We are absolutely drowning in documentation and code that seems legit and the only recourse is to lean on AI to help process the sheer quantity of it. I have a feeling that the fallout from this phase of the industry is going to be an exotic form of technical debt that is remarkable mostly in its enormity.
LLMs are prolific and they love to add shit. Truly capable engineers are able to achieve more business outcomes with less code / fewer moving parts.
I'd simplify to "Truly capable engineers are able to achieve more positive outcomes" - half of what makes a capable, dependable engineer is knowing what outcomes are needed and making them happen.
> But there is a correlation between output and LoC.
That is less true today than it ever has been, due to LLMs.
High-quality code and high-volume code are highly anti-correlated. Incidentally, low-quality code that is excessively long just so happens to be common complaint with AI-generated code.
Perhaps they tackle non-code-editing tasks like architecture, design, mentoring and code review (think staff and principal tasks)
> Every line of code removed is a line that was previously added
Yes. This os not a failure. Code has a surprisingly short half-life.
What would you keep from this?
Because they were added doesn't mean they were needed and even if the same person added and then removed them, it doesn't mean they are digging ditches to fill them.
The idea that "I would have written a shorter letter, but I did not have the time" also applies to code, and sometimes later you are blessed with more time than you had when implementing something under deadline pressure.
Huh? If LoC weren't needed then adding them was unnecessary and a waste of time. Someone who is known at an organization for removing unnecessary code screams inefficiency to me. It's paying one person to create a mess then another to clean it up.
My previous reply already addressed this?
I can't help but think you are being purposefully obtuse if you can't acknowledge the concept of developers creating known (and hopefully temporary) technical debt due to various forms of deadline related time pressure or changing requirements.
i tend to find that the most productive teams make better decisions and work fewer hours. the quality of decisions is such a huge force multiplier that it renders actual hours worked almost an irrelevant variable.
• sloc = Source Lines of Code
.. so I suppose nloc would mean Net LoC
It's okay, I'm sure the algorithm questions during the interview phase totally weeded out the fakers of systems knowledge right?
Often the solution to create those document to feed into these AI automations? Use AI. Its like ouroboros. Create docs using AI, then summarize and ingest using AI, explained by AI.
Same thing is going to happen with code. Create 1000s of line of code using AI. Then explain it using AI etc.
I've yet to have Opus 4.8 fail me with defensive explict code. Often it'll write code that is better than what I might have done. I imagine it would be a nightmare to go through one of the OOP debug chains with implict error handling, but when every function has a runtime assertion which is basically the contract for how it is supposed to work and exactly what to do if it encounters a corrupt state, then things are just so much easier with AI.
I do agree with you on documentation. The amount we have has exploded in the post AI world. Which is a little ironic since the assertion is frankly what you'll need to know and not the 10 pages of prose the AI autogenerated in the shared loop (microsoft's terrible confluence). It is what it is though, and at least it's easier to meet EU compliance rules now, since those are more about the bureaucracy than actual security.
I think this is an important point. Software engineers always had the right instincts on how to approach AI for coding -- cautiously. Execs got too coked up on LinkedIn puff pieces from nobodies and adver-prophesizing CEOs selling their tokens and chips that they forced something unnatural upon their orgs.
Now what we see in the software dev space is incredible levels of malicious compliance ("you want slop, I'll give you slop").
In all seriousness, though, I'm indeed curious about Anthropic's engineering practice, particular how they can achieve such level of autonomy.
Chalk up yet another echo of the 1920s Gilded Age? Between all these economic spasms and the simultaneous tilting towards fascism, I think there is way too much historical rhyming going on right now...
> perfect formatting and at least superficial plausibility
Basically, a library full of books that have nice covers is going to take time to see that all those books are just filled with ipsum lorem. Before, they coudln't stand up a fake library.
The issue comes down to time and effort.
Worse if it’s a mixture of good content and ipsum lorem.
> Code becomes precious when it is the only place knowledge lives.
Reading AI code all day is _agonizing_. Just, a horrible way to live, and it melts people's brains at the moment you need them to be the most capable.
Manual programming has this really productive and gratifying feedback loop, where you read the code, write the code, and fix it until it compiles/runs/does what you want. AI code not only does half that for you, but it makes the "click" at the end uninspiring because you're never sure if it's cheated a bit to get to that moment.
Trying to operate with AI-generated code as the only durable artifact of programming is a dead end for the industry. Charity points to (and correct discards) architecture diagrams/specs as an interesting space to work in. My suspicion is that it's closer to the thing that's hand-written: prompts, markdown plans, and other nudges. Focus on the thing that you, as a human, produce, and that's the basis for both the core loop of "did the AI follow my instructions" and it's higher-leverage when you go to code review.
By the time you get to the PR, you've probably typed enough to Claude that you can regenerate the code, but the current industry default is to just throw away all those sessions and ship the code. That's backwards!
I think it's less about "break it down" and more about "let's communicate at the same altitude."
I wrote a (bait-titled) post about it: https://tern.sh/blog/stop-reading-prs/
305 files +15075 −13110
153 files +21934 −8698
125 files +28120 −2398
43 files +11188 −63
118 files +21564 −647
These are the largest (6 of 35) in the past 30 days. added: 190079 removed: 39696 in the last 6 months
from one person.
Because if all your SWEs produce 5x more code, it also means they have to review 5x more code. But LLMs don't really help with code reviews. Then it becomes a Metcalfian paradox unless you just rubberstamp PRs, which is what is expected of you.
if youre being asked to rubberstamp prs thats a management skill issue
But it's also the exact sort of thing that LLMs are literally perfect for in my experience so there's really no excuse anymore. I've never seen Claude fail to turn a 5k PR into a well-decomposed Graphite stack.
I would, and all my training at Google told me to do that. But what I found after I left that comfortable box was that somehow this kind of practice is acceptable in the industry at large and you're expected to just Deal With It(tm). 5k lines isn't even high by what I've seen.
Worse the "code review" tools that people have access to in GitHub make this absolutely and totally unworkable to incrementally improve review. Messy merge commits full of "responding to code review" comments. Threads impossible to follow. Just bad tooling.
So a lot of shops, from what I've seen, are just yeeting it with very shallow reviews.
This is my observation pre agentic AI. LLMs just threw kerosene on that dumpster fire.
https://github.com/gitsense/gsc-cli/blob/main/internal/cli/r...
Notice how the code block header attributes the model. The UUID can be traced to the conversation so everybody can tell exactly how the code came about. For this to work though, you need to use my chat app as it ensures you can't tamper with things if you are truly serious about AI code provenance.
I also have a lot more human-focused method which is part of my CLI tool.
https://github.com/gitsense/gsc-cli
I am currently looking at making pi (https://github.com/earendil-works/pi) support AI code provenance, but for now if you want a more structured way to capture what you have done in an agent session that can be used in code reviews and be carried forward as knowledge that lives inside your repository, I have
gsc lessons
The basic idea is, after you have finished chatting/working with the agent, you would work with it to identify lessons worth carrying forward. You can store your session if you want, but really, the lessons should be something that can help you review code better and to prevent future mistakes.
I have a real working example at
https://github.com/gitsense/smart-ripgrep
This is a fork of the BurntSushi/ripgrep repository. It shows how you can use lessons to learn from past design decisions.
First product compares the code to the prompts and highlights places the agent made decisions you weren't involved in: https://tern.sh/docs/tours/
depending on your industry, you might be able to ship half-slop and then fix some bugs downstream though
At the very least apply it at a higher level: specification, proofs, anything but generating Rust/Java/C and then letting yourself or an agent babysit it.
When starting on a new codebase, how do you make yourself into a helpful contributor as quickly as possible? I go straight for the humans and their human docs. What problem was the system originally built to solve? What was the original design, and what were its biggest problems? Who is currently using it? If you know these, reading the code is much easier because you can guess why things were done the way they are.
Also, this blog post has gotten popular: https://blog.gpkb.org/posts/just-send-me-the-prompt/
I think Charity is observing a very old problem and expecting the new technology to lead to a new solution of some kind. I doubt she thinks even the current generation of tools are the end of the AI software development story. She's not saying we'll drop design docs right into Claude code and walk away (design docs aren't complete either, that's why when you're ramping up you also have to talk to people, read old tickets and postmortems, etc.)
What she's observing is that, in prod, people don't like infra where it's hard to tell how it got into is current state, and so infra-as-code is what we do now. She's also observing that, "it's hard to tell how it got into its current state" is the status quo with codebases, which other people have observed going back to "Programming as Theory Building" and earlier. And she's expecting that, analogous to infra, software development will somehow be done with tools focused on making "how the code got into its current state" clearer.
> When starting on a new codebase, how do you make yourself into a helpful contributor as quickly as possible? I go straight for the humans and their human docs. What problem was the system originally built to solve? What was the original design, and what were its biggest problems? Who is currently using it? If you know these, reading the code is much easier because you can guess why things were done the way they are.
This is the way but plenty of engineering teams don't have any human docs at all. Decisions are made in one engineer's head or in a chat that isn't saved. The spec was just a few notes in a ticket that was deleted during cleanup or lost when the team changed trackers. There's no map of the codebase or features, no ADRs, minimal observability. All you have is the code. You read the code to try and figure out what is going on then ping an engineer who made a recent commit to a specific area to ask if they remember why something was done the way it was. Someone makes a change and it breaks something on the other side of the codebase that they thought was totally unrelated, etc.
Even the most AI-positive teams prefer human discussion when things get that tough. Given enough time, things will "click" for humans. LLMs don't work that way.
Even a team of all-new unfamiliar devs forced to study an old codebase will eventually figure out what it was about and pick up tons of nuance the LLM cannot. This is the nature of writing. It exists in a time and place beyond the pure literal text. Humans live in this context and can get into the headspace of the original dev(s).
AI needs more discipline, yes. But theoretically that discipline can be learned much easier than becoming a good engineer.
Think of it this way... 20 years ago, to write good, scalable C code - you needed to 1) either be a genius, or 2) dedicated to the craft.
You need to learn dozens of tools like the back of your hand.
* ASan
* LSan
* UBSan
* TSan
* GDB
etc... God forbid if you needed to manually read DWARF files. Unless you're a pure genius, this is not feasible to master in a short amount of time. And in parallel, you need to learn how to design systems, too, otherwise, you're still not very good, and that's an almost completely orthogonal skillset.
Now, you simply need to be aware of the hazards in your language/framework, tell your LLM to test for them, have the infrastructure set up to see if they've adequately tested for those hazards, and maybe read the actual tests and implementation.
It is pretty easy to be able to read and understand Rust compared to debugging all the sorcery-like errors that come during Rust development... It is easy to see that you need a Loom test for certain scenarios, and to write a tool to detect if you did it.
Even if you're still working in C or Zig, it far easier to know and detect when you need to use those tools then to learn to use them all individually.
It is not hard to learn to read SQL. Almost ~50% of business professionals can. Python is barely harder. Rust can look like sorcery if you don't read a 50 page guide to understand to read it, but that's a VERY small price to pay compared to spending ~10 years learning the craft painfully by trial and error.
I'm not sure how you get from "LLMs work in mysterious ways" to "So we need more discipline" to "everything is fine."
I agree that everything is fine. I just don't think this is the clear path and thought process.
Anyone who has the determination to get things to actually work, and takes a little bit of time to understand what makes them not, should be able to leverage LLMs to work wonders.
In my opinion, LLMs are going to make things far more complicated, because the cost of building something complicated is becoming almost free.
Engineering was always about discipline and getting things to work. But you needed a set of prerequisite skills to have much value. Most of those are gone now.
It is simply far easier than before. It does require discipline, yes. But discipline is cheap compared to ~10 years of trial by fire.
Rather than dismissing humans for quality control, we should take an asymptotic approach, where humans verify less and less as more verifications are automated, but are never out of the loop. Get down to 1% of the things, then 0.1%, then 0.01% and so on.
Automate all the linting you can before the agent is allowed to make a PR, make sure it passes the tests, add custom linting for dumb AI-isms you’re sick of telling the agent not to do - yes, you can lint for that fallback & backcompat code you never asked for, you just have the agent generate a script that walks the AST and flags the problem by line and file, then put that in your pre-commit checks - the agent treats it like just another lint error. Now you never have to review for that thing again.
But you still have value!
Even when you automated everything you can think of, there’s still tremendous value in human review. It’s your last chance to fully understand the implementation before it melds with the codebase. You also pick up more antipatterns to add to your automated reviewer (the automated reviewer is just a long prompt with an ever growing list of bullet points)
And the asymptotic nature of QC extends to observability and production. You cannot really ever automate a loop directly from observability to code fixes? Even when the agent presents a fix to an unhandled exception in production - if it was bad data, should you clean it in a backfill? If a key business metric dropped off a cliff because of a bug, should you add an alert once you fix the bug?
And I‘ve quickly realized that it’s also much easier to follow that premise.
Not only because agents obviously help with writing documentation, test cases, DX tools, and so on.
But also because it feels so much more rewarding to know that someone — even if it’s just a soulless agent — actually cares to read and use and follow these.
I have always been the guy on the team who would write the tools and documentation, and it’s always been a bit frustrating to know that only half the team would care to read and use and follow them, at best.
I've been thinking about this a whole lot recently. So much of my intuition about software development is based on 25 years of accumulated experience on how long it will take to write different bits of code.
Should I add validation for this one edge-case which won't break everything but will make a little bit of a mess if someone hits it? If that's an extra couple of hours of code I might skip it. If it's one more prompt, why wouldn't I?
This new feature would be a lot easier to understand if there was a custom API explorer for it. There's no way I could justify investing in that... unless it's just 10 minutes with Codex, and it was: https://tools.simonwillison.net/datasette-extras-explorer#ur... (linked from the release notes https://docs.datasette.io/en/latest/changelog.html#extra-sup...)
That's just on the small scale. There are entire projects that I'd never previously have considered, because I don't need a custom SQLite SELECT query parsing library enough to justify spending a week or more building one. But now... https://github.com/simonw/sqlite-ast
People get VERY upset (and condescending) any time you suggest that being able to produce lines of code faster is a valuable thing. And sure, measuring output through "lines of code" is stupid.
But measuring "lines of verified code that deliver valuable" isn't stupid at all. That's the thing we can do faster now.
I find it’s less common for me in ruby, even refactoring bad ruby. Sure I can remove lines but bad JS/React balloons so fast.
My current org values this and my direct boss constantly praises those of us that try to remove more lines than we add. Very refreshing.
if you could look clearly at the progress from 2020 to 2023, as someone like Gwern did, and from 2023 to 2024 with the invention of reasoning modes, then it was not that hard to understand what would happen in late 2025. Opus 4.5 was not a surprise to anyone who was actually paying attention.
But people (including the author) still mistake the current state as a stable state and future gains as incremental. he says
“I am not asserting that all code will eventually be AI-generated to spec, bypassing human understanding”
I AM asserting that, and it’s incredibly easy to do so.
The question of “when” is separate.
AI demands more of your hours cranking out the same discipline, due to the volume of stuff that needs to be verified.
Normally the term "more discipline" is understood as increased rigor, not simply more work at the same level of rigor.
> A sufficiently detailed specification is runnable code.
In a way I think LLMs will enable the dream of 4gl and "sufficiently smart compilers"[c].
LLMs aren't smart, but they are capable. Especially capable of translation and transformation.
I can certainly see them help move the abstraction horizon at which we work - so that rigid high level descriptions of the desired logic/process along with the process for quality testing - become the relevant curated artifacts - and the generated go/rust/java/python/etc code become incidental and mutable; subject to constant rewriting as part of the deployment of systems.
[c] You know, the ones that take naive C/C++ and produce executables that fully leverage RISC/EPIC platforms to be better than CISC. See also: Intel Itanium
1. What a C compiler was
2. What a C compiler looked like
3. What the C compiler had to do at runtime to pass gcc’s torture suite through some sort of collaborative iteration (compile, run, did it get stuck at some torture suite test or fail?)
Remove 1 and 2, or replace it with imperfect business logic, and you’re left with a system that is built to _only_ pass the tests you supply it, or in the most extreme case, print(“unit and functional tests pass!”)
What happened in 2025 was this: the economics of code production were turned upside down. Instead of being very hard, time-consuming, and expensive to generate code, it became effectively free and instant. Lines of code went from being treasured, reused, cared for and carefully curated, to being disposable and regenerable, practically overnight.
Also I disagree with:
Lines of code are not the ideal artifact to review
> What does "ideal" mean here? When I was growing up "show your work" was the rule for all examinations. Why? Because we're working to improve mental models and thought processes for the next generation, not just products we will release tomorrow.
They're saying that the mental models and thought processes are incredibly important but that code is not the place for that work to live.
What I meant is that, insofar as some work has been produced with a human mind involved and where imperfect abstractions are used, one should not for whatever idealistic reasons push for reviewing the work at some coarser granularity than the details which are readily available. That's a way to foster and encourage mistakes, in both the work and in the mental model.
So when you say that code is not the place for that work to live (or more closely to the line I disagree with, that code is not an 'ideal' artifact to review), you are essentially purporting that there is a perfect abstraction that can generally be trusted, which I disagree is currently the case for an LLM spec versus produced code.
They’re important for discussion and brainstorming. They’re also important for sharing context before reviewing. But code is the only perfect representation in terms of semantics of what the computer will do.
You can have all the diagram and all the proses you want, but they’re still ambiguous.
I suspect the stance they described as one readers mistakenly took away from their previous article to in fact be their stance. Otherwise why dance around it so hard?
I bet that I know why!
I now do documentation driven development, and with very few exceptions I am committing code that is better written, better documented, easier to reason about and maintain, with less library overuse than I ever did as a senior lead with a smal team, and I’m doing it for 1/4 the price, at 4x the speed.
But it’s not vibe coding. Discipline is critical, as is deep systems understanding.
That question was answered decisively last November."
It's easy to forget that people said this exact thing about every model after GPT 3.5. This is a standard trick the industry uses to invalidate negative experience with LLMs. 'You are prompting it wrong' becomes 'you are using Gemini, but you should use Clade' which then becomes 'well, all of your criticism is now irrelevant, because everything is fixed in this new version'.
This "discussion" about capabilities is set up to be asymmetrical and basically non-falsifiable.
I really don't know how I'm supposed to reply to stuff like this.
You undermine your own point when you misrepresent the situation like this. Real human mathematicians, including at least one Fields Medal winner, have validated and complimented the result.
It doesn't even occur to you that the claim made by Open AI has two necessary components. The first one is that the conjecture had been disproven. (This is what had been verified by "real human mathematicians".) The second necessary part of their claim is that the work to disprove the conjecture was done mostly by their AI model rather than by people employed by Open AI.
I agree that AI demands more engineering discipline, but it also demands more domain knowledge, purpose, and intent. Suddenly we can actually accomplish most of our goals a little faster. I can take on work I couldn't before.
Before I even begin getting disciplined about engineering, I need to ask: does this work actually make sense? Should I do it? If it's done... What do I think will change for my team or organization? Will it have practical results that move us in the right direction?
The better you get at asking that question, the less you'll find yourself prompting and planning and shoving PRs into the chute. It's still somewhat difficult to find important work in many places.
Still, engineering discipline is and always has been critical when going ahead with important work.
My gut feeling is that many of us simply aren't doing important work, and the discipline might be nice but is ultimately irrelevant. The sloppers are doing a faster version of something they always have, and much of it will be lost to time just like our pre-slop work has been.
I find LLMs aren't as helpful when applied to well-thought and intentional work towards very specific goals in complex domains. They're still helpful, but, the deeper you go and the more specific you get, the more they tend to deliver results you can't use. If you're on the rails they can be incredible. Diverging from the track and having exacting requirements, eh, it gets pretty hit or miss and you can spend a lot of time herding a digital cat. This certainly demands a lot more engineering discipline.
Now that AI coding speed and performance outperformed most of human. But AI still need human to be commanded. Yes, you can let AI agent manage sub-agents but still, human is at the top of manager who order AI what should be written.
So human must command and final say on when it's done.
Is laziness still a good virtue in AI era?
If you buy that, then it follows that the more work you accomplish with AI, the "lazier" of a dev you are.
That is still an enormous virtue in the AI era. It is completely the opposite of what many AI-using programmers are doing, which is being lazy in the conventional sense, minimizing their individual energy expenditure at the price of increasing the overall energy expenditure.
Being big-picture lazy is a virtue. Being individually lazy is a vice.
In general most developers are going to find themselves fighting incentives which will color their opinion. AI isn't there yet but if you are going to abase your whole world view on a point on a graph and not on the trajectory you are in for a bad time.
Writing software begins with a solid design that is defensible. If you don't have that, the AI will produce slop.
Once you're happy with the design, you need a solid plan. If you don't have that, the AI will produce slop.
Once you're happy with the plan, you can set the AI loose, but don't get too complacent! Anything that you missed in the previous phases could very well lead to slop (although likely localized).
And then then, as your project matures and you gain more understanding of the space, you start to notice deficiencies in your model. This is where AI really shines: design and code changes to adapt to reality.
The author makes the wrong assumption though that the majority of people who are doing engineering want to do even more engineering.
It’s my experience that most technology workers just want a high paycheck and have some kind of association with being in tech and doing cool things
yeh, I can see how that is now mistaken for a definition of 'engineer' or 'hacker'.
I am sorry you never knew what engineering truly means.
Guess who the author is.
> > The enthusiasts are not wrong. We are starting to see real, non-imaginary, discontinuous leaps in capabilities from teams that lean in hard to working with AI. And this does not feel like a normal technology cycle where you can wait for the dust to settle; teams that sit this out while competitors are hustling could be out of business before the dust settles. That’s a real, existential threat.
It’s not imaginary. It’s real. This time it’s different. And on a higher level, the FOMO is real. It’s not imaginary. It’s even existential.
Why do they all write the same as well? It’s so emphatic.
> The tech is cool, but as a thinking, feeling, breathing human who cares about other people, it can be hard to get excited about anything that so many people are this upset about. It’s also hard to get excited about something when so many of the loudest voices are out there talking gleefully about putting everyone permanently out of work, and so many artists and writers and people from developing nations are talking openly about the impact on them.
> Hold your desire to jump in and berate me here, I beg you. Like I said, I will deal with the ethics and morality of using AI in my very next post. Be honest, your attention span is no more up for reading a 10,000-word essay than mine is up for writing one. (Can we blame AI for that too?)
More Inevitability Soothsaying. All our feelings are crashing with Existentinal Threat Reality.
Was this article written by AI? It's certainly stupid enough!
- Schema validation with appropriate size limits on all relevant fields.
- Authentication.
- Access control.
- Backpressure management and rate limiting in case a (possibly malicious) user tries to perform too many computationally expensive actions in a short time.
- Ensuring that the actions of one user doesn't throttle another user which is connected to the same process/host, e.g. using async constructs to avoid freezing the main process.
- DDoS mitigation.
- Avoiding race conditions.
- Designing a good database schema, with well chosen indexes, with deterministic IDs/idempotency to avoid double-insertion scenarios. You don't want to be forced to rely on overly complex queries with a lot of joins. This doesn't scale well and rarely necessary.
- Logging and error handling.
- Avoiding conflicts and accidental overwrite with old data when multiple users are editing different fields of the same resource concurrently.
- Efficient distribution of realtime messages.
- Scalability.
The list goes on and on... And every piece has to be implemented perfectly. This involves a huge number of carefully thought-out decisions.