, now that is less of a factor.
As a fellow cheese lover I would have loved for more geographical diversity, especially when it comes to sheep cheese. Ok, it didn't include Romanian telemea (I'm Romanian myself), but it could have at least gone for the Greek feta. Some Anatolian or Middle Eastern varieties would have also helped.
My first computer was a 486sx 25Mhz [1] The rig (tower, monitor, etc.) cost around $3,000. We got the SX instead of the DX because it was $500 cheaper. And I wanted a 16bit sound card. (Note that this is in 1992 dollars. Today it would cost over $7,000)
My parents didn't have a lot of money, but my great-grand father passed and they used some of the inheritance to buy the computer. I was instantly hooked. In hindsight I see how much of a gift my family gave me.
The announcement reminded me of article John Dvorak wrote around the same time. 1GB hard drives had just come out, and he asked what all the extra space would be used for. Even as a young teenager, I remember thinking how short sighted that comment was. That was before I realized how the tech press tends to get stuck in local optimizations, and can't understand the bigger picture.
It's all a good reminder that cutting edge today doesn't stay cutting edge very long, and the world figures out how to squeeze every ounce ounce of power out of hardware. (Also, yes, that leads to bloat...)
> In hindsight I see how much of a gift my family gave me.
True for many, many of us, I suspect. My family bought a 286 in the early 90s and it cost something like $2000 CAD then, which is nearly $4000 now; but salaries were lower then, this would have been something like 5-6% of my single income family's yearly post-tax earnings for the year, and if you think about it as the % of "disposable" income it was probably more like 60% of it for the year.
Obviously it paid off in that it set me on the path for my career, hard to make any other investment as good as that, but who would have known that at the time? I'm glad that there were so many ads positioning computers as being educational and not just game machines; even though in reality I think it was learning about the computer to make the games work that taught me way more than any educational software ever did.
Ha! Same for me: 286 in 9th grade (1990) for about $2k CAD. 286 was a bad call though as I think it was harder to expand compared with 386. I remember 1MB RAM but really only 640k usable. Had to change some BIOS settings to get to about ~700 kB?
I’ve been thinking a lot about these inflation-adjusted prices due to the big Apple Computer anniversary — an Apple // cost $5000 in 2026 dollars, meanwhile a $600 Macbook Neo cost $150 in 1980 cash!
What helped me reconcile this was an observation that we’ve inverted the prices of necessities and luxury goods. Rent and mortgage in particular were a much smaller slice of income back then, but luxury goods were very expensive, so one would save up for a year or two to buy a new TV or a computer for the kids.
Now the necessities take a much larger slice of our income, but TVs and computers are incredibly cheap. It takes very little money to get a nice computer, and not-buying it barely makes a dent in the bills. This isn’t a good thing.
I do disagree a little with your observation regarding the industry “squeezing every ounce of power out of hardware”. Beyond local LLM stuff, there’s basically nothing a modern computer can comfortably do that any laptop since the mainstreaming of SSDs can’t.
> You need something beefy if you're serious/professional about those
But you can get way better results with the lowest end computers than you could years ago.
Back in the 90s my grandfather used 3DS Max to map out his future apartment's rooms and start planning furniture, using renders to get an idea of how sunshine would look like at different times, etc. At the time, he did this on an expensive 486 that would take an entire day to render some of those visuals. Nowadays I can do the same with a free copy of Blender and any reasonably modern integrated GPU in probably under an hour.
You can just use SweetHome3D (GPL too) and call it a day. No need to mess with models, everything can be set and adjusted. Just design in a plane, set heights/widths for forniture, walls and the like, set a final render settings (hour of the day/sunny/cloudy and such) and even an Elementary kid could finish the work.
> In hindsight I see how much of a gift my family gave me.
Gotta tack on to this thread showing appreciation for parents. We could never afford new computers in the 90s, but luckily my dad could bring home obsolete equipment from work. We were thus always at least a generation behind. I remember my friend's Pentium feeling like sci-fi compared to our 386, but my goodness it completely molded my life!
Later, towards the end of the 90s, those sci-fi Pentiums were obsolete, so I got a few to run "that weird Linux stuff" on. Since it was considered junk, nobody cared what I did with it. To this day, if I happen to hear Metallica play and there's early winter's first smell of snow in the air, my mind will be transported back to that school night I secretly stayed up wayyy too late and discovered SSH for the first time. Haven't looked back.
Thank you, dad! I just hope general computing devices owned by regular people are still natural by the time my children come of age.
My grade school friend got a Nintendo and I wanted one so badly. My parents got me an Apple IIGS instead. I was a little disappointed about the Nintendo, but saw there were plenty of games on the thing, and of course it could do so much more than play games. That turned out to be a very good move on their part.
I wanted to link his columns "Microsoft Dot Nyet" and "New Architecture Needed" from circa 2000-2001 but it turns out they have been memory-holed. They should be somewhere in the wayback machine.
EDIT: At least one of them has not been deleted, just his name has been removed
Yikes, you're not wrong. And I guess he's never heard of security issues, what with his ROM idea. Neat for a console (where the ROMs are game cartridges, as they used to be) or an appliance not connected to the internet, not a general-purpose OS...
Pretty much the only thing I agree with is that computer architecture could use a complete rework (both from a software as well as hardware side, though primarily the former); as well as said rework being basically impossible in practice.
Mine neither although the grandparents were moderately wealthy but my mom understood very early on that it was a match for me and that computers would really take off.
Fun story: first BASIC I ever got was an Atari 2600 cartridge that came with some key of a "keyboard" in two parts you'd plug in the joystick ports. When my parent bought that Atari 2600 they tried it and spent the entire night playing "Tank Attack" on the TV in their bedroom. She only told me that years later.
Then as I was writing tiny BASIC programs on the Atari 2600 gaming console, she realized I needed a "real" computer, so she bought me an Atari 600 XL a bit later. Then I began salivating on the neighbours' Commodore 64, which I could see trough a window. And she thought: "If I buy the exact same computer as the neighbours, maybe my son and the neighbours shall become friends!". 42 years later one of our neighbor just went to visit my brother in another country and his brother we exchange Telegram messages nearly daily.
My mother was a stenographer. She used a 286 for processing docs. That baby wasssss alll mine during the day!!! All my friends had hacks for sys/bat/exe files to get wolfenstein at least to load. Best days of my life.
Wonderful gift, even though I gave it a pass and saved on the sound card.
Mine was the 486 DX 2/66.
The trouble with the 486 SX 25 was IMO that a fast 386 easily beat it. I was part of the demo scene back then and wanted to compete with the likes of Dust, Future Crew.
And: Doom! It could be displayed and run in 800x600 if I remember correctly on a DX 2/66.
This. My daughter is a high-school junior, and she's been asking for a laptop going into her senior year/college. This is exactly who Apple is going after.
I convinced my parents to get me a 2017 MBP for college, yes it was overkill for the day-to-day classes, but I ended up getting into iOS app development and was so fortunate to have a beefier-system. However, for a liberal-arts student the MBN appears to be a sweet spot.
I like the creativity, but I'm not sure it's needed. I've been building a large-scale database system using Opus 4.5, and targeting Rust. It's not perfect, but the Rust compiler is so helpful that Opus has solved a lot of problems on its own. I have around 100,000 lines of code, and have completed some major refactoring.
I am using a variation of spec-driven development.
This absolutely was NOT needed, just to be clear, it was just fun to do and, perhaps more importantly, it taught me a lot in the process of making it.
I also have a few geometric 3D printed objects on my desk that I made with openscad as "printing challenges" and then beat my head against my 3D printers for hours trying to actually print them. They did not need to be printed, they serve no purpose other than to be aesthetically pleasing and educational. :)
I once spent a couple hours debugging a perl cgi script. Nothing worked. Called in my colleague. Looks fine. We both were tearing our hair out. Sent it to the line printer, ordered pizza, and one of us read the code while the other typed it in. Couple hours later we finished and it worked.
> I suspect that there is a strong correlation between programmers who don't think that there needs to be a model/theory, and those who are reporting that LLMs are speeding them up.
I also strongly agree with Lamport, but I'm curious why you don't think Ai can help in the "theory building" process, both for the original team, and a team taking over a project? I.e., understanding a code base, the algorithms, etc.? I agree this doesn't replace all the knowledge, but it can bridge a gap.
I agree, the llm _vastly_ speeds up the process of "rebuilding the theory" of dead code, even faster than the person who wrote it 3 years ago can. I've had to work on old fortran codebases before and recently had the pleasure of including ai in my method and my god, it's so much easier! I can just copy and paste every relevant function into a single prompt, say "explain this to me" and it will not only comment each line with its details, but also elucidate the deeper meaning behind the set of actions. It can tell exactly which kind of theoretical simulation the code is performing without any kind of prompting on my part, even when the functions are named things like "a" or "sv3d2". Then, i can request derivations and explanations of all relevant theory to connect to the code and come away in about 1 days worth of work with a pretty good idea of the complete execution of a couple thousand lines of detailed mathematical simulations in a languages I'm no expert in. LLMs contribution to building theory has actually been more useful to me than is contribution in writing code!
From what I've seen they're great at identifying trees and bad at mapping the forest.
In other words, they can help you identify what fairly isolated pieces of code are doing. That's helpful, but it's also the single easiest part of understanding legacy code. The real challenges are things like identifying and mapping out any instances of temporal coupling, understanding implicit business rules, and inferring undocumented contracts and invariants. And LLM coding assistants are still pretty shit at those tasks.
You could paste your entire repo into Gemini and it could map your forest and also identify the "trees".
Assuming your codebase is smaller than Gemini context window. Sometimes it makes sense to upload a package,s code into Gemini and have it summarize and identify key ideas and function. Then repeat this for every package in the repository.then combine the results . It sounds tedious but it is a rather small python program that does this for me.
I've tried doing things like that. Results reminded me of that old chestnut, "Answers: $1. Correct answers: $50."
Concrete example, last week a colleague of mine used a tool like this to help with a code & architectural review of a feature whose implementation spanned four repositories with components written in four different programming languages. As I was working my way through the review, I found multiple instances where the information provided by the LLM missed important details, and that really undermined the value of the code review. I went ahead and did it the old fashioned way, and yes it took me a few hours but also I found four defects and failure modes we previously didn't know about.
Indeed, I once worked with a developer on a contract team who was only concerned with runtime execution, no concern whatever for architecture or code clarity or whatever at all.
The client loved him, for obvious reasons, but it's hard to wrap my head around such an approach to software construction.
Another time, I almost took on a gig, but when I took one look at the code I was supposed to take over, I bailed. Probably a decade would still not be sufficient for untangling and cleaning up the code.
True vibe coding is the worst thing. It may be suitable for one-ff shell script s of < 100 line utilities and such, anything more than that and you are simple asking for trouble
> I.e., understanding a code base, the algorithms, etc.?
The big problem is that LLMs do not *understand* the code you tell them to "explain". They just take probabilistic guesses about both function and design.
Even if "that's how humans do it too", this is only the first part of building an understanding of the code. You still need to verify the guess.
There's a few limitations using LLMs for such first-guessing: In humans, the built up understanding feeds back into the guessing, as you understand the codebase more, you can intuit function and design better. You start to know patterns and conventions. The LLM will always guess from zero understanding, relying only on the averaged out training data.
A following effect is that which bunderbunder points out in their reply: while LLMs are good at identifying algorithms, mere pattern recognition, they are exceptionally bad at world-modelling the surrounding environment the program was written in and the high level goals it was meant to accomplish. Especially for any information obtained outside the code. A human can run a git-blame and ask what team the original author was on, an LLM cannot and will not.
This makes them less useful for the task. Especially in any case where you intent to write new code; Sure, it's great that the LLM can give basic explanations about a programming language or framework you don't know, but if you're going to be writing code in it, you'd be better off taking the opportunity to learn it.
To clarify my question: Based on my experience (I'm a VP for a software department), LLMs can be useful to help a team build a theory. It isn't, in and of itself, enough to build that theory: that requires hands-on practice. But it seems to greatly accelerate the process.
People always wring their hands that operating at a new, higher level of abstraction with destroy people's ability to think and reason.
But people still think and reason just fine, but now at a higher level that gives them greater power and leverage.
Do you feel like you're missing something when you "cook for yourself" but you didn't you didn't plant and harvest the vegetables, raise and butcher the protein, forge the oven, or generate the gas or electricity that heats it?
You also didn’t write the CPU microcode or the compiler that turns your code into machine language.
When you cook or code, you're already operating on top of a very tall stack of abstractions.
Nah. This is a different beast entirely. This is removing the programmer from the arena, so they'll stop having intuition about how anything works or what it means. Not more abstract; completely divorced from software and what it's capable of.
Sure, manager-types will generally be pleased when they ask AI for some vanilla app. But when it doesn't work, who will show up to make it right? When they need something more complex, will they even know how to ask for it?
It's the savages praying to Vol, the stone idol that decides everything for them, and they've forgotten their ancestors built it and it's just a machine.
I agree with your sentiment. The thing is, in the past, the abstractions supporting us were designed for our (human) use, and we had to understand their surface interface in order to be able to use them effectively.
Now, we're driving such things with AI; it follows that we will see better results if we do some of the work climbing down into the supporting abstractions to make their interface more suitable for AI use. To extend your cooking metaphor, it's time to figure out the manufactured food megafactory now; yes, we're still "cooking" in there, but you might not recognize the spatulas.
Things like language servers (LSPs) are a step in this direction: making it possible to interact with the language's parser/linter/etc before compile/runtime. I think we'll eventually see that some programming languages end up being more apropos to efficiently get working, logically organized code out of an AI; whether that is languages with "only one way to do things" and extremely robust and strict typing, or something more like a Lisp with infinite flexibility where you can make your own DSLs etc remains to be seen.
Frameworks will also evolve to be AI-friendly with more tooling akin to an LSP that allows an MCP-style interaction from the agent with the codebase to reason about it. And, ultimately, whatever is used the most and has the most examples for training will probably win...
I agree with this sentiment. Perhaps this is why there is such a senior / junior divide with LLM use. Seniors already build their theories. Juniors don't have that skill.
